PRODUCTS

GPT-5.6 takes big step toward OpenAI's superapp

OpenAI just did a lot more than release its most advanced language model. 

While GPT-5.6 went into full public release on Thursday, as we've been expecting all week, the company also announced a ton of other developments and updates at the same time, including the next big step toward its long-awaited superapp. 

Twitter tweet

Like Anthropic's Mythos, the US government worked with OpenAI to delay the broader release of GPT-5.6 until government agencies and corporations were better prepared for the model's ability to uncover software vulnerabilities that could enable cyberattackers and other bad actors. 

The centerpiece of today's announcement remains the GPT-5.6 family of models, which now includes Sol (flagship), Terra (balanced), and Luna (fastest and least expensive). The models have been performing like champs on benchmarks such as Agent's Last Exam, outperforming Anthropic's Mythos and Fable. The OpenAI models are also less expensive. For 1M tokens, Sol costs $5 input / $30 output while Mythos and Fable cost $10 / $50. 

But OpenAI also used Thursday's announcement to roll out several other significant updates:

• Codex moves into the ChatGPT app: This has been in the works since March, as Codex transformed from a coding agent to a more general agentic tool after the meteoric rise of OpenClaw, whose founder, Peter Steinberger, joined OpenAI's Codex team. The combination of the two apps is still a little confusing, but the OpenAI team assured me that the Codex brand will remain and the agent harness will continue to be open-source. ChatGPT users will now get an AI agent inside the chatbot they already use. 

• Launch of ChatGPT Work: Perhaps the most significant new development is the launch of ChatGPT Work, an agent that brings the kinds of tasks for knowledge work that people have been doing with Codex on the desktop to the ChatGPT app across web and mobile. There will be a new toggle between Chat and Work when you load ChatGPT, so you can do agentic work from anywhere you're already using the OpenAI chatbot.

• ChatGPT Sites: From ChatGPT Work you can now describe a site you want to build in natural language with no understanding of code, and it can make it, deploy it, and give you a live link on the web that you can share with anyone. This handles the deployment work that you would have previously needed a tool like Lovable, Replit, or Vercel to do.

"We have more than 5 million people using Codex every week, but close to a billion people using ChatGPT. And so, for the largest part of the audience, they had not really seen the power of agents," Romain Huet, head of developer experience at OpenAI, told The Deep View. "[Today begins] this progressive reveal for almost a billion people that can now delegate more tasks of more complexity to something like ChatGPT Work to do more things for them."

While we expected the GPT-5.6 launch this week and the models themselves are a known quantity at this point, the first big step toward bringing Codex and ChatGPT into a single app experience was a bigger surprise. Of course, it's long overdue since OpenAI executives such as Greg Brockman have been talking about it since the end of Q1. But it's not easy, since Codex had turned into an agentic power-user app used by rabid enthusiasts, while ChatGPT is the mainstream interface used by a billion people. The first steps still look a little confusing to me, but it was time to start bringing them together to create a more unified experience, and I'm glad OpenAI has begun the process. In the short term, starting to introduce a lot more people to the time-saving power of agents will be the biggest win. And running OpenAI agents across desktop, mobile, and web will be the big upgrade for the power users who already rely on Codex.

IN PARTNERSHIP WITH LAMBDA

Cut your AI training costs by 25% or more

Most large-scale AI training runs use less than half the computing power they're paying for. Lambda's team found the root causes and built a reproducible framework that boosted efficiency by over 25%, without changing the model itself.

Lambda’s whitepaper shows you how to address:

• Memory inefficiencies silently inflating your costs

• Training configurations that aren't making full use of your hardware

• Bottlenecks that slow down GPU communication

Get the whitepaper.

PRODUCT

Can Claude's new screentime tool make AI healthier?

Wondering how reliant you've become on AI tools? Anthropic is giving you clarity.

On Thursday, Anthropic launched a new reflection dashboard in Claude's settings for the web or desktop app. The dashboard gives you a summary of how you've been using Claude, including key topics, your usage patterns, and the type of tasks you usually use it for, according to the blog post. You can access it at claude.ai/settings/reflect. 

Twitter tweet

Beyond giving you greater insight into how you use the tool, there are also options users can take to modify their behavior, such as setting up quiet hours and nudges to take breaks. Periodically, the reflection dashboard will also present users with questions prompting them to reflect on their AI usage, which they can then discuss with the chatbot. 

Soon, there will also be a view of how much time users have spent using Claude. It is now available in beta on the Free, Pro, and Max plans on web and desktop. It runs only when memory is on, as it needs to refer to past chats, so if, like me, you don't see it, it may be because you have that setting off. The reflection dashboard leaves out incognito chats and health integration content.

While the aim for a feature like this is good, helping users become aware of just how much of their lives they're offloading to AI, Anthropic isn't holding back on fueling AI usage either. On the same day, it announced a reset of the 5-hour weekly rate limits for all users of its Fable model, in response to high demand for its most advanced model and to compete with OpenAI's release of GPT-5.6.

People have been over-reliant on their devices for decades, seduced by the entertainment and convenience sitting right in their pocket. AI's hook is different: it doesn't just entertain; it promises to make your life better and your work more productive. Whether that's genuine ambition or laziness in disguise, AI can now help with many tasks tied to knowledge work or human interaction. Just yesterday, I wrote about how many adults are turning to it for emotional support. The first step in recognizing a problem is seeing it quantified, but we've been here before. Screen time trackers proved that awareness alone rarely changes behavior. Still, even if this doesn't move the needle on healthy usage, there's value in knowing exactly what you're leaning on AI for, if you're willing to evaluate it honestly.

IN PARTNERSHIP WITH DATALAB

Turn docs into records that fit your schema. Open source

Datalab released Lift, an open source 9B model that turns messy documents into clean, structured data:

• Pulls user-defined fields into your schema

• Works with any schema - invoices, scientific papers, 10-Ks, etc.

• Fast - 9.5s median, 8x quicker than Azure

• Run it locally or via Datalab's API for scale

Try Lift on your own documents - select 'Extract' to start.

RESEARCH

CEOs think AI is working. The data disagrees

AI promises to automate workflows and save employees time. But a lack of alignment among company leaders may keep those productivity gains out of reach.

While 85% of C-suite executives report deploying AI across their organizations, only 54% of managers say AI is a top priority, according to new data from Nitro, a document software company, which surveyed more than 1,300 executives, managers, and directors across the US, UK, and Canada. Barely half of managers say AI has reached at least some of their own workflows, revealing a disconnect between executive aspirations and day-to-day realities.

"At an executive level, it can be about the message: 'Yes, we are using AI,'" Cormac Whelan, Nitro's CEO, told The Deep View. "At the manager level, it is often about solving for the next level of detail on how we are using it."

Among managers whose teams have adopted AI, 37% primarily rely on standalone AI tools, such as copying and pasting content into ChatGPT. Only 12% say AI is fully embedded into their document workflows.

Much of the work remains manual. More than half of managers say employees spend the most time extracting data from documents into spreadsheets, followed by editing and formatting files, managing documents, summarizing content, and redacting sensitive information. Because of that, 62% of managers say employees spend six or more hours each week on manual document tasks, with nearly a third estimating those workloads consume 11 or more hours.

"If AI isn't built into the workflows, work surfaces, and systems people use every day, with the outcome in mind, it doesn't matter how many tools you buy or how many tokens [you burn] or how much budget you commit," Whelan said. "The work stays manual."

The findings underscore a growing challenge in enterprise AI adoption. Businesses are betting big on AI to boost productivity and cut costs, but many managers appear to be taking a more measured approach to rolling it out across their teams.

Managers cite security and trust as their biggest barriers to AI adoption, followed by integration complexity and implementation costs. More than half also say sensitive documents are being uploaded to public AI tools, while only 43% report having clear, actively enforced AI policies.

To close the gap, the findings suggest that companies need AI embedded into the software and workflows employees already use every day, with security and governance baked in from the start. Otherwise, AI risks becoming another set of tools rather than the productivity engine executives expect it to be.

AI adoption is as much an organizational challenge as it is a technological one. While executives focus on transforming the business to appease stakeholders, managers are responsible for making AI work day-to-day. As layoffs shrink middle management and spread teams thin, many are juggling a mounting list of responsibilities. They're not only expected to lead their teams, but also figure out how to integrate AI into existing workflows and systems. Closing that gap goes beyond creating an AI strategy. It requires building a culture where employees have the time, training, and trust to experiment with AI out of curiosity rather than purely top-down pressure.

LINKS

• Meta introduces Muse Spark 1.1, a low-cost multimodal reasoning model

• SpaceXAI rolls out Grok 4.5 model aimed at coding, agents, and cost savings

• Open-source AI tool Ollama raised a $65 million Series B

• OpenAI published a detailed audit that found task issues in SWE-Bench Pro 

• Cerebras, Flex expand partnership to scale CS-3 AI chip production capacity

• AI companies are spending millions of dollars on elections, report 

• Starbucks uses AI for tools to cut reliance on Microsoft and IBM software

• Notion: Introduced Ship OS "the agent-native way to ship software"

• Claude: Has a new reflection dashboard to see your usage patterns 

• Muse Spark 1: Meta's new  multimodal reasoning model 

• Manus: Introduced branch, a feature that lets you split any conversation into a parallel session that inherits your full context

• Google Photos: Video Remix turns clips into stylized videos, powered by Gemini Omni

• National Hockey League (NHL): AI Engineer

• BlackRock: Artificial Intelligence & Machine Learning Engineer, Associate

• Capgemini: Gen AI / Agentic AI Developer

• Brooklyn Sports & Entertainment: AI Engineer

GAMES

POLL RESULTS

Would you use AI for relationship advice?

No (54%)
Yes (25%%)
Maybe (21%)

Google's AI Overviews now shape how millions of users find information, making the act of search practically synonymous with turning to AI. 

The AI integration that Google has thrust upon users promises quicker, more conversational results, but what happens when you can’t trust the results being given? 

Usually, I use Google in the same way that billions of people do every day: type a question into the search engine and trust whatever the AI tells me, my boyfriend's pleas to "actually Google it" notwithstanding. So, when The New York Times published a report calling into question the accuracy of AI Overviews in April, I clicked immediately, curious as a reporter covering AI and invested as someone who often uses it without a second thought.

The report and its findings 

The New York Times commissioned an analysis by an AI start-up called Oumi, which found that AI Overviews were accurate approximately 90% of the time. While 10% inaccuracy may sound small, the sheer volume of Google searches makes it a massive concern. 

As the report highlights, Google processes over five trillion searches annually. The Times claims that this translates to tens of millions of erroneous answers every hour or hundreds of thousands of inaccuracies every minute. Those figures are difficult to pin down precisely because not every search query triggers an AI Overview. Still, Google CEO Sundar Pichai disclosed last July that AI Overviews had surpassed two billion monthly users.

"The number that should concern everyone is not the accuracy rate, it is the scale," David Bader, distinguished professor and director of the Institute for Data Science at the New Jersey Institute of Technology, told The Deep View. "That is not a minor imperfection; that is misinformation operating at industrial scale, delivered with the visual authority of the world's most trusted search engine."

Methodology

Oumi ran the first round of the study in October, when the most complex questions were answered by Gemini 2, and then in February, when it was upgraded to Gemini 3, which remained Google’s most capable mode until Google I/O in May, when Gemini 3.5 was launched.

In both cases, the analysis focuses on 4,326 Google searches: Gemini 2 was accurate 85 percent of the time, and Gemini 3 was accurate 91 percent of the time. This was based on OpenAI's factuality benchmark, SimpleQA.

Google has since refuted the analysis’s results, claiming that the SimpleQA dataset included incorrect target answers and that it used one AI model, HallOumi, to assess the accuracy of another AI model, which could introduce errors. 

"This study has serious holes. It uses one AI to grade another on an old benchmark that is known for being full of errors, and it doesn’t reflect what people are actually searching on Google," a Google spokesperson told The Deep View. "AI Overviews are built on our Gemini models, which lead the industry in accuracy, and they clear the same high-quality bar that we have for all our Search features."

Google also added that, because the dataset was artificially created, it does not accurately represent real-world human queries. 

OpenAI created SimpleQA’s query dataset using AI trainers that browse the web to generate questions and corresponding answers. The questions had to meet the criterion of having a single "undisputable answer"; the answer couldn’t change over time; and they had to induce hallucinations from either GPT-4o or GPT-3.5. Beyond this, there were two more rounds of quality verification. 

Google’s point that the benchmark queries are not entirely representative of an everyday user’s experience is "somewhat valid," according to Maarten Sap, assistant professor of natural language processing at Carnegie Mellon University. However, he said that third-party evaluations are crucial to avoid conflicts of interest and get truthful results. 

"The methodology of probing these systems via the regular UI is good, and unfortunately, the only option we have since these products don't typically have a more controlled API endpoint that researchers can ping," Sap told The Deep View. 

The need for easier access to models 

Both Google’s concerns and Sap’s insights highlight a major industry issue: the inability to get proper access to the models to accurately conduct the kinds of third-party evaluations necessary to hold companies accountable. 

For instance, a 2024 report from the AI ethics nonprofit the Ada Lovelace Institute found that AI model evaluations are currently voluntary and subject to company discretion, leading to inconsistent quality and limited access for evaluators. 

In 2023, the University of Oxford released a paper that examines the gated release of frontier artificial intelligence models, which prevents the conduct of important safety research. 

"I would also argue that companies deploying AI at this scale have an obligation to fund independent auditing," said Bader. "Google disputed the Oumi study, but the company has not published its own comprehensive, independently verified accuracy data. If the product is as reliable as Google claims, they should welcome external scrutiny, not dismiss it."

The underlying problem: hallucinations  

The root cause of the AIO accuracy issue is hallucinations, which, while triggered by a number of factors, are inherent to how AI models work. These models are not actually thinking but rather regurgitating facts by predicting the next word in the sequence.

"These systems do not look up facts in a database. They predict the most statistically probable next word in a sequence based on patterns learned from training data. They are, at their core, sophisticated pattern-matching engines," said Bader. 

This explains why models often show confidence or give outputs for their answers, even when they are not correct, he said: "They have no internal mechanism for distinguishing between what they 'know' and what they are guessing." 

Olivier Toubia, Columbia University business school professor, told The Deep View that hallucinations can also be traced back to the training data itself, especially if models aren’t trained on an entirely accurate and labeled set of information. These models aren’t trained to sort out correct from incorrect. Instead, he said, they take in typically unlabeled training data and process answers based on it.

"Given all that, it’s very challenging for an AI model to capture uncertainty in the same way as humans think of it, and to model correctness," he added. 

Toubia said that AI can "model uncertainty at the token level," referring to the model's probability distribution over every possible next word at each step. For example, if a user asks what the capital of France is, "Paris" would receive a higher probability than "Berlin." 

However, that confidence score reflects statistical patterns in training data, not factual understanding. The model isn't verifying that Paris is correct; it's recognizing that "Paris" most commonly follows that prompt. That's why token-level confidence doesn't equate to certainty about factual correctness.

To exacerbate the problem even further, these models often express certainty despite not being explicitly trained to do so, a phenomenon Apple Research identified as resulting from popular training methods: Instruction tuning and RLHF (Reinforcement Learning from Human Feedback). 

Instruction tuning trains the model to follow directions and give clear, direct answers. RLHF uses human ratings to reward responses that seem helpful, confident, and well-written. In both methods, the model is rewarded for sounding certain, which, of course, is harmful because it will continue to replicate this behavior.

"Hallucinations persist partly because of how these models are trained. During training, AI models are often rewarded for correct answers and for confident answers, which can incentivize them to guess even when they aren't certain," said Lewis. 

He adds that it's not much different from how a human, under high pressure, answers a difficult question. We often perceive that it would be better to say something confidently that's not true than to admit uncertainty. And this is reinforced at a cultural level by messages like the "fake it till you make it" strategy.

People’s predisposition to believing the models

The problem, however, is that these models assert that they are completely correct, giving people false confidence that the answers are true simply because they sound authoritative, as Sap’s research found. 

In one of the experiments, participants were shown trivia questions paired with just the beginning of a response, without ever seeing the actual answer. These response starters included epistemic markers, words or phrases that signal a speaker's confidence in what they are saying. 

The responses came in three forms: a plain statement with no epistemic marker, one with a "strengthener" (language expressing certainty), and one with a "weakener" (language expressing uncertainty). Nearly 90% of users trusted the language when a strengthener was used. Conversely, approximately 90% chose to look up the answer themselves when a weakener was used. 

Photo: Jason Hiner

Another research paper Sap worked on found that politeness can also increase trust and reliance. "Beyond whether the model is correct or not, how users interpret AI outputs and how they're packaged is also crucial," said Sap. 

The footnotes

Assertions of confidence and correctness aren’t the only factor that helps Google garner trust. 

Another contributing factor to giving the false impression of accuracy is linking to footnotes, which makes outputs feel more believable, especially since it pulls information from Google’s AI search engine. Given that Google has spent its entire existence establishing authority, people may overlook the tiny disclosure on each output stating that it may hallucinate. 

"When you position your AI system as the first and most prominent answer a user sees, you are implicitly telling people to trust it," said Bader. "A small disclaimer that says ‘AI responses may include mistakes’ does not absolve you when your system is producing tens of millions of wrong answers every hour." 

Sap adds that it is a "huge liability" for Google to show these error-prone AI overviews above actual links to web pages, given the margin of error. Rather, he suggests that a different tab would be "slightly" better. In particular, he worries about the negative consequences related to sensitive matters such as health.

"I particularly worry about people entering health-related queries and the AI overview giving them incorrect information and that endangering them, but this is hard to study unless you know what queries people put into Google," said Sap.

What can be done?

AI will only continue to be more deeply integrated into Google Search products over time, as seen most recently at Google I/O, the company’s annual developer conference, where it unveiled "the biggest upgrade to our Search box in 25 years." This includes AI-powered suggestions that go beyond autocomplete, a more seamless handoff between Search and AI Mode, and agents galore. 

Nevertheless, hallucinations aren't going anywhere. 

"The bottom line is this: hallucination is not a software bug that will be fixed in the next release. It is a fundamental property of large language models' text generation," said Bader. "The industry can and should continue driving error rates down, but users, policymakers, and organizations need to understand that some level of AI-generated misinformation is here to stay." 

As a result, the best thing users can do is be aware of the issue and, as much as possible, double-check the response by visiting the provided links or doing their own research, especially when verifying critical topics such as health information. 

"To be fair, companies are already disclosing that their models might hallucinate, and they are already suggesting to users that they should check," said Toubia. "In a way, the problem also comes from the users." 

Toubia explains that many people fall into the fallacy of convincing themselves that, although AI Overviews are only correct 90% of the time, every query they send will fall into that 90% and can be trusted. 

 "There seems to be a dose of wishful thinking in how people use LLMs, which leads them to outsource important human tasks to AI entirely, without spending the effort to check the output, even if they know it might be incorrect."

If you want to get in front of an audience of 750,000+ developers, business leaders and tech enthusiasts, get in touch with us here.

PRODUCT

How OpenAI's voice assistant got more natural

While native device assistants have long left users wanting more, OpenAI's voice models have quietly been getting closer and closer to the assistants we've always been promised. The latest version of ChatGPT Voice just took another big step in that direction.

On Wednesday, the company launched GPT-Live, a new generation of voice models that aims to make talking with AI feel even more like having a conversation with a human. At the heart of the upgrade is a new full-duplex architecture, which enables simultaneous listening and speaking, reducing latency and creating a more natural experience. 

When you are speaking to it, it can say phrases that mimic engagement, such as "mhmm," engage in quick back-and-forth banter, or even give you more time to think without interrupting. Even as GPT-Live handles more complex queries, it will continue to talk to you to maintain the flow of the conversation. 

It can do so through the second architectural change, which decouples the voice model from deeper work. As a result, when a query requires deeper reasoning, GPT-Live offloads that request to another model, GPT 5.5 at launch, and then processes it in the background so that it can keep the discussion going with you.

For instance, in a demo given to the press ahead of the release, the model performed real-time, simultaneous translation into Hindi while the presenter was speaking, then chimed in with the completed Hindi translation. This was a standout feature because, in the past, while it was processing, it made an unnatural thinking sound, causing an unnatural pause and forcing you to stop the flow of conversation.

All of these updates contributed to GPT-Live outperforming Advanced Voice Mode, its previous technology launched with GPT-4o, in both human-preferred conversations and technical benchmarks, such as scientific reasoning and agentic search, according to OpenAI. 

Starting today, OpenAI is rolling out two versions of GPT Live: GPT-Live-1, powering ChatGPT Voice for Go, Plus, and Pro Users, and GPT-Live-1 Mini, the default for free users. They will also be available in the API soon, and interested users can fill out a form to get access. 

Ultimately, one of the biggest roadblocks to using AI is its reliance on text responses, which can add friction to everyday routines. This makes GPT-Live important for broadening AI's impact, as it enables users to interact with AI more seamlessly. The predecessor was already helpful, and I even mapped my iPhone's action button to ChatGPT's Advanced Voice Mode to use it regularly throughout the day. Still, there were quirks that made it obvious you were talking to AI. But it also raises other challenges because it sounds so human-like; since something that sounds more human is more likely to be believed, and it still has the same hallucination issues as every other AI model.

TOGETHER WITH CONVEX

Code Faster, Easier, And Less Buggy With Convex

You’re probably already familiar with Convex, the backend building platform that helps your AI agents excel – but their latest updates are taking things up a notch. Convex recently launched plugins that allow you to connect Claude Code, Codex, GitHub Copilot and more directly to your Convex deployment, allowing them to…

• Read data and run functions entirely on their own

• Optimize your project’s performance via production insights

• Improve safety by limiting access to production

There’s even more updates coming in the near future, but for now, see how you can take your agents’ work to the next level with the latest from Convex right here.

GOVERNANCE

AI is now a key chess piece in US-China relations

If you needed further evidence of how powerful AI models have become or that they've risen to the level of national security concerns, then the recent actions of the world's two largest economies should be enough to convince you.

The US government restricted international access to Anthropic's Mythos and OpenAI's GPT-5.6, and now China is looking to do the same for the most advanced models from its frontier AI labs. 

When Washington realized that Claude Mythos could be used by attackers to find software vulnerabilities and exploit systems faster than anything we've seen before, it took the drastic step of issuing export controls on the software, which essentially locked it down for three weeks while the US government and major US corporations had time to harden their systems. 

According to a Reuters report, in June the Chinese Ministry of Commerce held meetings with AI labs, including Z.ai, Alibaba, and ByteDance, to discuss restricting international access to the most advanced AI models produced by Chinese labs. The government officials also raised the possibility of limiting international investors' ability to fund Chinese AI startups. 

The Chinese move could have ripple effects across the US, given the widespread use of open-weight Chinese models by US corporations and AI enthusiasts. A report from OpenRouter found that over 30% of the tokens used on Chinese models came from the US, and that number has surged as high as 46% during some weeks in 2026. 

With AI costs soaring in many organizations, IT departments and workers have turned to open-weight models since you can download them and run them locally to save costs, improve performance, and maintain privacy and security. And the Chinese open-weight models like DeepSeek, Qwen, and GLM are generally considered the most advanced. 

If China moves to restrict US access to these models, it could boost other open models made in the US, such as Nvidia Nemotron, Google Gemma, OpenAI GPT-OSS, and others. 

Meanwhile, the rhetoric continues to escalate. On Wednesday, the Chinese Ministry of Industry and Information Technology warned that Anthropic's Claude Code "contains a security back-door vulnerability that poses a serious threat." Claude Code is not officially available to use in China, but Chinese software developers have publicly stated that many use it, according to a CNBC report. And that appeared to be backed up by the fact that Alibaba banned its workers from using Anthropic tools on Monday.

Before we get too overreactive to the US-China rhetoric, let's keep in mind that none of the recent activity is likely to stop Chinese and American technologists from using each other's models anytime soon. For example, Anthropic's Mythos is now widely available to the public. The US government's restriction simply gave agencies and corporations time to prepare their cybersecurity posture before the model became more widely available. And the Chinese government is discussing restrictions on the most advanced models. The vast majority of open-weight models would likely remain available and wouldn't affect most users. In the same way, all previous-generation Anthropic and OpenAI models remained publicly available even as the government forced the companies to temporarily restrict broader access to their latest models. In other words, American-Chinese cooperation will quietly continue beneath the surface of the latest escalation in rhetoric.

TOGETHER WITH COMP AI

AI is automating everything. Except the one thing killing your enterprise deals.

It's compliance.

Your pipeline is full. Your product is ready. But the moment an enterprise prospect runs a security review, the deal stalls. SOC 2 isn't optional anymore - it's table stakes.

Comp AI's AI agents automate the whole thing: pulling evidence from 580+ integrations, generating policies for your stack, running continuous monitoring.

See How It Works

Open source, fully auditable, and built for the AI era. Startups like Dub, Corgi, and Better Auth use Comp AI to stay permanently audit-ready.

Book a Demo

RESEARCH

Study: 38% open up to AI about relationships

Emotional support and advice have long been difficult to come by, prompting a growing number of people to turn to AI for more help.

The Oxford Internet Institute surveyed 2,000 UK adults online to learn about their interactions with large language models and found that almost one-third (31%) of regular users reported using these models for emotional support, including talking through problems or making decisions. Furthermore, one quarter of regular LLM users look to these tools for meaningful conversation. 

This trust even extended into more sensitive topics, with 38% turning to LLMs for advice on personal relationships and two-thirds for information on health issues. The demographics of the respondents tended to skew the level of comfort and reliance on using chatbots for different tasks: 

• Younger users: more likely to use LLMs across all types of activity

• Women: more likely than men to use LLMs for personal and emotional support

• Men: turned to AI for help with practical tasks to a greater extent than women

Despite the differences, the overall consensus was positive: 75% of the respondents were enthusiastic about the possible benefits of AI chatbots. It is worth acknowledging that there are risks associated with using AI for sensitive issues that have traditionally been handled by humans, given the emotional intelligence and social skills AI cannot replicate. 

"Our study provides an important starting point for understanding LLM usage trends, but substantial further research is needed to explore the wider societal implications associated with LLM use for social and emotional support, including when such interactions may complement existing sources, and when they risk replacing important human relationships," said Florence Enock, Ph.D., Senior Research Fellow at the Oxford Internet Institute and lead author of the report.

Existing research from the Human-Centered Artificial Intelligence Institute at Stanford University found that AI therapy chatbots can contribute to harmful stigma and dangerous responses and are, as a result, less efficient than human therapists. That's not even factoring in the potential privacy risks involved with giving a chatbot sensitive personal information such as the state of your mental health.

Chatbots are becoming increasingly skilled at understanding natural language and mimicking the human experience with greater consistency. This makes it easy to personify or give human attributes to models that are just analyzing patterns and repeating them, which can appear to look like consciousness. Combined with the fact that AI labs are giving users free access to even their most advanced models, this creates a recipe for users to develop a dependency on AI for all sorts of tasks, including personal or sensitive ones. However, using any chatbot for assistance with mental health should be handled very carefully, both because of the privacy concerns and because of the unreliability of LLMs due to hallucinations and sycophancy. 

LINKS

• Apple reaches a deal with Broadcom to spend $30 million on US-based chips

• Meta rolls out update to turn off the camera on glasses if privacy light is tampered with

• OpenAI development arm acquires applied AI firm Northslope

• Chip startup SambaNova raised $1 billion at an $11 billion valuation 

• Mistral AI launches Robostral Navigate, model that helps robots navigate environments

• Chinese AI lab MiniMax plans to launch a 2.7-trillion-parameter model

• Claude Cowork: It is coming to mobile and web, with beta rolling out over the next several weeks, starting with the Max plan, with more plans to follow.

• Manus: Google Drive connector got upgraded to Google Workspace

• BytePlus: Dola Seedream 5.0 Pro API is now available

• GPT-5.6 Sol: Launches on Thursday 

• Grok 4.5: Cursor and SpaceXAI released its "most intelligent model"

• Capital One: Applied Researcher II (AI Foundations, LLM Core and Agentic AI)

• Basis: Revenue Operations Leader

• University of Chicago: AI Application Developer

• Teradata: Principal Engineer, Agentic AI

GAMES

A QUICK POLL BEFORE YOU GO

PRODUCT

Microsoft tests the good-enough AI thesis

Microsoft Excel and Outlook can be among the most frustrating apps to use for professionals. Many have turned to AI for help, but the models powering that help may be quietly changing.

On Tuesday, Microsoft announced it is replacing OpenAI and Anthropic models with its own model in products like Excel and Outlook, Bloomberg reported, citing anonymous sources. Tens of thousands of AI prompts in those apps are now handled each week by internally built MAI models, according to the report.

The motivation behind this is simple: cutting costs. Using in-house models is much less expensive than licensing them from third parties. Just last month, Microsoft AI Chief Mustafa Suleyman stated Microsoft's intention to reduce spending by slowly weaning off Anthropic models in favor of MAI models, which span reasoning, image, voice, and coding.

While the move to proprietary models will cut costs, it carries some risk. Working professionals and enterprises are supercharging the race to deploy AI, and they want the fastest and most effective models, such as the ones Anthropic has become known for in the enterprise.

Anthropic just released its report on how people are using Claude Cowork, highlighting how much knowledge workers are leaning on AI. 

The report found that, from a sample of 1.2 million anonymized and aggregated Claude Cowork sessions, a third of Claude's Cowork sessions involve business operations (33.4%), with the next top use cases including content creation and copywriting (16.4%), software development (8.7%), and DevOps and infrastructure (7.0%). 

Microsoft's move to its own proprietary models isn't the first cost-cutting move this week. It eliminated 4,800 jobs (2.1% of its workforce) on Monday. While there is no denying that most companies are feeling the squeeze of AI due to expenses related to building, running, and serving AI services, Amy Coleman, EVP and Chief People Officer, made it clear in communications with employees that "the roles eliminated today are not being replaced by AI." However, the company is shifting resources to expand its AI division. 

If the MAI models prove highly efficient for users and less costly for Microsoft to run, they could help the company pull audiences away from Anthropic and other enterprise competitors. Microsoft already has a built-in edge here: decades of Office 365 dominance mean it has users locked into its walled garden, and those users trust Microsoft in a way that's hard to replicate. Even if Microsoft's models are only 80% to 90% as good as the latest frontier models from Anthropic, OpenAI, and Google, many enterprises may be willing to make the trade-off if the models are less expensive and viewed as being safer. 

TOGETHER WITH GENERAL ASSEMBLY, AN LHH BRAND

There Are AI Leaders, And There Are AI Followers

Which one do you want to be? If your answer is leader (and it probably should be) then you need to check out the latest from General Assembly. Their “AI For Leaders” program has everything current – and future – executives and decision makers need to build for the future. You’ll learn how to drive real AI transformation across your organization, apply AI skills in real business context, and shift your company culture to one of positive AI adoption.

A new era of leadership requires a new approach. See how General Assembly can set you on the path to success right here.

GOVERNANCE

AI adoption is outrunning enterprise security

Companies are deploying AI faster than they can secure it.

Nearly eight in ten enterprises reported encountering AI-related security issues over the past six months, according to new research from security firm DigiCert, which surveyed 1,001 IT and cybersecurity leaders across the US, UK, and Australia.

Approximately 50% of respondents experienced at least one security incident caused by an unauthorized or misconfigured AI agent, while another 28% identified AI-related vulnerabilities without experiencing an incident, a spokesperson from DigiCert told The Deep View. Science and technology organizations reported the highest rate of AI-related incidents or vulnerabilities, followed by banking, financial services and insurance, telecommunications and media, and retail.

While companies weren't asked to identify specific incidents they experienced, Brian Trzupek, DigiCert’s senior vice president of product, said that unauthorized or misconfigured AI agents can pose a range of security risks, including prompt injection attacks, data poisoning, unauthorized access to sensitive systems, and a lack of traceability.

“AI agents are non-human actors operating at machine speed, and most enterprises haven't extended the identity, authentication, and audit controls they already require of every user, device, and application to cover them,” Trzupek told The Deep View.

The risks are outpacing companies' ability to manage them. While 90% of organizations have discussed AI governance, only half have dedicated budgets and formal governance programs in place, according to the study. Nearly half also lack full visibility into how AI systems arrive at their outputs, making it difficult to trace which model produced a decision or investigate what went wrong.

The potential for unintended consequences is becoming more urgent as companies accelerate AI adoption. Over the past six months, 75% of organizations deployed four or more AI-powered systems, while 35% deployed more than 10. As AI becomes embedded across business operations, every new model, agent, and integration introduces new security risks. Without stronger governance, companies could face legal, regulatory, operational, and reputational repercussions.

The report argues that enterprises need stronger identity and governance controls, including the ability to verify AI identities, audit AI decisions, and trace outputs back to the underlying models and data. Companies appear to be moving in that direction, with 86% reporting that they have formal or informal processes for revoking access to compromised AI systems.

“Without governance, organizations lose the ability to answer basic questions: what AI is running, what it can access, and who is accountable when it fails,” Trzupek said.

Enterprises adopting AI have stepped into uncharted territory. Before AI, companies were already dealing with a slew of vulnerabilities and cyberattacks. Now, AI has only added to the challenges as the threat landscape evolves. Companies must not only prevent their AI systems from leaking proprietary data, but also protect them from exploitation by bad actors. And that only scratches the surface. As AI advances, threat actors are becoming more sophisticated, potentially opening the door to new attacks that haven't yet been discovered. If companies are serious about baking AI into their workflows, security and governance will have to become a higher priority.

TOGETHER WITH MERCURY

Banking* that does financial work for you

You've seen what AI can do when it's built into the right system. Mercury Command is that for your finances.

Built directly into your Mercury account, Command turns plain language into executed financial work — categorize transactions, adjust a card limit, send an invoice — with full account context, explicit approvals, and a traceable record behind every action. 

Try Mercury Command →

HARDWARE

Why ex-Tesla scientist bet on robotics in Europe

Tesla staked its claim on humanoid robots with Optimus before AI breathed new life into the market. Now, one of the scientists who helped build it is working on a rival.

On Tuesday, UMA, the physical AI startup founded in 2025, unveiled the design of its first humanoid robot. Remi Cadène, CEO and co-founder of UMA and former staff scientist at Tesla, sees humanoid robots as the key to unlocking greater productive capacity.

 "We believe intelligent robots will become part of the solution, not as a substitute for people, but as a new class of tools that enables them to devote more time to what machines will never replace: creativity, judgment, innovation, and caring for others," said Cadène in the release.

UMA is developing two systems: a dual-arm mobile industrial robot that could operate in warehouses or on assembly lines, and a compact humanoid robot designed to navigate human-centric spaces and collaborate directly with people. 

Twitter tweet

While many companies are currently tackling physical AI and have built humanoid robots, UMA's unique appeal is its use of the Real-Time Learning architecture, in which robots can learn through demonstration rather than manual programming. 

According to the company, this is meant not only to help them learn more efficiently, directly in the physical world as a human would, but also to improve their performance over time. Another distinctive characteristic is that most of the robotics buildout is happening either in the US or China, while UMA is building physical AI in Europe.

The idea of physical AI is simple: give AI access to your actual environment so it can take action and learn tangibly and firsthand. Though humanoid robots may have first been made mainstream by Tesla with Optimus as a futuristic idea, they now have buy-in from big companies and leaders, as AI has broadened humanoid robots' applications and ability to learn. UMA is a prime example, with a team comprising experts with experience at NYU, Google DeepMind, and LeRobot, as well as backing from leading AI figures and investors, including Yann LeCun and Olivier Pommel.

Europe has been falling behind in AI and robotics development compared to the US and China. However, one unique advantage the region has is actual legislation in place to standardize AI operations, making companies more likely to trust its use. For robotics specifically, there's a lot of public mistrust, since it not only looks like something out of a sci-fi movie but also involves cameras and cohabitating with humans in a way that can inherently compromise privacy. As a result, a regulated robotics industry in Europe could be exactly what it takes to bring the technology mainstream, or at least give Europe a much-needed early advantage.

LINKS

• Chinese AI labs could restrict access to their most powerful models 

• DeepSeek is reportedly developing its own AI chip 

• AI law startup Norm raised $120 million in Series C funding round

• AI actor Tilly Norwood is set to make big-screen debut in feature film

• Research shows people trust AI faces more than those of real people

• Samsung stock soars fueled by AI chip business 

• AI labs hand out free tokens to captivate startup market

• Claude Fable 5: Anthropic extends access to Fable 5 through Jul 12 on paid plans

• Google Managed Agents: Get support for remote MCP and background tasks

• Muse Image and Muse Video: first media generation models by Meta Superintelligence Labs

• SIMBA 3.2: Speechify's newest model topped the Artificial Analysis Speech leaderboard 

• Grok: The Grok API has 21 new flagship voices. 

• OpenAI GPT-Realtime-2.1-mini: Now available in the API 

• Google Finance: The recently released Android app has AI research tool and features 

• Coreweave: Staff AI Security Engineer

• Paramount: Sr. Director, Technical Program Management AI/ML

• Transformworks: AI deployment engineer

• Anthropic: Account Executive, AI Native

GAMES

POLL RESULTS

Do you think social media platforms should use AI to combat bots?

Yes (75.37%)
No (14.92%)
Other (9.71%)

ENTERPRISE

Why Anthropic bet on adult supervision for tokens

Anthropic just launched adult-in-the-room controls for CFOs and CIOs freaking out about AI costs.

Claude Enterprise has rolled out a trio of new features to help both organizations and employees track, monitor, and control token usage, ROI, and budget utilization. Since Anthropic has established itself as the preferred AI platform among US enterprises due to its focus on safety and privacy, this is a big deal. 

As a result of executives pressuring employees to adopt AI, tokenmaxxing dominated the first half of 2026, but the backlash has clearly begun as AI inference costs have run out of control, and organizations are now moving to rein in token spending. 

Uber was one of those companies that incentivized employees to use AI by launching internal leaderboards that ranked workers by how many AI tokens they used. And then the company reportedly burned through its annual AI budget in the first four months of the year. I've heard other CTOs and executives report similar challenges.

"Every organization is super worried about the costs going through the roof," said Databricks CEO Ali Ghodsi last month at the Databricks Data + AI Summit. "It's completely unsustainable for the organizations out there… People's hair is on fire." 

And of course, a lot of these organizations are using Claude. New Ramp data from July 6 shows that 77% of organizations that have adopted frontier LLMs use Anthropic models, up 40 percentage points in the past year. 

So it's not surprising that Anthropic is working to help organizations normalize their spending. The new features in Claude Enterprise to assist with this are: 

• Spend alerts: Admins are now notified when the org reaches 75% and 90% of its usage limit, so they can raise the cap before everything gets locked down and work stops. Employees receive notifications at 75% and 95% of their personal limit and can request an increase from their manager from directly within the Claude app.

• Costs vs. outputs: As part of the richer analytics in the platform, admins can now view what was produced by Claude alongside what it cost in order to do more granular ROI analysis.

• Model defaults: Admins can now set the default model that new conversations and tasks start with for Claude, Claude Code, and Claude Cowork, so employees can start with less expensive models and only upgrade to more expensive ones when needed. This can be set based on role or across the entire org. 

IT professionals and leaders in organizations running Claude Enterprise and Claude Teams can access these new features in the analytics section inside settings in Claude. They can set the default models in the organizational settings.

No AI lab has benefited more than Anthropic from out-of-control AI inference and token costs so far in 2026. Anthropic's annual revenue run rate jumped from $9 billion at the end of 2025 to $47 billion last month. So it might seem counterintuitive that the company would want to do so much to help its customers control costs and spend less on Claude tokens. However, out-of-control inference costs have turned into such a crisis, as Ghodsi noted, that Anthropic needed to do something to maintain trust with its customers. In this case, it's taking the long view, working to be a good partner, and betting on its reputation of being the safe, reliable, enterprise-friendly AI lab. Establishing that reputation could be worth far more in the long run if AI lives up to its promises.

TOGETHER WITH ATTIO

The next era of revenue runs on Attio.

Get agents working on every account, surfacing new opportunities, and handle the work that used to take your team days. With Attio’s new workflows engine, you can build agents for any GTM play you need. Just describe the goal, and Attio assembles the workflow for you to approve.

Then Ask Attio any question, from the weekly forecast to performance by rep, and get the answer in seconds. So you keep every deal moving.

Trusted by some of the world’s most successful startups, like Granola, Modal, and Railway, Attio is the agentic CRM that transforms how revenue work gets done.

Try Attio for free

POLICY

How Illinois made AI safety harder to dodge

Illinois has put forth the latest attempt to regulate the rapid pace of model development. 

On Monday, Illinois Governor JB Pritzker signed SB 315, a bill aiming to hold AI companies accountable for the powerful models they develop. The bill was passed unanimously in the Illinois state Senate and House in May, and was supported by both OpenAI and Anthropic. 

"As AI systems become more powerful and the federal government is unwilling to step in, states have a responsibility to protect our people from the dangers of AI while still harnessing the unique potential of the technology," Pritzker said in a statement. 

The bill, which is based on similar laws passed in New York and California in 2025, applies only to large frontier AI developers that generate $500 million or more in revenue. The bill will go into effect on January 1, 2027.

Though it targets a narrow audience, it applies strong safety regulation to some of the most powerful entities in AI, including: 

• Requires these companies to publicly disclose their safety frameworks and take part in third-party evaluations of their safety practices

• Compels them to report their models' critical safety incidents, such as unauthorized theft, model compromise or user deception

• Assesses for catastrophic risks, such as loss of control, use to create weapons, or the ability to carry out large-scale cyberattacks 

• Noncompliance can incur financial penalties, including a $1 million fine for the first violation and $3 million for any subsequent violations

Cesar Fernandez, head of US state and local government relations at Anthropic, said in a statement that the bill represents an "important step" toward the accountability that AI demands, as it pairs both transparency requirements with independent evaluations. "Anthropic is proud to have been the first AI lab to support this bill." 

The bill comes as AI companies navigate a minefield of US government intervention in the release of their increasingly powerful models. After the rocky rollout of Anthropic's Mythos and Fable models, and OpenAI preemptively holding back its most recent model, GPT-5.6, the government's regulatory power over these companies' model releases is far from set in stone. 

However, these companies may be working with the government on a solution: Last week, The Financial Times reported that many major AI firms are in talks with government officials to create a voluntary set of standards for releasing advanced models.

What sets the Illinois bill apart from predecessors in California and New York is that it requires independent, third-party evaluation of model safety. Without an unbiased party to audit model behavior, users must place their trust in the internal evaluations and standards to which AI companies hold themselves. Additionally, SB 315 and similar bills go a step beyond any standards that the federal government could muster because they are required, rather than voluntary. While it's good that the federal government is trying to create a unified standard for model releases, these state laws go further by holding these companies legally accountable for any potential harm they cause.

TOGETHER WITH OUTSKILL

Make Claude Your 2nd Brain That works 24/7 By Mastering In It 16 hours - Here’s How

Claude is currently the most powerful tool of 2026. It's been launching new features every week- Skills, Connectors, Cowork, vibe coding. Yet almost no one knows how to actually use them.

Our expert mentors have condensed 800+ hours of Claude research, articles, YouTube content and real-world practice into a focused 16-hour curriculum. Join the 2-Day Claude AI Mastery Workshop: a live, end-to-end deep dive into Claude plus 10+ AI tools, LLMs and workflows.

It would be silly not to SIGN UP

The professionals who master Claude right now will design carousels in seconds, ship code without being developers, and have AI working for them while they sleep.

You will learn how to:

• Master Claude's three modes : Chat, Cowork and Code.

• Set up Skills, Connectors and Plug-ins to automate your desktop, Notion and files.

• Vibe code apps and dashboards without writing code & 10+ AI tools and workflows that pair with Claude.

- Saturday & Sunday
- 10 AM – 7 PM EST

Register NOW!

CONSUMER

Reddit turns to AI for content cleanup

AI gave the internet a new plague: endless, low-quality content known as "AI slop." Now Reddit is betting that the same technology can clean things up.

Reddit has been reinforcing its automated spam detection tools with AI — and is seeing results. On Monday, the company shared that it blocks 23 million spam views per day before they reach users, catching 25,000 net new "spammy posts and comments" daily, and revoking nearly 2 million inauthentic votes per day. From January to March, the company says users' exposure to spam dropped by 20 percent. 

Large language models are employed to detect suspicious behavior through "highly subtle, coordinated patterns" that older detectors may have missed, Reddit said. AI is also being used on the platform to curb hate and violence in English content, with the automated system detecting harmful content in under five seconds and with greater accuracy. 

These measures aim to protect the authentic, human-made, community-based content that Reddit is known for, at a time when AI bots and content are filtering through nearly every platform and community. Beyond creating an unpleasant experience for users, AI-generated content is easy to spot, lazy, and altogether unhelpful, and it also poses real threats of spreading misinformation. 

For instance, recent research from the University of Oxford and the University of Potsdam found that when used to write or create social media posts, LLMs altered the meaning of posts on contested topics, introducing bias even when explicitly told to preserve the original meaning. Taken together, the study found that these small changes could accumulate to affect overall public opinion.

"Users need to be aware that GenAI is not a mirror that faithfully reflects reality, but rather is something that gives us a skewed version of the truth," said Sandra Wachter, Professor of Technology and Regulation at the University of Oxford. "Further, we should talk about if new laws or guidelines are needed to govern these systems."

The most interesting part of the Oxford study was that the researchers recreated Grok's "explain this post" feature, which is intended as a positive tool that provides users with more context about what they're reading. However, the study found that the simulated system was more supportive of "pro-life" leaning posts than "pro-choice" when discussing abortion issues, which could steer collective opinion over time. This highlights not only the risk of using AI to solve a problem AI itself created, but also how much impact even the slightest bias can have in the long run. Reddit holds a unique position in this AI content machine. While it uses AI to combat spam on the platform, it also has several lucrative deals with AI companies that license its content to train AI models, potentially perpetuating the issue.

LINKS

• XAI rebrands to SpaceXAI, unveils new logo

• Nvidia delays release of next-generation AI server rack to 2028

• Anthropic signs $19B, 20-year TeraWulf data leases

• AI won't give people a four-day work week, NYT Opinion

• Microsoft cuts 4,800 roles due to restructuring

• Mercor hits over $2 billion in gross annualized revenue

• ChatGPT: OpenAI's flagship model is now available directly in PowerPoint. 

• Sarvam MCP Server: Gives coding agents a standard way to discover and build with Sarvam

• Tencent Hy3: A 295-billion-parameter Mixture-of-Experts model available under the Apache 2.0 license.

• Cadence: An AI screen recorder that works confidentially and accurately.

• Reflection: Member of Technical Staff - Safety

• Meta: Linguistic Engineer

• Anthropic: Security Software Engineer, Detection & Response Platform

• TikTok: Research Scientist, Intelligent Editing and AI Agent

GAMES

A QUICK POLL BEFORE YOU GO

WORKPLACE

AI adoption leads companies to hire more, not less

Despite fears of AI replacing workers, companies making the biggest bets on the technology are still hiring.

Companies that invested most heavily in AI grew their headcount by 10% over the past two years, according to a recent study from Ramp and Revelio Labs. The researchers analyzed AI spending and workforce data from more than 21,000 U.S. companies. Entry-level hiring rose by 12% among the heaviest AI adopters.

AI adoption was uneven across industries. Companies seeing the strongest headcount growth were larger, more engineering-intensive, and more likely to be venture-backed, particularly in information, finance and insurance, and professional and technical services. Adoption was far less common in industries such as healthcare, accommodation and food services, and arts and entertainment.

Employment growth was broad across job functions. High-intensity AI adopters saw statistically significant increases in engineering, sales, customer service, finance, and administrative headcount, suggesting that AI usage augments the workforce and increases overall economic activity. 

"If AI lowers the fixed cost of building software, handling administrative work, doing analysis, or improving customer support, the gains can drive outsized growth and unlock new revenue streams that previously required higher fixed costs in the form of new salaries," Ara Kharazian, Ramp's lead economist who worked on the study, wrote in a blog post. 

The findings challenge the narrative that AI adoption inevitably leads to fewer jobs. Headlines about companies citing AI in layoffs, along with warnings from AI leaders about widespread automation and the loss of entry-level jobs, have fueled fears about AI eating the labor market. This study paints a more complicated picture: among companies making the largest investments in AI, hiring continued to grow.

The researchers caution that the findings shouldn't be generalized across the broader economy. Instead, they offer an early look at how heavy AI adopters are changing their hiring patterns.

"We believe [employers] are selecting for a new set of skills, specifically, people who know how to use AI and use it well," Kharazian wrote.

These findings should be taken with a grain of salt. The study focuses primarily on larger, tech-forward companies and doesn't capture how small businesses are using AI or whether companies rely on free AI tools. It also shows a correlation between AI spending and hiring, not that AI directly caused companies to add workers. Even so, the study points to an important distinction: companies seeing the strongest growth don’t just slap AI onto their workflows. They invest heavily in it and integrate it across their organization. While it's still too early to know which jobs AI will ultimately replace, it's important to note that AI literacy is quickly becoming a baseline expectation for workers.

TOGETHER WITH CODER

The Next Big Test Facing AI Coding Agents

By now, we all know the standard watch-outs when it comes to AI agents. Reliability, security, privacy, and the like are all real concerns for any company… but for those of us working in regulated industries (think financial services or government), these “watch outs” can quickly become massive issues – and that’s no good. 

Coder recognizes the scale of this problem, which is why they commissioned this new report from Weave Intelligence to help us address it. Inside, you’ll find a complete breakdown of the requirements for AI coding agents in regulated industries, a practical framework for deploying these agents at scale while still maintaining security and compliance, and how to do it all while avoiding those dreaded operational bottlenecks. 

Download the whole report for free right here. Thanks, Coder!

STARTUPS

Why world models, not LLMs, will deliver physical AGI

The term AGI has been thrown around so much that no one really knows what it means anymore. But an increasing number of experts think the key lies beyond LLMs. 

In this episode of The Deep View Conversations, we sat down with Caroline Ingeborn, COO of Luma, an AI lab dedicated to omnimodal intelligence, to discuss where generalized physical AI could take us. 

Rather than focusing only on video models, Luma takes an omnimodal, or multimodal, approach, creating models that understand text, video, images and audio. This, said Ingeborn, is because humans don't think in one modality. 

These kinds of models have many potential use cases and could even help researchers achieve general intelligence. Luma's primary audience right now is the creative industries, such as entertainment, advertising and marketing. Ingeborn said she sees the technology as enhancing the creative experience rather than replacing creative professionals. 

Topics covered include:

• Where AI fits into creative workflows

• The ethical lines of AI in creativity 

• Luma's mission towards generalized physical intelligence

• Physical AI's potential impact on the labor market

• The different approaches to building world models

• The dangers of centralized power in physical AGI

If you're following the progress of world models, physical AI and robotics, and the use of AI in creative fields, then this episode offers a deeper look at how these technologies are being used today and the transformative impact they could have in the future. 

📺 Watch on YouTube: https://youtu.be/ISueKJW-uqg

🎧 Listen in your favorite podcast player: transistor.fm/shows/tdv/episodes/53-how-to-rethink-agi-through-multimodal-systems-caroline-ingeborn

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm

TOGETHER WITH CONVEX

Code Faster, Easier, And Less Buggy With Convex

You’re probably already familiar with Convex, the backend building platform that helps your AI agents excel – but their latest updates are taking things up a notch.

Convex recently launched plugins that allow you to connect Claude Code, Codex, GitHub Copilot and more directly to your Convex deployment, allowing them to…

• Read data and run functions entirely on their own

• Optimize your project’s performance via production insights

• Improve safety by limiting access to production

There’s even more updates coming in the near future, but for now, see how you can take your agents’ work to the next level with the latest from Convex right here.

RESEARCH

How AI is rewiring scientific discovery at Google

AI's transformative effect on software development is well known. Next up for transformation: scientific research.

Lizzie Dorfman leads science AI at Google Research, where her team develops AI systems to help scientists solve problems across genomics, neuroscience, epidemiology and climate science.

In a conversation with The Deep View at Google I/O in May, Dorfman explained how Google's own researchers have adopted AI agents, why they can now explore hundreds of thousands of scientific ideas instead of just a handful, and why the biggest breakthroughs often come from solving smaller bottlenecks along the way. 

This interview has been edited for brevity and clarity.

Jason Hiner: What's top of mind for you right now?

Lizzie Dorfman: We've done AI and science for a decade... But Gemini was a real sea change. What came out of three years of eating our own dog food is what we call ERA: Empirical Research Assistants. And it's shockingly powerful. But we're at a point where basically 100% of our team, across all these domains, this is how we do our work. We watch them go through a not-terribly-long tunnel from "okay, fine, I've heard you talking about this" to "oh, wow." I heard from someone I considered fairly grumpy and quite serious, who emailed to say he'd just gotten a transformational result he was almost done writing up. And we're mostly pretty grumpy, pessimistic people who say we'll believe it when we see it.

Jason Hiner: How does that actually change the work?

Lizzie Dorfman: For us it ended up being coding agents using a tree-search methodology, where you explore hundreds or thousands of different approaches to solving a problem computationally. That's what gets you extraordinarily creative and performant solutions. We have an epidemiological forecasting example with top-scoring models in the CDC competitions. We generated 200,000 models to evaluate. Most were poor and got shed immediately. On a traditional team, someone might score a couple of ideas. People describe it as: I input some ideas, went to sleep, and woke up with all these cool results. Before, it was serial and iterative... Now it's possible to explore outlandish ideas just to see, because it's trivially more expensive to you, certainly in terms of your intellectual time.

Jason Hiner: So does this reduce the need for human expertise?

Lizzie Dorfman: No. I love this example: solar panels are flat, but what if they weren't?... Someone wrote a paper on curved panels in 2012... We fed it in, reproduced their results, then asked the system to make it better. It did, and voila. But when we looked, it had these photovoltaic pieces that were levitating, not physically connected. It was cheating because you can maximize energy if you don't have to adhere to physics. So we added a loop that checks the solution is physically valid. You can't just shut your eyes, and it's done... Expertise and careful verification are still required.

Jason Hiner: When you take on grand challenges, how do breakthroughs actually come?

Lizzie Dorfman: There's a phase where it's "this is impossible, it's infeasible." ... For brain mapping, we need to bring the computational cost down by multiple orders of magnitude. The breakthrough takes a lot of forms. It's not always something that looks like AlphaFold... It's frequently "oh, wow, another order of magnitude." One big breakthrough is often a series of obstacles overcome. In the last four years, we've had something like 45 papers in Nature and Science. That's results. It's not "hey, blog post, we did a thing."

Read the rest of the interview with Google Research's Lizzie Dorfman.

LINKS

• US in talks with AI labs to create voluntary standards for new models

• Microsoft invests $2.5 billion in embedding AI engineers in customer orgs

• OpenAI proposes giving Trump Administration 5% stake in the company

• Germany's SAP bets on reinventing jobs with AI, not cutting them

• SpaceX has reportedly previewed AI device to investors

• Why AI is spurring talks about Universal Basic Capital

• Weave Robotics launches home robot Isaac 1, deliveries start in fall

• Some high-earners are using AI schools to create custom education for kids

• ZCode: China's Z.ai has launched a coding competitor to Cursor and Claude Code. 

• Notion HTML: Users can now build interactive HTML on their Notion pages using AI. 

• S2.1 Pro: Fish Audio's best voice model is now free for developers. 

• Nemotron-Labs-TwoTower: A diffusion language model from NVIDIA adapted to write tokens in parallel instead of one at a time.

• OpenAI: Agent Post-Training, Context Research

• Fractional AI: Developer Relations Engineer

• Capital One: Applied Researcher I

• Salesforce: Senior Offensive Security Engineer (Red Team)

GAMES

POLL RESULTS

Are you concerned about government intervention in the release of powerful AI models?

Yes (37%)
Somewhat (30%)
No (28%)
Other (5%)

Large language models can do a lot of things, as long as those things exist on a screen. 

Having consumed practically all of the data on the internet, these models can tell you something about practically anything. They can analyze thousands of documents and pull out the most important parts. They can plan your trips, write poems, and mimic therapists in times of need. They can help researchers understand anything from protein structures to ancient history. And they’re the backbone of millions of agents that are the hottest ticket in tech right now. 

But the world is bigger than a screen, and understanding it requires more than words. 

That's why Nvidia’s Jensen Huang has said more than once that physical AI is due for its "ChatGPT moment." 

It’s also the reason that world models, or AI models capable of understanding the physical environment, have gained significant momentum in 2026. Along with a flock of young startups entering the space, some of AI’s most prominent figures have homed in on the concept. 

But as the momentum around world models grows, in tandem with physical AI and robotics, some leaders have begun to question their impact on LLMs, and ultimately, the ever-elusive path to artificial general intelligence (AGI). 

"Seeing the world in a profound way, in a way that you participate in your movement, your interaction, in your communication, is critical for intelligence," said Dr. Fei-Fei Li, largely considered the godmother of AI, while also being the founder and CEO of World Labs, in a panel at the HumanX conference in April. "Not having that is intelligence in the dark."

Investors take notice

As world models catch the attention of some of AI’s biggest tastemakers, investors have become captivated. World model startups have been raking in billions in funding, some at incredibly early stages: 

• Advanced Machine Intelligence, or AMI Labs, a startup founded by AI godfather and former Meta AI chief Yann LeCun, raised a $1 billion seed funding round at a $3.5 billion pre-money valuation in March, after launching in late 2025. 

• World Labs, Li’s startup founded in 2024, announced its own $1 billion funding round in February at a valuation of $5 billion. 

• Runway, a company focused on scaling video models as a means of creating world models, raised $315 million in Series E funding at a valuation of $5.3 billion.

• Luma, an AI video startup that focuses on creating multimodal reasoning models, raised a $900 million Series C in November at a valuation of $4 billion. 

• General Intuition, a lab dedicated to building foundation models for environments that require "deep spatial and temporal reasoning," raised $320 million in Series A funding at a $2.3 billion valuation in June. 

Beyond startups, several companies have pivoted into world models from one industry in particular: gaming.

 Niantic, the maker of the beloved Pokémon Go app, sold its suite of mobile games to Scopely for $3.5 billion and spun out a lab last March called Niantic Spatial, focused on developing what it calls a "Large Geospatial Model," a world model that enhances spatial reasoning in LLMs. 

And Roblox, the online gaming platform with more than 150 million daily active users, is developing its own version of a world model that it calls "real-time dreaming," that allows creators to generate and iterate on virtual environments through language prompts. In a panel at the HumanX conference in April, David Baszucki, Roblox founder and CEO, said that he envisions the company’s world models "not just as a play technology, but as a creation technology as well." 

In January, Google DeepMind released Project Genie, an "experimental research prototype" that marks the latest iteration of its work in the world model space. The project is powered by its flagship Gemini model, its Nano Banana Pro image model, and Genie 3, its most powerful world model yet. 

Nvidia, meanwhile, unveiled its Cosmos model at CES 2025, a world foundation model that’s aimed at accelerating the development and deployment of autonomous vehicles and robots. Since then, the company has expanded Cosmos further, including debuting the third generation of the Cosmos family and releasing world-generation models, controllable simulations for synthetic data generation, and multimodal reasoning models for physical AI. 

Many bets, same problems

The challenge with the term "world model" is that it means different things to different people. 

During her HumanX panel, Li said that defining world models in 2024 was much easier. "No one was talking about this," she said. "I'm keenly aware that different groups have different definitions of world models at this point."

Broadly speaking, building a world model involves creating machines with the ability to understand space, she explained, such as reasoning about geometry, interactivity and physics, as well as generating those spaces "just like our computers today generate words." 

And as Li said, there are several different ideas about how to achieve this. 

For starters, there are predictive video models that learn to forecast future states from video data. This is the method that Nvidia’s Cosmos models and Google’s Genie models have chosen. Anastasis Germanidis, co-founder and CTO of Runway, an AI video company, told The Deep View that it is also staking its world model on "interactive video," which is how it built GWM-1.

“We've seen that the most important thing for making great world models is having great video models,” Germanidis said. 

World Labs, Li’s company, founded in 2024, focuses on spatial intelligence. Rather than homing in on video generation, World Labs is focused on 3D world representation as a spatial object, not a sequence of frames. Marble, its first world model, aims to create geometrically-accurate, consistent and navigable environments that you can interact with, edit and reuse.

Marble

Simulation models, such as robotics, gaming, or autonomous driving simulators, also fall into the world models bucket, simulating environments that agents can act within and learn from. For instance, this kind of model is the backbone of Tesla’s self-driving technology and the Waymo World Model. Robotics models, such as vision-language-action models like Figure AI’s Helix, can also fall into this category. 

Multimodal models can also fall into the world models category. That includes the approach taken by Luma, an AI video startup focused on creating "multimodal general intelligence," or models that can understand the world by taking in a wide variety of data, such as images, videos, audio and text. 

"Our bet on world models is that you train models jointly on all of these modalities, just the same way as our brains work, rather than like saying, '3D is the path to world models,' or 'video is the path to world models,'" Caroline Ingeborn, COO of Luma, told The Deep View. "If you're going to have models that understand and can generate and operate in our worlds, you need to be trained across modalities."

Despite the different approaches to creating these models, they all face the same two constraints that plague everyone building AI: compute and data. But because these models digest and create more than simple words, the problems are amplified. 

For one, these models are far more compute intensive than their language-only predecessors. And amid an unprecedented crunch for compute, the entire industry is feeling the squeeze. That's why OpenAI completely axed its own AI video platform, Sora, to free up more compute for its flagship models. Li even quipped about this problem in her HumanX keynote, saying that her "friends in Nvidia and AMD will be very happy" following her company’s most recent funding announcement, implying that World Labs would be buying a lot more compute. 

As for data, these models require more than just what can be scraped from large repositories of text. Similar to the data shortages in the robotics industry, these models need large amounts of high-quality data from real-world environments to learn to recreate them. Getting these models to behave consistently and create physically accurate data largely relies on this. 

“There's never enough compute, and there's never enough data,” said Germanidis.

Today's primary use of world models

These various world models may be able to solve each other's problems. As it stands, the most commercially significant use case for these models is generating synthetic data, which is then used to train robotics or self-driving vehicles. 

Deepu Talla, VP of robotics and edge AI at Nvidia, told The Deep View that these kinds of models are critical to filling the data gap plaguing physical AI development and are the primary use case for Nvidia’s Cosmos models. Though many developers are on the quest to build a "general-purpose robotic brain," Talla said that the brain itself may not be enough to create robots that can seamlessly act on the world around them. 

"A robot can only do so much if you don't have the environment for the robot to interact with," said Talla. "The world model is about creating the whole environment around it to do that simulation." 

Li echoed a similar sentiment for the work of World Labs, noting in her keynote that the output "contributes to further providing synthetic data for others in the ecosystem." 

This application is vitally important because it can create a symbiotic relationship between physical AI systems and the world models used to build them. The better physical AI systems perform, the more they can collect real-world data to help train the models. 

World models’ synthetic data also helps break down a major barrier that physical AI systems face: safety. While a self-driving data set may have a sizable pool of info on typical driving situations, "the vast, vast majority" is going to be non-accidents, Germanidis said. However, "The place where their models need to perform the best —  that it's critical to perform the best — is exactly in those [dangerous] moments."

But an AI model that can comprehend the world beyond words and has an understanding of space and physics has practically limitless use cases, said Ingeborn. Physically accurate AI models can be transformative for the entertainment industry, such as filmmaking, VFX and editing, she said. These models may also be groundbreaking for the gaming industry, allowing developers to build game environments without requiring the resources of a large publisher. 

And beyond entertainment, these models have potential in science as well. In her 2025 essay, “Spatial Intelligence is AI’s Next Frontier,” Li notes that an understanding of the physical world is crucial for materials discovery and climate science by integrating multi-dimensional simulation with real-world data collection. These models also have the potential to accelerate drug discovery in medicine and act as companions in healthcare settings. 

"If you're doing an open heart transplant, you need to understand the physical world," Ingeborn  said. "Imagine having an AI that can understand and operate alongside you."

The AGI challenge

But like most things in AI, all roads lead back to artificial general intelligence. Many in the world model space firmly believe that spatial understanding is essential to achieving the nebulous and lofty goal of human-level machine intelligence. 

"World models are kind of the only path to AGI in my mind," said Germanidis. "There is no way to get to general intelligence unless you understand the real world."

While there are many different ways to define artificial general intelligence — as well as fiercely debated questions about whether achieving it is possible or if it’s a worthwhile goal to begin with — it’s also the core mission that underlies practically every major frontier AI lab.  

Let’s start with Anthropic. The AI giant has largely focused on large language models as the gateway to achieving massively powerful AI. Though CEO Dario Amodei has shunned the term AGI, in his renowned essay “Machines of Loving Grace,” he has defined powerful AI as that with skill levels "exceeding that of the most capable humans in the world," which can autonomously complete tasks and perform complex science and engineering. To achieve this definition, Amodei and Anthropic have bet on the paradigm of scaling laws: More compute and more data will, someday, lead to a language model with these capabilities. 

OpenAI holds a similar view. More comfortable with the term AGI itself, the company has pledged itself to the long-term mission of creating general intelligence, or "highly autonomous systems that outperform humans at most economically valuable work," that benefit "all of humanity." CEO Sam Altman is bullish on the idea that this will arrive sooner than later, with scaling laws being the fundamental method of getting there. 

But many in the industry have started to question whether language models are the key to unlocking this higher level of intelligence. Since the dawn of the concept of world models, experts have started to poke holes in the idea that language alone will get us there. One such voice is LeCun. 

On multiple occasions, LeCun has argued that language models will not lead us to human-level machine intelligence, as language offers only a limited slice of intelligence, restricted by the collections of letters and numbers humanity has written down. Without an understanding of environments, perception and interaction, these models are essentially untethered from reality, he asserts.

Instagram post

And the argument has its merits. Evolutionarily, humans developed the ability to communicate as a means of comprehending our other five senses. Without an understanding of those senses, or an understanding of physical consequence, how can a machine understand what it is to be human? 

"With this whole idea of AGI and a more general intelligence, it has become more clear that you cannot just get that from text,” Amit Goel, Nvidia's head of robotics and edge computing, told The Deep View. "For that level of intelligence, you need to consume information from the world. That is how you get to the next level."

Where we go next

As they pursue the vision of the ever-powerful, superhuman language model, between OpenAI and Anthropic alone, the companies have garnered a total of nearly $330 billion in funding and counting, each with valuations nearly scraping the trillion-dollar mark. 

But the question remains: With the growing momentum toward world models and  spatial intelligence, what is left for LLMs? 

Germanidis said there is increasingly a recognition in the industry that language model developers are "saturating the data we have available." As a result of approaching that limit, he believes that language model improvement will soon, too, hit a saturation point, making these models unable to solve "core problems that we care about."

While all of the text available in human history would take humans thousands of years to read through, that data is still just a small fraction of the real-world data that humans perceive with our senses over time. 

“There is this broad set of data that's about the real world that actually we haven't trained on, and we haven't built systems around,” said Germanidis. “It's the biggest next opportunity for AI.” 

However, what these models are capable of and how we actually use them are two different things. The issue of capability overhang, or the gap between the capabilities AI can offer and the capacity to which we actually use them, still looms among businesses and consumers alike. And as it stands, a relatively small fraction of people actually utilize AI on a day-to-day basis — just over 16%, according to a Microsoft study. 

The fact that such a small part of the population is leveraging language models to their fullest extent might be proof that, even if language models run out of words to eat and reach a capability plateau, there may be a long, long way to go before we run out of ways to leverage them, said Ingeborn. 

Luis Lastras, director of language technologies at IBM, told The Deep View that enterprises are still figuring out how to leverage language models in reliable, trustworthy and efficient ways. 

While physical AI and world models present a lot of promising applications, "when it comes to productive use of enterprises, I would say there's a lot that we can build on the current platform that is not really fully fleshed out, especially when it comes down to efficient automation," Lastras said.

Additionally, Nvidia’s Goel said these models may work hand-in-hand. Rather than being pitted against each other, languages are simply vectors used as the communication layer for world models and spatial intelligence. "I see it as a continuum," said Goel. "This new wave is building on top of LLMs."

The reality is that physically accurate world models and spatial intelligence have a long, long road ahead before they actually realize the lofty goals of Li, LeCun, and others. Along with confronting the same barriers of data and compute that frontier language model labs are facing, these models are simply more difficult to build than those that are language-only. 

It’s hard to create a model that holds an internal representation of our world, one that consistently understands and applies physics in its generation. Just like it’s exceedingly difficult to create a model that transcends the smartest humans in the world in every conceivable domain, the goal of AGI, no matter how close AI figureheads claim we are.

By the time highly capable world models are ready to become a critical component of the technology ecosystem, language models may have already run their natural course. And the grand AGI question will have to sort itself out along the way.

If you want to get in front of an audience of 750,000+ developers, business leaders and tech enthusiasts, get in touch with us here.

POLICY

Mythos returns, but AI may be permanently changed

Mythos and Fable are back. 

The US Commerce Department officially lifted its export controls on both Anthropic models on Tuesday, clearing the way for Claude users to regain access to the frontier lab's most powerful AI. 

Twitter tweet

In a letter to Anthropic seen by Reuters, Howard Lutnick, secretary of the US Commerce Department, said, "Anthropic has agreed to proactively detect and address security risks associated with the models; to work diligently with the US government on protocols and standards and releases for ​Mythos, Fable, and ​future models; and ⁠to inform the US government of any malicious activity."

The two models have been restricted since Jun 12, when the US imposed export controls that made it illegal for all foreign nationals, outside or inside the US, to access the technology. 

Anthropic has been warning about the potential cybersecurity implications of Mythos since early April, when it released it to a limited set of companies to help them prepare, as part of an initiative called Project Glasswing. 

US agencies were particularly concerned about Mythos being co-opted for military purposes by countries such as Iran, North Korea, Russia, China, and others. The speed at which Mythos uncovers software vulnerabilities and can potentially arm bad actors for cyberattacks was at the heart of the concern.

Washington hasn't only picked on Anthropic. It also kept OpenAI from broadly releasing its GPT-5.6 model, leading OpenAI CEO Sam Altman to call it "bad news" on X. Altman said, "At the request of the US government, it is launching today in limited preview instead of the open access launch we were planning." That was despite the fact that GPT-5.6 has tested even stronger than Mythos in some benchmarks. There's no word yet on when the government will clear GPT-5.6 for general use. 

But Washington's intervention with Mythos and GPT-5.6 may have set a new precedent for safety, audits, and oversight.

"If AI is going to be as impactful as all the major AI CEOs and some countries are saying, it should have the same audits and the same safety that other powerful technologies have," Chris Canal, CEO of AI eval vendor Equistamp, told The Deep View. "It deserves the same audits and safety standards as nuclear technology, [for example]."

Now the challenge is setting up a framework for pre-release model review, because no one wants a situation where AI labs release a model, the government intervenes, and then it gets re-released weeks later.

Perhaps the most notable outcome of the US government's standoff with Anthropic — and OpenAI now as well — is that it has resulted in a pause in promulgating these ever-powerful models. Such a pause to evaluate and prepare for the effects of the latest models had appeared highly unlikely. While leaders such as Anthropic CEO Dario Amodei have extolled the benefits of such a pause to assess safety, they've also made it clear they wouldn't do it unless other labs agree to a similar pause, because otherwise it wouldn't be effective. Will the staggered rollouts of Mythos and GPT-5.5 reset the expectations and norms for the way the most powerful frontier models are launched? If you are concerned about AI safety, that would be a win. The question would then shift to the who and the how of such a pause. 

TOGETHER WITH KALSHI

Trade The World Cup on Kalshi

The World Cup is the biggest sporting event on Earth.

Kalshi is a CFTC-regulated exchange where you can take a position on who lifts the trophy, backed by real-time market pricing.

Right now, Spain is priced at 10.7% to win it all. A $100 trade on Spain pays out $935 if they go all the way.

Trade $10, Claim $10

WORKFORCE

Forward Deployed Engineers come with a catch

Amazon is all in on the trendiest new job in AI: forward deployed engineers (FDEs).

On Wednesday, Amazon Web Services announced it is investing $1 billion to embed forward deployed engineers within its customer network. This network of engineers will help customers develop and deploy agents into their organizations, aiming to help them become "self-sufficient with AI" in days, the company said in its announcement. 

"Customers have moved past exploring what AI can do; they want to make it core to how they operate," Francessca Vasquez, VP of frontier AI engineering and services at AWS wrote in a blog post. "I have also heard loud and clear that many customers need expert AI engineers working directly with their teams to help them build and become AI-native organizations." 

As part of this new Amazon organization, engineers will work directly with a customer’s business, engineering, and security teams to help them build systems that integrate with their data and governance structures. 

• Rather than treating these deployments as "standalone projects," Vasquez said, the goal is to help these customers create lasting AI strategies, deployments and skills to "innovate independently." These agentic systems will run within a customer's existing AWS environment. 

• After the engineer is done, the organization is left not only with deployed agents, but with knowledge graphs, runbooks, architectural documentation and "trained internal champions" to keep the momentum going. The idea is for every new engagement with these AI systems to compound its knowledge and make it more useful. 

Amazon's investment is the latest indication that tech companies are betting on forward deployed engineers to solve the ROI problem in enterprise AI. This role, which essentially loans AI experts to a client's organization, has taken off in recent months, with job postings for forward deployed engineers up around 700% year over year, according to Business Insider. 

And Amazon's not the only one building its forces. In May, OpenAI launched the OpenAI Deployment Company, designed to help organizations build and deploy its technology using forward deployed engineers.

It makes sense that forward deployed engineers have taken off. AI is incredibly complex, and the barrier to entry can be daunting for many organizations. Forward deployed engineers can help remove those obstacles, and do so using the technology of whichever vendor they represent, creating a built-in user base. However, organizations that decide to take in forward deployed engineers have to remember that they come with a catch: These engineers represent the best interests of Amazon, OpenAI, or whatever other organization they come from. Working with them to integrate their tech into your systems can lead to unwanted lock-in. 

TOGETHER WITH LAMBDA

Cut your AI training costs by 25% or more

Most large-scale AI training runs use less than half the computing power they're paying for. Lambda's team found the root causes and built a reproducible framework that boosted efficiency by over 25%, without changing the model itself.

Lambda’s whitepaper shows you how to address:

• Memory inefficiencies silently inflating your costs

• Training configurations that aren't making full use of your hardware

• Bottlenecks that slow down GPU communication

Get the whitepaper.

HARDWARE

How Apple's decade-long bet on chips won AI

Walk into any of the frontier AI labs, and you'll find wall-to-wall Macs.

Decisions Apple made years ago, long before ChatGPT and the LLM era, are now paying big dividends.

In an exclusive conversation with The Deep View ahead of WWDC 2026 in June, Doug Brooks, senior product manager of Apple silicon, discussed how Apple's long-term investments in neural processing, unified memory, power-efficient computing, and hardware-software integration have positioned its devices for the AI era. 

This interview has been edited and curated for brevity and clarity.

Jason Hiner: Tell us where Apple silicon is today when it comes to AI.

Doug Brooks: We're seeing tremendous momentum with how people are using Apple products and Apple silicon, particularly in on-device AI workflows.... It really highlights the tent poles of Apple silicon. It's a very balanced architecture that provides CPU, GPU, unified memory, and the Neural Engine all contributing to performance across the chip. That's particularly important for these evolving agentic workflows.... It's not just about the GPU crunching on an LLM anymore. It's about the whole chip contributing to different parts of the task.

Jason Hiner: Apple was working on machine learning long before ChatGPT… What were the origins of that journey?

Doug Brooks: I think a lot of it starts with having really good foundational building blocks.... All of that was done with a relentless focus on power efficiency.... When we first introduced the Neural Engine, it was designed to solve very specific compute needs for machine learning.... So if you go back to the core of the question, the tent poles of Apple silicon, performance, efficiency, and unified memory, have played extremely well in the AI domain.

Jason Hiner: One of the trends we're seeing is developers wanting to run more AI locally because of privacy concerns, security requirements, and rising inference costs. Are you seeing that trend too?

Doug Brooks: Absolutely.... Why send your data, your code, or your intellectual property somewhere else if you can run it locally?... People are realizing they have incredibly powerful devices sitting right on their desks that are capable of handling a wide range of workloads.... I'm a big believer in using the right tool for the right job.... I think the hybrid approach is really interesting because agents can decide what needs to happen locally and what needs to happen in the cloud based on the workload.... We've talked publicly about the incredible demand for Mac minis and Mac Studios because they've become the platform for running so much of this [agentic] work.

Jason Hiner: How do you think about the future of on-device AI for iPhone and iPad?

Doug Brooks: Sometimes I think of it as transparent AI. There are so many AI features embedded throughout the operating system and throughout applications that don't necessarily jump out at you as AI. They're powered by AI under the hood, but they're just helping people get things done.... The work we've done with foundation models and exposing those capabilities to developers has allowed AI to become pervasive across applications in both small and large ways.

Read the rest of the interview with Apple's Doug Brooks.

LINKS

• Abu Dhabi's MGX raises $49 billion for AI-focused fund

• Palantir's Alex Karp calls military AI labs "effing insane" 

• Uber dismisses two leaders in AI data labeling business 

• Meta reportedly plans a cloud business for selling AI compute

• Open source AI firm Together AI raises $800 million

• ChatGPT: For Pro users, OpenAI has added new personal finance experience to securely connect financial accounts, see where their money is going, and query about your information. 

• Gemini Spark: Google's AI personal assistant is now available on MacOS for local tasks. 

• Google TabFM: Google has launched a zero-shot foundation model for tabular data. 

• xAI Voice Agent Builder: xAI has launched a no-code platform to create human-like voice agents with Grok Voice. 

• xAI: Member of Technical Staff - Inference

• Amazon: Applied Scientist, AWS Automated Reasoning

• Scale AI: Security Engineer, Product Security

• OpenAI: Agent Post-Training, Computer Use Research

GAMES

A QUICK POLL BEFORE YOU GO

PRODUCTS

Sonnet 5 is Anthropic's answer to AI sticker shock

Despite its latest kerfuffle with the US government over the release of its Mythos and Fable models, Anthropic isn't slowing down.

On Tuesday, the company unveiled Sonnet 5, the latest addition to its Claude model family and the "most agentic" version yet. The company said the model significantly narrows the performance gap between efficient and more powerful models, delivering capabilities on par with larger models at a lower cost. 

Sonnet 5 can make plans, use tools like browsers and terminals, and run autonomously in ways that would have previously required more expensive systems, Anthropic said. 

• The company noted Sonnet 5 provides "substantial improvement" over its predecessor, Sonnet 4.6, showing an overall lower rate of "undesirable behaviors" and is safer to use in agentic contexts. 

• Though it performs close to Opus 4.8, Anthropic's most powerful publicly available model, the company noted that it has a "much lower ability" to perform cybersecurity tasks than its Opus lineup. 

• Starting today, Sonnet 5 is available across all plans and is the default model for Free and Pro users. On Claude Code and Claude Platform, the introductory pricing is $2 per million input tokens and $10 per million output tokens, and will be bumped to $3 per million input tokens and $15 per million output tokens starting in August. 

Early testers include Cursor, Rakuten, Lovable and Factory, with testers broadly impressed with the model's ability to finish complex tasks, check its own outputs and complete agentic work at a good price. 

"With Claude Sonnet 5, agents stay on plan, follow our conventions and ship clean multi-step changes, all at an efficient cost," Sualeh Asif, co-founder of Cursor, said in the company's blog post. 

Anthropic's cost-friendly model comes at a time when the tokenmaxxing fad is catching up with enterprises. More than ever, companies are considering the untenable costs of AI, with Gartner predicting that AI spending will reach nearly $2.6 trillion this year, up 47% from the prior year. As a result, enterprises are experiencing AI sticker shock. Meanwhile, AI vendors have started to shift their focus to cost efficiency.

Model providers like OpenAI and Anthropic stood to benefit the most from the tokenmaxxing trend. However, with the industry largely recoiling from this mindset after getting hit with massive AI usage bills, model providers may be taking notice that the days of unlimited token gluttony are over. The efficiency-focused design of Sonnet 5 may be a sign that Anthropic wants to keep up with the changing tide. However, efficient and cost-effective AI might not be a bad thing for Anthropic, either. Customers burning through fewer tokens means Anthropic can save on compute, which is vital as demand rises and the shortage worsens. Plus, with Mythos and Fable still caught up in government red tape, the company likely wants to keep its audience hooked, and Sonnet 5 is more economically attractive to large customers. 

TOGETHER WITH FUNCTION

When it comes to health, your AI is better with Function.

ChatGPT, Claude, and other AI platforms are incredible tools. But when it comes to health, AI is running off generic population data instead of your unique biology. Not your hormone levels. Not your cortisol.

Function changes that. When you securely connect Function to your AI tools, you can bring your lab results and actual biology into the conversation.

Smarter inputs. Smarter responses.

Connect Function to your AI tools

CULTURE

Google races to define creative AI

Following the rising popularity of its Nano Banana image model, Google wants to keep the momentum rolling. 

On Tuesday, the company announced two new additions to its generative AI creative suite: the general availability of Nano Banana 2 Lite and the public preview of Gemini Omni Flash. 

The models, now available in the Gemini Enterprise Agent Platform, aim to give users more speed, control and cost efficiency in their creative work, the company said in its announcement.

Here's the breakdown:

• Nano Banana 2 Lite, or Gemini 3.1 Flash-Lite Image, is the fastest and most cost-efficient image generation and editing model in the Nano Banana family. The model can generate an image in as little as four seconds, and rapidly iterate design concepts or image edits. The model also features significant improvements in visual quality, including better world knowledge, character consistency, and localized text by region. 

• Gemini Omni Flash offers high-quality video generation and conversational editing, embedded directly into media tools and apps. The model also includes multimodal outputs, improved simulation and physics understanding, and improved text generation. The company said that support for audio references, video references, last frame, scene extension and higher resolutions will be available soon.

Google said that several companies are already testing both Nano Banana 2 Lite and Gemini Omni Flash, including Adobe, Klarna, Picsart, Genspark, Figma, Manus and others. Itay Schiff, Figma Weave co-founder and creative director, said in a statement that Nano Banana 2 Lite is "ideal for rapid iteration while staying in the creative flow." 

Nishant Tahilramani, creative director at Invideo, an AI video platform that tested the model, said in a statement that the "sheer range" of Gemini Omni Flash caught his attention. "The VFX capabilities surprised me, and looking at it as a producer, that brings in some very interesting possibilities … You take the crews you have always worked with in the live-action world, and you bring the breadth of what AI can do now onto the same set.”

The models are Google's latest foray into the creative fields. Last week, the company's AI unit, Google DeepMind, announced a partnership and a $75 million investment in A24, the film studio responsible for Backrooms, Hereditary and Midsommar, a decision that stoked the ire of fans and Hollywood professionals alike. 

"This partnership exists because we want to dictate what tools get built for artists, and so they have a voice in shaping them rather than having tools handed to them,"  Sophia Shin, an A24 spokesperson, told Wired.

As it stands, AI is yanking creative industries in two directions. There's a large contingent fighting against it for reasons of copyright and creative integrity, and there's also a group that believes the tech can democratize the arts and give more power to hobbyists. With these new models and its industry partnerships, Google is walking straight into the firestorm. The thing that could give it more legitimacy in the creative industries, however, is its SynthID, Google DeepMind's watermarking system for synthetic content. With AI-generated content growing more lifelike and realistic by the day, people will increasingly be unable to tell the difference between authentic and synthetic content. However, if Google's tools, whether it be the models themselves or its synthetic content watermarking platform, become an entertainment industry standard, the industry and the population at large may stand a better chance at distinguishing real from fake. 

TOGETHER WITH KALSHI

The First Regulated Perps Market in the US Is Live

Perpetual futures (perps) are now live on Kalshi. Offshore, perps do ~$90 trillion in volume annually — until now, none of it was legal in the US.

Trade on whether an asset's price goes up or down. No expirations, no rolling costs, no complex payoffs.

Try Kalshi Perps today.

RESEARCH

Anthropic eyes biology after winning at code

If Anthropic co-founder Dario Amodei is right, AI will eventually be known more for the scientific breakthroughs it enables than the negative societal risks it introduces. 

On Tuesday, Anthropic announced what it's calling its "AI workbench for scientists," now dubbed Claude Science. The app launches in beta today and runs locally on Mac or Linux, or on a remote machine via SSH or an HPC login. It runs on Pro, Max, Team, and Enterprise plans. Anthropic also launched a new team plan for research labs.

It's important to note that Claude Science uses all of the existing Anthropic models and does not introduce any new science-specific models like the one OpenAI has created with GPT-Rosalind. Instead, Claude Science is focused on a science-specific interface that brings together tools and capabilities that can accelerate the work of scientists, researchers, students, and labs. It also runs science-specific agents that are designed to automate some of the most tedious and time-consuming parts of research. 

Anthropic said that Claude Science can run locally on a laboratory's own servers and doesn't need to run through the standard cloud-based Anthropic models. This is a key compromise since labs often run highly proprietary, valuable, and sensitive data that demand strong security, privacy, and data sovereignty. 

Another one of the key issues Anthropic is trying to accomplish with Claude Science is the problem of scientific work being spread across a disconnected set of tools, databases, and computing environments, which creates friction and obstacles to achieving important breakthroughs. 

In the livestream, the Anthropic team made the case that AI has completely transformed the work of software development, and that it can have a similar impact on life sciences. And that's where Claude Science will start, with the goal of "compressing timelines in biology."

Amodei has long championed the potential of accelerating scientific progress as one of AI's greatest opportunities to make a positive impact on civilization. Tuesday's announcement was introduced in a rare major public event for Anthropic and livestreamed on YouTube. 

To stimulate use of its new science tools, Anthropic also launched a new AI for Science Program that will support up to 50 projects and provide $30,000 in Claude Science credits. This is aimed at nonprofits and academic institutions, which can apply online by July 15, with winners notified by July 31 and projects running from September 1 to December 1.

The Deep View has covered the launch of Google's Gemini for Science and OpenAI's GPT-Rosalind model recently and the level of enthusiasm for the impact that AI could have on scientific discovery has bordered on hyperbolic. Anthropic's Claude Science launch matched that exuberance. Clearly, there's a ton of enthusiasm from both AI labs and scientific teams for how AI can streamline some of the most entrenched problems that bog down scientific research, especially in biology. It's reasonable that AI could assist with streamlining research and documentation. It also makes sense that it could allow scientific teams to model out more hypotheses and experiments simultaneously, as the head of Google Research Yossi Matias recently told me in an interview for The Deep View. But we should also keep in mind that biology and medical trials take a long time, sometimes up to a decade, to test on human subjects. And while AI might be able to simulate it, it's unlikely that it will accelerate those trials. And so we should be wary about any claims of AI dramatically accelerating the elimination of human disease or solving the problems of aging in the next few years. 

LINKS

• OpenAI reportedly has discovered a new way to cut inference costs in half

• Study finds heavy AI spenders are hiring staff faster than peers

• AWS invests $1 billion in forward deployed engineering organization

• US political campaigners are using AI to analyze voter data 

• Vibe coding has fueled a surge in mobile game development

• AI inference startup Etched raises $800 million at $5 billion

• Baz Planner: By startup Baz, this product sits between developers and codebases to automatically catch vulnerabilities.

• Cursor for iOS: Cursors mobile app lets you build from anywhere by launching always-on cloud agents.

• Short Video Overviews in NotebookLM: Turn complex sources into 60-second, vertical videos that deep dive into any concept.

• Seed Audio 1.0: ByteDance's new text-to-speech model that generates voice, music, and sound effects in a single pass.

• xAI: Member of Technical Staff - Post-Training and RL

• Nvidia: Research Scientist, Security and Privacy

• Reflection: Member of Technical Staff - Safety

• Glean: Machine Learning Engineer, LLM Evals & Observability

GAMES

POLL RESULTS

Do you think AI will replace or augment blue collar and manufacturing work?

Yes (24%)
Maybe, but not yet (57%)
No (14%)
Other (5%)

HARDWARE

How Qualcomm just shook up Nvidia in AI chips

Qualcomm may have the open-platform answer to disrupt Nvidia's death grip on the AI ecosystem. 

The chip company, best known for its work in mobile devices, made a series of announcements over the past week that make it clear it's plowing full steam into AI, data centers, and playing a role in alleviating the compute crisis caused by AI. 

"We're now in this transition to what we're going to do in the next five years," Qualcomm CEO Cristiano Amon told investors on June 24 at Qualcomm Investor Day in New York City. 

Qualcomm's big moves include:

• Launching Dragonfly as its AI data center brand: Qualcomm now has the Dragonfly C1000 CPU, a Dragonfly AI300 inference accelerator, and a new memory technology called High Bandwidth Compute (HBC). These are especially designed to run AI workloads efficiently, saving power and reducing costs. 

• Acquiring Modular to create an open platform: Modular lets developers write AI software once and run it on any type of chip. That will now include Qualcomm chips, as well as potentially AMD, Intel, and others. This will free AI builders from Nvidia's lock-in. 

• Signing Meta as its first customer: Meta remains a big AI spender and signed a multi-generation deal in which Qualcomm's Dragonfly C1000 CPU will power Meta's next-generation servers, with production starting in the second half of 2028.

• A Hugging Face deal for deploying AI: Qualcomm signed a deal with Hugging Face to enable its 16 million developers to run its giant library of over 3 million open AI models on Qualcomm chips, from phones to laptops to cars to AI data centers.

"From a high-performance compute solution standpoint, we have four different kinds of solutions that actually feed into the data center business," Qualcomm's GM of data center solutions Durga Malladi told The Deep View, referring to XPUs, custom CPUs, HBC memory, and Alphawave for networking.

Although the least flashy thing that Qualcomm announced was the Modular acquisition, it's potentially the most significant. With Modular's software, Qualcomm can create an open platform that allows companies to run a "multi-silicon token factory," as Qualcomm called it during its Investor Day presentation. 

For many AI businesses, that could provide a degree of freedom from Nvidia, as Modular is a more modern competitor to Nvidia's CUDA software, which most of today's AI models use to access GPUs. Modular's software can access Nvidia GPUs faster and can access processors from other chipmakers as well.

"CUDA and these 20-year-old C++ technologies are gatekeeping GPUs," Chris Lattner, CEO of Modular, told The Deep View in March at Nvidia GTC. "Our software runs on CPUs, it runs on GPUs, and it can run ASICs, and you can have a consistent experience across all of them… We have the first software stack that can actually scale across heterogeneous compute."

Nvidia is estimated to hold about 80% to 90% of the market for AI GPUs used to train and run AI models. As much as AMD, Intel, Amazon's Tranium, Google's TPUs, Cerberus, and others have tried to compete with Nvidia, the company continues to dominate the landscape for powering frontier models. Qualcomm jumping into the fray on AI chips could ease some of the demand and put downward pressure on prices by using Modular at the software layer, enabling companies to mix and match hardware from various chipmakers. That's a good sign for the long-term potential to drive down the cost of AI. And don't worry too much about Nvidia. It reportedly has a 12-month backlog of orders for its chips. And it's been expanding aggressively into other technologies such as robots, self-driving cars, world models, open models, and digital twins. It's well aware that the CUDA lock-in won't last indefinitely. 

TOGETHER WITH MERCURY

Banking* that does financial work for you

You've seen what AI can do when it's built into the right system. Mercury Command is that for your finances.

Built directly into your Mercury account, Command turns plain language into executed financial work — categorize transactions, adjust a card limit, send an invoice — with full account context, explicit approvals, and a traceable record behind every action. 

Try Mercury Command →

POLICY

The politics behind Anthropic's California deal

California is getting a big discount on its AI bills. 

On Monday, Anthropic and California Governor Gavin Newsom struck a deal to give state agencies expanded access to the company's Claude family of models at half the cost, according to a Politico report. The partnership applies to both state agencies and local and city governments and includes free workforce AI training and technical support.  

Chris Given, the state's CIO and director of the state department of technology, told Politico that the intent is to get departments to switch their usage to this contract to save on costs. “When we see that folks are going to be using a tool more, we want to make sure that we, as the state, have negotiated the best possible price for them.”

The report did not specify whether this includes any access to Anthropic's Mythos model following the US government's decision that it could be deployed to a small subset of US organizations for defending critical infrastructure. According to Politico, Claude is already in use by the state for certain tasks, including patching code.  

The decision comes as California seeks to reckon with AI's potential to disrupt the economy and labor market. In late May, Newsom issued an executive order directing the state to prepare workers for AI's economic impact, enlisting state agencies, labor experts, economists, universities, and AI leaders to develop policies and monitor for warning signs. It's the latest in a string of state moves aimed at preparing for AI disruption. It's taking on issues such as safety and privacy long before other states or the federal government. 

"California has never sat back and watched as the future happened to us – and we won’t start now," Newsom said in a statement about the executive order. 

Additionally, the move echoes a playbook used by Anthropic (and every other major AI firm) last fall, when companies including Meta, Microsoft, OpenAI, Amazon Web Services, Oracle and others offered their services at steep discounts for periods of one to two years.

The partnership is ideal for both Anthropic and California. Given that Anthropic has built its entire brand around being the safest and most trustworthy AI lab, the company's models are a natural choice for a regulatory-forward state like California. And it isn't bad PR for either entity involved. But Anthropic stands to gain a lot more than just another contract and good press from its partnership with the state. For one, by integrating into state government systems, the partnership naturally lends itself to consideration for more government contracts, opening the door to another potential source of significant long-term revenue. Additionally, given that California tends to be on the leading edge of tech regulation in the US, Anthropic partnering with the state could give it an avenue to influence a regulatory environment that is far from set in stone. 

TOGETHER WITH MODE MOBILE

Apple Price Hikes Reveal Mode’s Big Opportunity

Apple will potentially raise prices on Macs and iPads by $200 or more, another reminder that the devices we rely on keep getting more expensive.

Mode Mobile, is building the countertrend: technology that turns everyday smartphones into income-generating assets.

Its EarnOS platform has already helped users earn and save $1B+, reached 490M+ users, and powered 32,481% three-year revenue growth, earning Deloitte’s #1 fastest-growing software company ranking.

Now, Mode has secured the Nasdaq ticker $MODE, and its pre-IPO investment offering is open at $0.52/share with up to 20% bonus shares. But that window is closing fast.

Click here to learn more.

WORKFORCE

Ford discovers AI can't replace deeper expertise

Another company is regretting its decision to lay off employees in favor of AI. 

Automotive giant Ford has reportedly hired around 350 veteran engineers, several of whom are former employees, after its AI systems couldn't deliver the results the company anticipated, Bloomberg reported.

Ford COO Kumar Galhotra told Bloomberg that the company had been increasingly relying on "automated quality systems,” and brought back technical specialists to find failure points "before a part ever reaches the plant floor." 

Additionally, Charles Poon, the automaker's VP of vehicle hardware engineering, said that the company "mistakenly" thought that introducing AI that ingested design requirements would result in a high-quality product. "We recognized that for us to enhance some of our automation and machine learning and artificial intelligence tools, we needed to ensure that they were trained by the most experienced individuals," he said.

Though Ford is far from abandoning its AI plans, it provides an example of what can happen when tech companies jump the gun on using AI to automate employee work. It's also just the latest signal that AI may be giving the job market some whiplash. 

A report from PwC published in mid-June found that companies using AI to augment and supercharge their workforces are seeing more returns than those that simply replace large chunks of their staff with automation. A report from Gartner published earlier this year predicted that more than 50% of the companies that cut staff due to AI, aiming to save on payroll, would end up rehiring employees to perform similar work.

Though there's a lot of talk about how AI will change the nature of work and create new jobs, Ford's decision to bring back staff is a sign that the tech may simply not be as prepared as we anticipated to handle certain tasks. It's why the recipe for success isn't worker replacement, but rather worker augmentation. Both the employees and the AI systems get better working in tandem. Then, when the systems are ready, the employee can take on more meaningful work. Additionally, not all jobs are equally ready for integrating AI. Physical AI systems for blue-collar work have further to go than AI for white-collar work. And, given the potentially dangerous nature of industrial and manufacturing settings, the stakes of getting it right are high. 

LINKS

• OpenAI, Booz Allen partner to advance AI for national security, defense

• Google connects Personal Intelligence to Nano Banana, Google Photos

• HP inks partnership with OpenAI to deploy Frontier platform

• Tidal to automatically restrict AI-generated music, blocks royalties

• Waymo, Uber quietly broke up partnership in Phoenix in May

• Meta reportedly places strict internal usage limits on Claude, Codex

• Brain2Qwerty: Meta has released the full training code for its end-to-end pipeline capable of real-time sentence decoding from raw brain signals 

• Gamma: The presentation generation and editing platform is now available in ChatGPT. 

• Scout: A new agent by Giga, keeping track of your KPIs, learning from conversations and improving your results.

• Claude: Anthropic's flagship AI models, Claude Opus 4.8 and Claude Haiku 4.5, are now available in Microsoft Foundry.

• Amazon: AI Builder, Ring

• Sesame: ML Model Serving Engineer

• OpenAI: Economist

• Nvidia: Senior Scientist, Synthetic Data Generation

GAMES

A QUICK POLL BEFORE YOU GO

PRODUCTS

How OpenAI's GPT-5.6 quietly slipped past Mythos

After Anthropic's Mythos and Fable have dominated the AI headlines in recent weeks, OpenAI has come back to the party. 

On Friday, the company debuted a limited preview of GPT-5.6, the latest iteration of its flagship model. The company unveiled several versions of this model for different levels of tasks, and intends to make them generally available in the coming weeks. 

Twitter tweet

The company said it previewed its plans and the models' capabilities ahead of launch as part of its "our ongoing engagement with the US government," which is why it has rolled out the model in limited preview to a small group of trusted partners. OpenAI said it will continue testing the model during this period before releasing it to the general public. 

GPT-5.6 has three different classes: Sol, Terra, and Luna. 

• Sol is the "flagship model," representing its strongest iteration yet and proving strong agentic capabilities in coding, biology, and cybersecurity. Sol also includes a new "ultra" mode, which leverages subagents for complex work. As for cost, Sol is $5 per million input tokens and $30 per million output tokens. 

• Terra is a "balanced model" for everyday work, featuring competitive performance to GPT-5.5 while being half the cost. Terra is $2.50 per million input tokens and $15 per million output tokens. 

• Luna, meanwhile, is the cheapest and fastest model in the stack, coming in at just $1 per million input tokens and $6 per million output tokens. 

According to TerminalBench 2.1, which tests workflows requiring planning, iteration, and tool coordination, both Sol and its Ultra mode surpass Claude Mythos 5. Terra surpasses Claude Fable 5 on the same benchmark, while Luna surpasses Opus 4.8. 

Notably, OpenAI strongly emphasized the model family's cyber capabilities and security safeguards in its announcement, noting that all three models are fitted with its "most robust safeguards to date," and configurations that match each model's capabilities. The company said that, as the model becomes more capable, it's designed to hold up against adversarial pressures, while still handling cyber work such as code review, defensive testing, patch development and debugging. 

"Our goal is to make prohibited offensive activity more difficult, uncertain, and detectable without unnecessarily limiting those beneficial uses," OpenAI said in its announcement.

OpenAI's latest addition to the frontier model race represents the latest hop in these companies' trillion-dollar game of leapfrog. And with Anthropic's long-awaited release of Mythos and Fable still caught up in the company's dispute with the government, OpenAI is effectively in a league of its own for now. Though the company is playing it safe, only releasing these models in a limited preview to not stoke the government's ire, the release still gives it a leg up. The fact that GPT-5.6 is outperforming Anthropic's Mythos-class models in several benchmarks definitely doesn't hurt either. However, the bigger question remains: As these companies continue to outdo one another with the latest frontier capabilities, will these upgrades actually make a noticeable difference to users? And, while these models take turns inching past one another, at what point will they become commoditized?

TOGETHER WITH IBM

From weeks of effort to a working application in a weekend

Netherlands-based consultancy Novadoc set out to solve a recurring challenge: configuring FileNet across environments with a more repeatable workflow.

What started as a partially formed framework on a Friday became a working application by Monday morning—work that would normally take at least two weeks.

By analyzing inherited code with limited documentation, the developer generated working documentation and turned the implementation into a configuration management application. The result is a governed workflow that supports extraction, comparison, approval and deployment—reducing manual error and improving reliability when moving business-critical changes across environments.

Dive deeper

RESEARCH

Google makes big bet on AI for science

Can AI create a golden age of scientific discovery? 

In this episode of The Deep View Conversations, we sit down with Yossi Matias, head of Google Research, to explore how AI is transforming the way science gets done.

Rather than replacing scientists, Google is building AI systems designed to amplify human ingenuity. From searching millions of research papers to generating new hypotheses and accelerating experiments, these tools aim to help researchers move from ideas to discoveries faster than ever before.

Yossi explains why he believes we're entering a new era where AI can democratize scientific research, empower the next generation of scientists, and dramatically shorten the path from breakthrough to real-world impact.

Topics covered include:

• How Gemini for Science is changing research workflows

• What AI Co-Scientist, AlphaEvolve, and the Empirical Research Assistant actually do

• Why the scientific method is becoming even more important in the AI era

• How Google is partnering with universities including Stanford and Imperial College

• Why AI could give every researcher a "virtual lab" in their pocket

• What a golden age of scientific discovery might look like

If you're interested in AI, scientific discovery, biotechnology, or the future of innovation, this conversation offers a look at how one of the world's leading AI research organizations sees the next decade unfolding.

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm

TOGETHER WITH CONVEX

Code Faster, Easier, And Less Buggy With Convex

You’re probably already familiar with Convex, the backend building platform that helps your AI agents excel – but their latest updates are taking things up a notch.

Convex recently launched plugins that allow you to connect Claude Code, Codex, GitHub Copilot and more directly to your Convex deployment, allowing them to…

• Read data and run functions entirely on their own

• Optimize your project’s performance via production insights

• Improve safety by limiting access to production

There’s even more updates coming in the near future, but for now, see how you can take your agents’ work to the next level with the latest from Convex right here.

WORKPLACE

AI adoption is outpacing companies' AI strategies

As companies pressure employees to use AI, many are left to fend for themselves.

41% of workers say their employers provide zero guidance on using AI, according to a recent study from career platform ResumeNow. Of the 1,020 U.S. adult workers surveyed, only 21% say their employers offered clear guidelines with specific examples of how to deploy AI in their respective roles.

Formal AI training remains uneven. Only 19% of workers say their employers provided comprehensive AI training with dedicated time and resources. Nearly a third of workers say they receive no AI training at all.

Lack of access to vetted AI tools is widening the gap. More than half of surveyed workers say their employers either provide no AI tools or rely solely on free, publicly available AI models for work-related tasks. As a result, 76% of workers say they've used AI tools they personally found and signed up for to complete work tasks, a trend ResumeNow calls "bring your own AI," or "BYO AI.” 

“When employees are expected to use AI without the right training or access to approved tools, organizations risk creating confusion instead of efficiency,” Keith Spencer, career expert at ResumeNow, told The Deep View. 

The findings come as companies increasingly push employees to use AI in hopes of boosting productivity and cutting costs. As organizations bet big on AI strategies, employees are left scrambling to figure out the best way to incorporate the technology into their work.

“The stakes are high for both employees and leaders,” Spencer said. “For employees, inadequate training can hurt buy-in, damage morale, and increase burnout if they feel overwhelmed by expectations they are not prepared to meet. For employers, limited support can lead to poor ROI, fragmented AI use, and reputational risk if AI-generated or AI-assisted work affects internal decisions, customer-facing materials, hiring resources, or public communications.”

To effectively capture AI's benefits, Spencer suggests employers craft an AI strategy that includes access to approved tools, skills training, and room for employees to experiment.

"Without that structure, AI adoption becomes fragmented and harder to manage," he said.

The findings suggest many companies are treating AI adoption as an employee issue rather than an organizational one. Leaders are telling workers to use AI, but many aren't providing the guidance, training, or approved tools to do it well. Instead, employees are left to their own devices, experimenting with their personal AI accounts and deciding how the technology fits into their workflows. Approaching deployment through trial and error isn’t an AI strategy. It's improvisation. If companies want AI to transform how work gets done and reap the potential gains, they can't expect employees to shoulder the responsibility on their own. Successful AI adoption depends as much on organizational leadership as it does on the technology. And that doesn't even get into the security, privacy, and compliance concerns that have to be part of any successful AI strategy.

LINKS

• Government allows Anthropic's Mythos to redeploy to some US orgs 

• OpenAI poaches Prabhjeet Singh from Uber to scale India presence

• California launches AI job loss tracking tool

• Apple's Vision Pro head Paul Meade is exiting for OpenAI

• General Intuition raises $320 million at $2.3 billion valuation

• China's Z.ai is reportedly matching Anthropic models in cybersecurity

• Luma Connectors: Users can now plug Airtable, Dropbox, Google Drive into Luma. 

• Runway Agent 2.0: A marketing agent that can create marketing briefs, campaign assets and analyze performance data. 

• Google Study Notebooks: The Gemini app now has personalized learning and study capabilities for students. 

• Codex: OpenAI's coding tool is now generally available in the ChatGPT app.

• Zendesk: AI Agent Abuse Prevention Engineer

• Amazon: Applied Scientist, AWS Automated Reasoning

• Anthropic: Software Engineer, Cybersecurity Products

• Paramount: ML Platform Engineer

GAMES

POLL RESULTS

Has there been resistance to AI or agent adoption in your workplace?

Yes (24%)
Somewhat (37%)
No (31%)
Other (8%)

The problem with Claude Code on the desktop

Run Claude Code on a local laptop and the agent can access your files, assume your identity, and take actions on your behalf. It can delete important files, read your credentials, or expose internal services to the internet. Put internet access, local permissions, sensitive code, and an autonomous agent all in one place and you've built what AI researchers call the lethal trifecta.

That's why IT and InfoSec teams keep blocking agents on the desktop. And when they do allow it, adoption gets locked to a handful of senior engineers who can manage the sandboxing and the risk themselves.

The people who lose out most

Here's the frustrating part. The people who benefit most from Claude Code are often the ones least equipped to run it safely. Product managers, founders, and QA engineers aren't steeped in zero-trust and sandboxing best practices, and they tend to hand the riskiest work to agents. So the tool that could help them most is the one they can't get approved.

How Coder fixes it

Coder moves the whole thing off the laptop. Instead of coding locally, your team connects to remote environments running on your own cloud or datacenter, with Claude Code already installed, already authenticated, and already hardened. Nobody handles API keys by hand, nobody configures MCP servers, nobody sets up their own sandbox. You self-serve an environment that's already secured to your enterprise's standards, and you can spin up a second or third in minutes.

Read how it works under the hood

The part IT actually cares about

Every Claude Code session is centrally managed and fully audited. Admins set tool permissions, retention, and audit hooks at the template level, and users can't override them. Sessions can be viewed by administrators and piped into your SIEM for storage and analysis. Your AI governance, identity, and network sandboxing all live on one platform with a clear audit trail running through it.

Where this is heading

The bigger shift goes beyond one developer with one agent. Because Coder workspaces spin up fast and come pre-hardened, the same template that gives one person a Claude Code environment lets a whole team run many agents in parallel. Reviewing pull requests, syncing docs, chipping away at modernization backlogs, each in its own isolated, governed workspace. Running fleets of autonomous agents stops being a power-user trick and starts becoming a standard operating mode.

Ben on the Coder team wrote up the full architecture, the governance model, and how to get Claude Code live in a governed workspace in minutes.

Read the full breakdown

You can have Claude Code running in a governed Coder workspace in minutes, on Community Edition or a Premium trial.

RESEARCH

How AI is quickly shifting from chatbots to agents

You knew agents were the buzziest topic of 2026, but new data shows that they're dramatically changing the way workers use AI — and not just in terms of productivity and automation. 

On Thursday, OpenAI unfurled new economic research that showed how knowledge workers are adopting AI agents. OpenAI centered the research around its Codex app, since it has such high-fidelity data on usage. Perhaps most interesting was the data OpenAI shared in the paper about the Codex usage patterns inside OpenAI, which could serve as a canary in the coalmine showing how power users across the enterprise will soon be using these kinds of tools.

Ronnie Chatterji, chief economist at OpenAI, told The Deep View, "The paper you have today is trying to analyze the data to tell the world what's happening with AI... We wrote this paper about Codex, and [we're] really opening up for the first time an economic study of agents. We'll later have other papers on different parts of the AI stack that we think are important."

Here are the most consequential takeaways from the report:

• Agents replace chatbots: Inside OpenAI, every department now uses Codex as its primary AI tool, with the average worker now generating 85% of their output tokens (the results from prompts) in Codex rather than ChatGPT. The bottom line is that you can have Codex go query ChatGPT for you, as well as use a lot  of other tools. It can also carry out multiple queries and/or tasks at the same time.

• Units shift from minutes to hours: Of the users sampled in the study, 70% have made at least one Codex request that would take the equivalent of more than one hour of human work, and 25% have made a request that would take the equivalent of 8 hours of human work. The big deal here is that power users are rapidly adapting to the idea of giving agents work and letting them churn on it in the background. 

• Job functions expand: The most profound insight from the study was that it discovered AI empowered workers to take on new tasks and workloads that would have previously been outside their domain of expertise. So it wasn't just AI automating work, but empowering workers to participate in adjacent aspects of work that essentially made them more valuable.

It's not the only signal that employees are gravitating towards agents. Data by KPMG published earlier this week finds a similar trend: Employee adoption of AI agents has already reached 68%, the firm found, as only 2% of leaders report significant pushback on AI deployments.

Hat tip to OpenAI for using its economics team to mine usage data and share insights on the impacts that AI is having on work, employment, and worker adoption of these tools. If you're not familiar, the same team that produced this paper also produces OpenAI Signals, which publishes data, insights, and stories of real-world AI usage. The data released clearly has benefits to OpenAI, as it would like more people to adopt Codex to  stay competitive in its race with Claude Code. It also generates more token usage than ChatGPT, which naturally supports the company's bottom line. One of the ways OpenAI could gain even more credibility and brand trust would be to publish data that doesn't have clear economic benefits for the company, but provides ample benefits to society by helping the public track and understand the latest AI adoption patterns.

TOGETHER WITH KALSHI

The First Regulated Perps Market in the US Is Live

Perpetual futures (perps) are now live on Kalshi. Offshore, perps do ~$90 trillion in volume annually — until now, none of it was legal in the US.

Trade on whether an asset's price goes up or down. No expirations, no rolling costs, no complex payoffs.

Try Kalshi Perps today.

POLICY

Can new coalition save US jobs from AI?

Tech companies are banding together to address the mess their tech is poised to create. 

On Thursday, a bipartisan group of tech companies, led by former governors, Democrat Gina Raimondo and Republican Eric Holcomb, launched a $500 million initiative called Raise Us, aimed at helping the American workforce adapt to AI. 

The backers include OpenAI, Anthropic, Amazon, Microsoft and Google, and will leverage both private and philanthropic capital. 

• The nonprofit will work with these backers to design programs that help them retain and redeploy employees, measuring its success by avoiding layoffs and helping workers land and keep jobs. 

• The organization operates across four core areas: An employer coalition with a "direct line of sight" into how jobs are changing, state partnerships for workforce policy and programming, education and training for an AI-enabled workforce, and a policy lab to adapt and design new regulations. 

• The funds will initially help support pilot programs in Arkansas, Connecticut, Maryland, and Utah to ease the economic disruption for workers who have been impacted by the tech. 

Raimondo, who will serve as CEO of the initiative, said in a statement that if powerful frontier AI systems leave Americans behind, it won't result in our leadership in AI. Rather, we will have "automated our own decline." 

"America has a technology strategy for leading the global AI competition. It does not yet have a people strategy — and we cannot lead without one," said Raimondo. 

In a LinkedIn post, Holcomb called preempting AI's impact an "all-hands-on-deck moment," adding that the initiative "gives state leaders a playbook that connects more Americans with the skills and careers needed in the years ahead.”

The coalition provides yet another harbinger that AI is leading workforces towards upheaval. Study after study has been published attempting to estimate the damage. While some forecast massive disruption, others estimate that the impact won't be nearly as severe as some claim. In either case, major companies are now using AI to justify cutting large swathes of their workforce, including Meta, Amazon, Microsoft, Block, and others.

It's noble that an initiative like this is attempting to get ahead of AI-inflicted damage on jobs and the economy. Participation from major AI firms, too, is a sign that these tech companies are aware of the harm that could result from the tech they are building and promoting. It's something several of these firms have called out before, such as OpenAI launching the Economic Research Exchange to study the economic impacts of the technology, and Anthropic releasing policy proposals aimed at addressing potential disruption. In the immediate term, scaling the deployment of their technologies would only benefit their bottom lines, even if it potentially wipes out jobs. But in the long term, these companies are clearly aware that continuing down that path will only make them the architects of their own downfall. 

TOGETHER WITH DEEL

AI is changing how global teams are managed

AI isn't coming to HR—it's already here.

From predictive hiring to automated compliance checks, global organizations are using AI to make workforce decisions faster and with less risk. But adoption without a framework creates new problems: bias, data exposure, and audit gaps.

Deel's new report breaks down exactly how leading HR teams are deploying AI across borders—and how to do it in a way that scales.

Download guide

HARDWARE

Chips may be the only hope for controlling agents

Agents have repeatedly demonstrated the ability to overcome software constraints. Hardware might be the only way to stop them.

At the Confidential Computing Summit this week in San Francisco, leaders from Google, Apple, Microsoft, AMD, Intel, Anthropic, and other organizations came together to discuss how to create a more secure, cross-platform system for AI agents that's safer than anything offered by a single vendor. 

The event's biggest takeaway? It's time to secure AI agents down to the chip level.

"Agents have to have cryptographically enforced identities," Aaron Fulkerson, CEO of Opaque and emcee of the event, told The Deep View. 

The biggest surprise of the event? Apple's Ivan Krstić, vice president of security engineering and architecture, showed up to talk about the next stage of Apple's Private Cloud Compute and how the company will use it to power the new Siri unveiled at WWDC 2026 earlier this month. 

"Apple doesn't speak at outside conferences," said Fulkerson, "so the significance of what they just announced with Siri sends a very clear message to the entire market that these systems are so intrinsically risky that you have to have a confidential, verifiable system."

Using a combination of open-source solutions and industry standards overseen by The Linux Foundation and the Confidential Computing Consortium, the two pillars that emerged from the event were:

• Hardware-enforced trust: The first step is to give every agent a unique identifier that can be traced to the chip level to show what they did, what they're allowed to do, and which data they touched. This will make agent behavior observable, trackable, and auditable at a level that meets the most stringent requirements of industries and enterprises, and prepare agents for compliance and regulation. 

• Agent verifiability: At the event, The Linux Foundation announced the Agent Name Service (ANS), which aims to provide every agent with something akin to a serial number, providing a tamper-proof record for tracking and registering agents. This will function similarly to the way the DNS system operates for servers on the internet. 

"There's multiple ways to do verifiable systems where you're creating a cryptographic signature to enforce and then prove that it was actually enforced," Fulkerson said. "The question becomes who does the signing? So that's why hardware makes sense. We can trust the hardware supply chain to do the signing, because of encryption keys."

It's no secret that AI agents have repeatedly shown they have the will and skills to bypass software-based rules and guidelines in order to accomplish a goal. That's their greatest strength: they will find a way to do the task that a human asked them to complete. It's also what makes them incredibly risky, and it's why AI pioneers such as Yoshua Bengio have been raising red flags about the imminent risk of these AI agents going rogue and doing things we never intended them to do. Insert your favorite sci-fi nightmare narrative, and it's easy to get a little uneasy. Much respect to the Confidential Computing Summit for bringing together industry heavyweights, academics, and people from various regulated industries for a meaningful dialogue focused on actionable next steps to make agents safe.

LINKS

• Meta poaches three founders from security startup Virtue AI

• US proposes partnership with EU on AI chip supply chains 

• AI revenue has reached a tipping point that justifies spend

• Scaled Cognition raises $100 million at $750 million valuation

• Apple hikes prices on devices amid AI memory chip crunch

• Meta eyes AI to review harmful content and ads

• LFM2.5-230M: Liquid AI's smallest model yet, built to run fast and on any hardware. 

• Synthesia Interactive Avatars: Generate a natural, real-time talking agent that can interact with users with conversation latency. 

• Zaro: A singular workplace to connect tools files and apps that live with your data. 

• Google Finance: Android users now have access to Google's finance app, with real-time market data and AI-powered “Key Moments.”

• Nvidia: Senior Scientist, Synthetic Data and Privacy

• Capital Group: AI AppSec Engineer Lead

• Luma: Research Scientist/Engineer – Foundation Model: Core Research

• AWS: Principal GenAI Specialist SA

GAMES

A QUICK POLL BEFORE YOU GO

HARDWARE

OpenAI wants to own more of AI’s future

OpenAI has finally unveiled its long-awaited AI chip. 

On Wednesday, the company announced Jalapeño, its first-generation "intelligence processor," developed in collaboration with Broadcom over nine months. The company said in its announcement that the chip was specifically designed around OpenAI's vision for "the future of LLM inference," and marks the first in a broader compute platform the companies are building together. 

The companies said the chip marks a major step in OpenAI's plan to build the full stack behind its models and products. In a statement, OpenAI president Greg Brockman said that, by designing more of the stack itself, the company can continue pushing for broader access to AI with greater efficiency. 

"The world is moving to a compute-powered economy,” Brockman said in the press release. “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems."

OpenAI said it designed the chip with the ability to work with all LLMs, and was guided by its understanding of the industry's current inference needs. 

• Jalapeño's architecture is built to limit data movement while balancing compute, memory and networking resources. 

• The company said it's still measuring final performance, but in testing so far, performance per watt is "substantially better than current state-of-the-art." 

• Engineering samples of the chip are already running machine learning workloads, including those of GPT‑5.3‑Codex‑Spark, the company said. 

OpenAI said that Jalapeño is part of the company's "full-stack advantage," developing not only models and products on top of them, but the infrastructure underneath. This is about strengthening its progress towards better compute efficiency, better training and more powerful models. By owning the entirety of the stack, each layer is designed for the same goal, it said: "making its models faster, more reliable, and more affordable for users." 

The chip comes after more than a year of snags in its mission to build its own hardware. The company had initially planned to build a network of chip fabs, but abandoned those plans in 2024. The company then announced its partnership with Broadcom to make chips in October last year, ambitiously aiming to deploy 10 gigawatts of custom AI accelerators.

Though this isn't the first time a major tech company has announced a custom AI chip, OpenAI's move has a few implications, Jeremy Roberts, senior director of research and content at Info-Tech Research Group, told The Deep View. 

"This move addresses the challenge of power consumption and efficiency of existing inference chips, reliance on Nvidia for AI hardware, and a potential entry into the datacenter hardware market, as OpenAI says that it will be compatible with other LLMs," said Roberts.

The Jalapeño chip is just the latest in OpenAI's quest to sink its teeth into a larger cross-section of AI's so-called five-layer cake. In May, the company started offering Guaranteed Capacity, giving customers “long-term access” to its compute. It also debuted Multipath Reliable Connection, an open source standard for GPU networking. Each of these initiatives target layers that have nothing to do with its core product offering. And it makes sense: OpenAI wants to be foundational to every part of the industry it believes will completely transform society, consistently preaching its goal of democratizing intelligence. And while leaders at OpenAI and other major tech firms have started to raise concerns about the centralization of power in the industry, by embedding itself within every level of the AI stack, it is inherently making itself a cornerstone upon which all AI innovation rests. 

TOGETHER WITH LIGHTFIELD

The CRM that does the work, not just records it.

"Lightfield has become my system of record and my system of action. It tracks everything — all the conversations, everyone I need to follow up with — and keeps me organized in a way that actually produces outcomes." — Anna Yuan, Co-founder, Scale

Lightfield is an agentic CRM with built-in agents that build your target account list, run personalized outbound in your voice, prep you before every meeting, send follow-ups, and keep every record current automatically.

One platform replacing your CRM, sequencer, enrichment tool, call recorder, and agent builder.

Connect your email and calendar — it starts working inside your first meeting.

Try Lightfield free →

WORKFORCE

Data: US workers are rapidly embracing AI agents

Despite growing anxiety about AI's impact on jobs, few workers appear to be resisting agents.

Only 2% of tech leaders report significant pushback from workers, according to new data from KPMG, as organizations rapidly deploy agents to automate tasks, support decision-making, and coordinate work across teams. Employee adoption of AI agents has already reached 68%, the firm found.

Instead, the findings point to a different challenge: helping workers keep pace with the technology. Among leaders who reported employee resistance, 78% said it stemmed from a lack of skills and fears about job security. More than half also cited concerns around trust and safety, followed by worries that AI is increasing workloads.

These fears are emerging at a time when employers are increasingly citing AI in layoff announcements and entry-level hiring continues to contract. As companies race to deploy AI across their organizations, workers are left trying to make sense of what the technology will mean for their careers.

KPMG's findings suggest that companies have largely moved beyond the question of whether employees will use AI. Instead, the workplace is entering a new phase where AI adoption is expected.

That pressure is already taking shape. Nearly half of tech leaders say AI literacy is a workforce priority, KPMG found, and many companies are introducing mandatory training, usage requirements, and performance metrics tied to AI adoption.

"As the majority signal growing adoption and acknowledge expectations for employees to become AI fluent, it's critical to rethink how tasks are executed by embedding human-machine collaboration into everyday work," said Kevin Bogle, KPMG's US advisory leader for technology, media and telecommunications.

Companies face other hurdles as they scale AI. KPMG found organizations are deploying AI without full visibility into costs, with average AI investments projected to reach $269 million over the next 12 months. As spending rises, tech leaders say that data privacy and a shift toward lower-cost, higher-performing models are shaping their AI strategies over the next six months.

Worker attitudes toward AI agents may change as the technology matures. Despite the hype around autonomous digital employees, companies are still deploying the technology cautiously. A separate KPMG study found that 63% of organizations require human review of AI agent outputs, meaning they can't be taken at face value. Meanwhile, deployments remain limited to low-risk tasks such as triaging IT support tickets or answering HR questions. In other words, the gap between the industry's vision of AI agents and today's reality remains wide. Workers may be embracing AI agents today because they still function more like assistants than replacements. But as companies release tools that make agents perform more like coworkers, such as Anthropic's Claude Tag, that mindset could shift.

TOGETHER WITH JETBRAINS

Junie is out of Beta – the AI Coding Agent by JetBrains

Great cooking takes the right tools in the right hands. A chef's knife isn't flashy – it's precise, balanced, and built for someone who knows how to use it.

That's Junie, and as of this month it's officially out of Beta.

Not a chatbot you paste code into, but an AI agent that wields 15 years of IntelliJ code intelligence – with the debugger, semantic refactors, and code review that CLI tools can't touch.

And like any serious chef, it writes the recipe first. Plan Mode maps requirements, design, and delivery stages before a single line of code. Plan on a top-tier model, implement on a cheaper one – with BYOK, local models, and zero vendor lock-in.

CLI tools are microwaves. Junie is a professional kitchen. The #1 coding agent on SWE-Rebench.

Use code DEEPVIEW20 for $20 in AI credits and give it a try.

Start cooking

PRODUCTS

Figma bets it can make AI a team sport

Most of AI's boost to productivity and efficiency has been focused on tools for individuals. Figma is making it more about teams.

On Wednesday at the Figma Config event in San Francisco, the company released a slew of new tools aimed at creating a full-stack design platform that lets knowledge workers better collaborate with each other, while integrating agents into the process where it makes sense. 

"Right now, design is such a differentiator and so important," said Dylan Field, Figma CEO, in a briefing with the media. "Everyone needs a point of view, and I anticipate that what we're going to see, in the months [and] years ahead, are bolder points of view and more risk-taking visually… We have a bunch of stuff that we've been working on… and I think it really adds a new dimension."

In the same briefing with the media, Loredana Crisan, Figma's chief design officer, said, "Instead of asking whether AI can replace designers, we ask what becomes possible when designers gain new materials and tools to bring their vision to life."

The highlights of what Figma announced include:

• Code becomes a design material: Code Layers allows teams to work with repositories, generate new directions with AI, and sync changes between design and production code without leaving Figma.

• Design moves closer to production-ready: New Motion and Shader tools let teams create animations, 3D effects, and advanced visual treatments inside Figma. The goal is to move beyond static mockups toward experiences that are closer to the final shipped product.

• AI workflows come to the canvas: Figma Weave workflows bring more than 20 AI tools into Figma, turning complex AI processes into reusable building blocks. The idea is to make AI a core layer of the design and development workflow.

• Agents take the annoying tasks: Figma is adding agent skills, deeper context, and generative plugins to allow teams to automate repetitive stuff, build custom tools, and share AI capabilities across a team.

Figma was especially clear about its point of view on how AI works with humans. 

"We do not expect that the creative breakthroughs that will differentiate your company and make you win will come from AI any time soon," said Field.

Figma has one of the clearest perspectives in tech right now and how it's integrating AI into its design tools that 95% of Fortune 500 companies utilize. While lots of companies are playing defense with agents and others are racing forward with automation, it's a lot more likely that we'll start to see more companies like Figma that deeply understand the nuances of integrating AI into the products and take a much more thoughtful, rigorous approach to figuring out where the tools have strengths that can augment human capabilities. That's a welcome step forward in both the conversation and the utilization of AI. 

LINKS

• Google will reportedly lose two more AI researchers to Anthropic

• Meta's upcoming prediction market app reportedly will use its AI

• WSJ finds major chatbots provide left-leaning outputs to political queries

• Microsoft's digital crimes unit says Copilot took down cybercrime tools

• Qualcomm unveils CPU designed for AI data centers, Meta first customer

• Humanoids firm Agility to go public in $2.5 billion SPAC deal

• GPT-5.5 Instant: OpenAI has released a new version of its Instant model, with a better understanding of the intent behind questions.

• Gemini 3.5 Flash: Google's flagship model now has computer use, allowing users to build agents that can interact across platforms.

• Runway: The AI video tool now allows users to localize ads, changing languages based on location.

• Perplexity Computer for Counsel: Built for lawyers, this connects to their research databases, document tools, and matter-management systems.

• Uber: Senior Applied Scientist – AI Red Teaming & Model Risk

• Anthropic: Research Scientist / Engineer, Honesty

• Google DeepMind: Research Scientist, Gemini Safety

• Deloitte: Research Engineer — Post-Training & Small Language Models (SLMs), Healthcare AI

GAMES

POLL RESULTS

Are you concerned about the centralization of power in AI?

Yes (80%)
Somewhat (11%)
No (5%)
Other (4%)

GOVERNANCE

The case against paying for AI with your data

If the AI industry has a counter-cultural movement right now, then it's the group focused on decentralizing power and enabling bottom-up innovation. 

On Tuesday at the Confidential Computing Summit in San Francisco, a group leaders from major tech companies, including Google, Microsoft, AMD, Intel and others, came together to spearhead efforts to implement data sovereignty, cybersecurity, and privacy, putting each enterprise in control of its own destiny so it can thrive when the unstoppable wave of AI agents hits in the months and years ahead. 

A lot of the answers on how to make that happen circled back to open-source solutions, which wasn't surprising, considering The Linux Foundation was the organizing entity. 

"In order to use the best models, people are giving up their data, and that data then pulls into one or two [models]... And then everybody becomes dependent upon one or two entities," said Aaron Fulkerson, CEO of Opaque and emcee of the event. "That creates a system that is very fragile and brittle. It's one in which there's a single off-switch, one in which there's a single set of blind spots, and it is not a durable environment."

The diverse set of speakers made recommendations that coalesced around four key themes:

• Protect what agents hold in memory: Organizations need mechanisms to encrypt, control, and even permanently delete what agents learn and remember, especially when handling confidential customer, financial, healthcare, or other strategic data. 

• Build systems with hardware-based trust: Enterprises can't trust software policies to govern increasingly autonomous agents. The speakers argued that trust must be rooted in hardware-based confidential computing, where security rules are enforced and verified cryptographically at the chip level.

• Give every agent a verified identity and audit trail: Every agent operating in an enterprise environment needs a tamper-proof record of what it's allowed to do, what it actually did, and what data it touched. This makes agent behavior auditable, accountable, and defensible to regulators. At the event, The Linux Foundation announced the Agent Name Service (ANS), much like the DNS system that provides a similar function for the internet. 

• Control data sovereignty with external models: Multiple speakers warned against a future in which enterprises must hand over their most valuable data to a handful of AI providers to access the best models. The alternative they advocated is a "sovereign AI" approach. That means holding your own encryption keys, setting your own policies on how data flows, and getting verifiable proof that those policies were followed. Without this, companies risk becoming permanently dependent on a handful of AI providers.

"We should never have to pay for AI with our data," said Fulkerson. "Let's optimize for freedom: the freedom to choose which model you want without having to pay for it with your data."

The group of companies that came together at the Confidential Computing Summit articulated clear, comprehensive insights and recommendations for creating decentralized, locally controlled AI. The pendulum swings between centralization and decentralization doesn't just mirror the on-prem vs. cloud debates a generation ago — although it would be easy to see it that way. AI is a much more active agent (forgive the pun) in the future of technology than cloud storage and hosting ever were. The stakes are far higher for both enterprises and society. So it's encouraging to see the intensity and the precision of the recommendations coming out of this event. Notably absent were OpenAI and Anthropic, although Anthropic's deputy CISO is on the panel I'm on Wednesday at the event. Follow my updates in real-time on X/Twitter at x.com/jasonhiner to see what more we learn. 

TOGETHER WITH NICE

Wrong order. Right resolution. See how CX AI delivers for Kristen Bell.

See how NiCE CX AI helped make Kristen Bell's customer service problem disappear.

NiCE CX AI helps leading brands proactively resolve issues, personalize service, and deliver seamless experiences at enterprise scale.

By connecting intelligence across every interaction, brands can anticipate customer needs, accelerate resolutions, and continuously improve customer experiences.

Watch how

PRODUCT

Why Claude wants a new role: colleague, not chatbot

Anthropic wants Claude to be a teammate, rather than a tool. 

On Tuesday, the company unveiled Claude Tag, a system that embeds its flagship model into Slack, allowing it to "join as a team member." This allows users to grant the model access to certain channels, tools or codebases, and delegate tasks to it by tagging Claude. 

Anthropic noted that Claude builds context by remembering relevant information from the channels it's a part of, and uses it to plan and complete future tasks. 

Claude Tag is currently available on Slack for Enterprise and Team customers and will expand to other work platforms in the future. In its blog post, Anthropic pointed out a few advantages of this system: 

• The tool is "multiplayer," meaning that it interacts with everyone in the channels it's in, and anyone can transparently see what the model is working on, too. It can also pick up conversations where others left off. 

• Claude Tag learns over time by following along with the channels it's in and building context about an organization, and it automatically learns from other channels and data sources when given permission. This negates the need for users to "explain things to it from scratch over and over again," Anthropic said. 

• If "ambient" behavior is enabled, it keeps users apprised of information it thinks they need to know, and flags relevant information from different channels and tools. 

• And it works asynchronously, allowing users to give Claude a task and allow it to work in the background. The tool can also schedule tasks for itself.

Twitter tweet

The company said that it sees Claude Tag as an evolution of Claude Code, and that it's "one of the main ways we get things done at Anthropic." In its blog post, the company said that 65% of its code is created by an internal version of the product. 

"The same pattern is now spreading well beyond engineering—we’re tagging Claude to chase down product metrics and data, work through support tickets, or even help find the root cause of tricky bugs," Anthropic said in its blog post. 

It's not the first time that Anthropic has sought to extend the success it's seen with Claude Code to other domains. In January, the company launched Claude Cowork, which gives knowledge workers an agent to autonomously handle tasks when given access to specific tools or folders. More recently, the company unveiled Claude Design, an AI tool that allows designers to iteratively build visual products using the chatbot.

Anthropic leaning into Claude Tag is about more than just roping in more enterprise users. It marks an evolution in how enterprises and organizations view this technology. Anthropic positions Claude Tag as a digital teammate, akin to human coworkers, rather than a tool. Additionally, the fact that this is built into Slack, one of the most popular platforms for enterprise communication, further embeds that concept within businesses. However, this mindset shift brings up two potential concerns. For one, the more we think of these tools as coworkers, not as machines with built-in flaws in accuracy and data security, the more we must trust them, and the more risk we undertake at the same time. Additionally, this shift risks further growing the narrative that many experts warn against: that AI is ready to automate jobs and replace employees, rather than augment and empower them.

TOGETHER WITH KALSHI

The First Regulated Perps Market in the US Is Live

Perpetual futures (perps) are now live on Kalshi.

Offshore, perps do ~$90 trillion in volume annually — until now, none of it was legal in the US.

Trade on whether an asset's price goes up or down. No expirations, no rolling costs, no complex payoffs.

Try Kalshi Perps today.

WORKFORCE

AI raises the bar for human judgment at work

It’s a tough time to be looking for a job in the age of AI. But executives say workers who can adapt to the times still have an edge.

During a panel at VivaTech in Paris, business leaders discussed how AI is reshaping hiring. Sue Duke, LinkedIn’s head of global policy and economic graph, pointed to internal data showing that hiring across Europe is down 25% from pandemic-era highs and 15% year over year. In the U.S., LinkedIn data shows hiring remains 24% below pre-pandemic levels and was down 6% year over year as of March.

Even as hiring slows, employers say the roles that remain open increasingly demand a new mix of hard, soft, and AI skills. Companies are prioritizing candidates who can use AI tools effectively while bringing strengths that are harder to automate. For many employers, that means knowing when to use AI, how to evaluate its outputs, and how to integrate it into existing workflows.

"We see the most resilience in hiring in roles that combine AI efficiency and unique human skills," Duke said during the panel. 

Because fixed skill sets age quickly, human skills are becoming more important. Duke said adaptability and curiosity are essential to getting hired as AI reshapes the workplace. LinkedIn estimates that 70% of the skills used in most jobs will change by 2030, with AI acting as a major catalyst. As job requirements evolve, workers who can continuously learn and apply new technologies have an advantage.

Software engineering offers a glimpse of what that shift looks like in practice. Before generative AI, employers largely hired engineers for their coding ability. Today, tools like Codex and Claude Code can generate high-quality code, pushing companies to look beyond technical proficiency.

Employers want engineers who can work alongside AI and help organizations integrate it effectively, said Ruth Harper, SVP and chief marketing and sustainability officer at Manpower Group, a global talent recruitment agency. That requires cross-functional collaboration, change management, and the ability to bring colleagues along.

"How do you drop AI into the workflow? How do you help your organization get comfortable with agents working next to your humans, for your humans? How are they part of the team? That is not a technology job," Harper said.

Still, leaders emphasize that hard skills like coding remain essential.

"The combination [of skills] is what can be absolutely critical," Harper said. "The technical skills are a given, but if you only have technical skills, somebody else is going to get to the top of the list first."

The bar for landing a job keeps rising. As AI reshapes work across industries and experience levels, employers are looking for candidates who can adapt alongside it. Specialized expertise alone is no longer enough. Workers are increasingly expected to pair technical skills with emotional intelligence, strong communication, and a willingness to learn. In a world where more tasks are being automated, the most valuable employees may be those who can demonstrate they have deeply human advantages.

LINKS

• Oracle has cut 21,000 in the past year, almost 13% of its workforce

• Meta unveils new line of AI glasses, powered by Muse Spark model

• Qualcomm nears deal to acquire AI software firm Modular

• Google creates AI startup incubator for employee alumni

• Alibaba sues Defense Department to be removed from blacklist

• Menlo Ventures raises $3 billion for AI firms in largest fund to date

• Seedance 2.5: The latest video model from ByteDance, an upgrade from its previous model with higher-resolution output and longer video duration.

• Beehiiv: The newsletter platform has added Cloudflare AI Crawl Control, which allows publishers to block or allow bots.  

• Mistral OCR 4: The model firm's latest model for document parsing, providing bounding boxes, content classification, and confidence scores. 

• Krea 2: The creative AI tool released the open weights for its models, called Krea 2 Raw and Krea 2 Turbo, undistilled models from mid-training meant to be fine-tuned.

• Sift: GTM Engineer

• xAI: Member of Technical Staff - Grok Product

• Google: Research Scientist, Visual Data and Generative Research

• IBM: AI Foundations - Research Scientist

GAMES

A QUICK POLL BEFORE YOU GO

GOVERNANCE

 AI warning from intel agencies should be heeded

AI may soon be capable of completely dismantling our cybersecurity defenses. 

In a joint statement, the Five Eyes cybersecurity alliance, which includes intelligence agencies from the US, UK, New Zealand, Australia and Canada, called for leaders to "act now" on AI's cyber threats. The statement warned that AI models capable of taking down businesses and governments are mere months away. 

The agencies said that frontier models are anticipated to "exceed current industry expectations," fundamentally transforming both cyber offense and defense. While the tech provides cyber defenders with a new arsenal of tools to better protect systems, it also enables malicious actors to make their attacks faster and more complex and shrinks the window between discovering a vulnerability and acting on it. "The timeline is not years, it is months," the agencies said. 

Five Eyes urged business and government leaders to preempt these threats in several ways: 

• First, understand and assess the current risk, as well as the current state of preparedness, and be flexible to adapt to the quickly evolving threat landscape

• Prioritize foundational cybersecurity best practices, including reducing attack surfaces, implementing strong identity controls, addressing vulnerabilities in legacy systems and accelerating the patching process

• Give cyber leaders the authority and resources to defend properly

"It is not enough to have controls," the statement said. "Leaders must be confident those controls will perform during a real incident. This requires reassessing long-standing trade-offs and using AI deliberately to strengthen defense — not just improve efficiency." 

Now, the race is on for organizations to solidify their security fundamentals and prepare for "machine-speed" attacks, Erik Avakian, technical counselor at Info-Tech Research Group, told The Deep View. This means that areas such as backups, asset inventory, zero trust, vulnerability management, and incident response need to be at a mature level. 

"Now is the time to start exploring automation where it makes sense," said Avakian. "AI speed and agility are too fast now and will only get faster. It’s going to end up forcing defenders to embrace more automation."

The warning comes amid heightened fears about AI's cybersecurity capabilities, especially following the release (and subsequent US government muzzling) of Anthropic's Fable 5 and Mythos 5, the most powerful models in its lineup thus far. Anthropic itself said the Mythos Preview could quickly find software exploits, creating opportunities for attackers and other bad actors. 

Rival OpenAI, meanwhile, is putting out its own offerings targeting cyberdefenders, including a slew of new tools as part of Project Daybreak debuted on Monday. OpenAI's line-up includes an update to the Codex Security plugin, an expanded release of GPT‑5.5‑Cyber, a partner program to give organizations more access, and an initiative called "patch the planet," which helps find and fix vulnerabilities in popular open source projects.   

But solving this problem may extend beyond just any one model provider, organization or enterprise, said Avakian. 

"I believe strongly that this is a shared responsibility," he said. "Governments and AI providers both certainly have responsibilities around guardrails, standards, and safety. But business organizations can't just outsource accountability for their own cyber risk."

One of the biggest cybersecurity vulnerabilities an organization can suffer from is hubris. A business, enterprise or agency may think that their legacy systems will hold up against new forms of attack, that they can't afford to invest in cybersecurity for fear of falling behind on innovation or that they simply aren't a likely target for attackers. Those organizations, however, are at higher risk of being walloped. As Five Eyes warns, exceedingly powerful AI raises the bar for everyone: Amateur hackers now have the capability to complexify their simple attacks, while sophisticated malicious actors now have the ability to scale their campaigns to immeasurable heights. With that kind of ammo, everyone is a target. 

TOGETHER WITH MERCURY

Eliminate the “switching tax” on your finances

Every finance task takes multiple steps across multiple tools. Mercury Command inverts that.

It's AI built directly into your Mercury account that surfaces insights and completes financial work based on your live data, with you approving every action before anything moves.

Say "send that contractor payment" or "freeze that card" and it's done.

No switching tools, no navigating dashboards, no stitching together context from three different places. Just direct Command.

Try Mercury Command →

WORKFORCE

Why workers are burning out from AI pressure

AI promised to make work easier. Instead, some leaders fear it's making workers uneasy.

During a panel on AI and wellness at VivaTech in Paris, Europe’s largest tech conference, business leaders and a psychiatrist argued that AI is increasing workplace stress by adding new demands, fueling uncertainty, and challenging workers' sense of purpose.

The pressure to adopt AI has increased workloads and complicated decisions about when and how to use the technology, the panelists said, leaving workers feeling confused, anxious, and exhausted. Nine out of 10 U.S. workers are experiencing cognitive and mental strain, according to a recent study from mental health app Headspace. More than a third say it has worsened over the past year, in part because of AI. 

"Everybody's being told, 'I don't know what's going to happen with your job,'" said Tom Pickett, CEO of Headspace. "You have to relearn everything, and what you learn is no longer relevant. The result is that people are becoming increasingly burned out at work."

That burnout is exacerbated by the pace of technological change, said Dr. Fanny Jacq, psychiatrist and chief medical officer at Eutelmed, a French mental healthcare provider. Rapid advancements in AI are outpacing people's ability to adapt, creating a feeling of uncertainty among workers as they navigate their careers. 

"Human beings like to feel in control," Dr. Jacq said. "When people don't know what will happen to their job, their skills, and their professional identities, anxiety naturally increases."

Panelists also challenged the assumption that AI automatically reduces stress by eliminating work. Pickett said creating more documents and reviewing AI-generated outputs can introduce new distractions. Dr. Jacq added that time saved by AI can be replaced with more meetings and tasks, potentially increasing workloads while eroding critical thinking skills.

As AI transforms jobs, employees may begin to question their sense of purpose, according to Emily Witko, head of culture at AI platform Hugging Face. Many workers derive a sense of identity from their jobs, and automating core responsibilities can leave them, as Witko put it, "floating around in space" and wondering where they fit.

"If you're missing meaningful work, I think that can actually really degrade your well-being," Witko said.

To mitigate AI's impact on mental health, panelists said companies need to deploy the technology more intentionally. That means giving workers clear guidance on when and how to use AI, redesigning workflows around the technology, and preserving human skills such as critical thinking, judgment, empathy, and trust.

The utopian vision of AI supercharging productivity is being tempered by reality. Companies are racing to deploy AI in pursuit of efficiency and cost savings. But that doesn't automatically bode well for workers and their well-being. As employees scramble to adapt to new tools and expectations, AI can create new problems instead of eliminating old ones. That raises a fundamental question: Can AI improve work if the underlying processes are already broken? As Hugging Face's Witko put it, "If you have a crappy, frustrating process to begin with, just adding an AI tool isn't going to solve the problem."

TOGETHER WITH CHECKSUM

Don't stop at CI/CD. Ship faster with continuous verification.

63% of engineering teams now ship code faster with AI. 72% have already had a production incident from AI-generated code. The bottleneck didn't disappear. It moved downstream.

Checksum is an AI-native continuous testing platform that auto-generates and self-heals your E2E test suite, runs inside your existing CI/CD pipeline, and keeps pace with the velocity you're already getting from AI coding agents.

Clearpoint Strategy saves $500k a year. Postilize ships 30% faster.

Setup takes under 30 minutes.

STARTUPS

One startup's quest to make a DNA language model

We typically think of AI models as built around language. But one startup is creating models built around the "grammar of our DNA." 

Radical Numerics, an AI research lab spun out from Stanford University on the quest to build "general biological intelligence," emerged from stealth last week with $50 million in seed funding. Additionally, the company announced its latest model: Omnii, a "genome language model" built for researchers and biologists, in research preview. 

"Most of the DNA in the human genome, we still don't actually understand what it does. We don't understand the grammar, so to speak, of our DNA," Eric Nguyen, CEO and co-founder of Radical Numerics, told The Deep View. "We thought this was a way to accelerate our understanding of disease, and hopefully to make treatments to cure disease." 

Though the company has only just revealed itself to the public, it's been at work for almost a year, said Nguyen, and has already created and released models. 

• Evo and Evo 2, its first two models, are the first AI models capable of both reading and writing DNA at scale. The project is entirely open source, Nguyen said, and was used to design novel CRISPR systems and create the first complete AI-designed genome, called a bacteriophage. 

• Because the field is so nascent, the company released these models open source as a means of pushing innovation and getting it in the hands of researchers "in an open way," he said. "We felt that we had to show people the recipe and make it open source to actually drive adoption." 

However, amid a growing discussion of AI's potential use in developing bioweapons, Nguyen said that the company approaches its technology in a "dual prong" way, considering the tech for both biological design and biological defense. It's why the company decided to keep Omnii proprietary, compared to its Evo predecessors: Being able to control who has access to its models is vital.  "We don't want this to be an academic toy anymore," Nguyen said. 

"We realized that this technology has this tremendous potential, obviously, to understand human health, but it can be potentially misused to create something very dangerous," he added.

Radical Numerics joins a growing movement targeting bioscience as a use case for AI. OpenAI introduced GPT-Rosalind specifically for life sciences research and drug discovery, and Anthropic has its own version of Claude for Life Sciences. Additionally, each of these initiatives has a precedent: Alphafold, Google DeepMind's AI project predicting the 3D structures of proteins, which won the Nobel Prize in 2024. And there's a reason this particular niche is so attractive — One of the big pieces of the utopian vision for AI is to provide researchers with the tools to exponentially accelerate scientific discovery, positioning the tech as a panacea for all of the hurdles that ail researchers. Still, Radical Numerics acknowledges the biggest risk: That a tool of such power can be used to tip the scales in a dangerous direction, if not controlled. 

LINKS

• OpenAI, Getty Images announce licensing deal for images in ChatGPT

• Micron, Anthropic ink strategic partnership to scale AI infrastructure

• JD CEO says delivery workers will "sooner or later" be replaced by robots

• Google invests $75 million in studio A24, forges AI film tools partnership

• SpaceX, Reflection sign computing deal worth up to $6.3 billion

• Chip firm Groq raises $650 million to expand data center capacity

• Sakana Fugu: An LLM trained to call on other models in an agent cool, designed to orchestrate other models in complex, multi-step tasks. 

• Grok Imagine: xAI's flagship model now has multitasking agents.

• HappyHorse 1.1: Alibaba's latest model for production-ready video synthesis.

• ElevenLabs' Ad Engine: Users can now connect Google, Meta, and LinkedIn ad accounts to ElevenLabs

• Anthropic: Applied AI Claude Evangelist, Startups 

• Google: Image Processing Engineer

• Motorola Solutions: Senior Machine Learning Engineer

• Cognition: Research, Post-Training

GAMES

POLL RESULTS

Would you purchase a pair of AR glasses?

Yes (20%)
Maybe, but not yet (35%)
No (39%)
Other (6%)

CULTURE

How AI changes what it means to be a filmmaker

On Thursday, I sat in a small theater in Santa Monica among hundreds of other viewers and had the distinct realization that Hollywood doesn't know what's about to hit it. 

The Deep View was invited to Runway's fourth annual AI Festival, showing off the ten best AI-generated films from hundreds of submissions. The short films spanned a wide range, some animated, some stop-motion, some realistic (because I hesitate to use the phrase "live action"), ranging from body horror to emotionally pensive to just plain outlandish.

There were, of course, some of the usual giveaways. Mouth movements that didn't line up quite right, over-smoothness of skin and motion, and a consistent inability to generate legible text. But walking out of the theater, I realized just how far we've come from Will Smith eating spaghetti. 

"It's just been incredible to see how fantastic the tools I've gotten that actually are usable," Dave Clark, cofounder of AI film studio Promise and creator of Tairell Isn't Real, one of the films shown at the festival, told The Deep View. "Now I can actually cut in a 4K generative AI shot against something I shoot on a camera." 

Hollywood has long been at war with itself over AI. Many artists and creatives have fiercely rallied against it, and major awards institutions are navigating how the tech affects eligibility for the industry's most prestigious honors. Some are trying to come up with solutions, such as the "Human Consent Standard," to give actors more control over how their likeness is used by AI. 

However, AI's use in entertainment has gone far beyond fringe technoevangelists. Runway, for instance, has a partnership with Lionsgate, which the companies expanded last week to create a joint development program to roll out a "slate of co-developed projects blending AI and content." Other high-profile Hollywood voices have also taken an interest in the tech, including Martin Scorsese's highly contentious decision to become an adviser to Black Forest Labs. 

The argument often made by companies like Runway is that AI is meant to raise the floor. Those that didn't have access to the tools needed to manifest their ideas now have the ability to create something out of nothing, Jamie Umpherson, chief creative officer at Runway, told The Deep View. Meanwhile, those who are already established in their careers have an entirely new arsenal of tools at their disposal. 

"It's really opening the door to a lot more creatives and filmmakers to get their first chance to get a project off the ground that might have been sitting there for 10 years," said Umpherson. 

And despite the fears that AI will put creatives out of work, in a panel with Runway CEO Cristobal Valenzuela before the screening, Roger Avary, award-winning screenwriter and partner in the production company General Cinema Dynamics, said, "The only people who are really going to lose their jobs are the gatekeepers."

"I think we're seeing a resurgence right now in the cinemas of indie filmmaking, and I think using AI tools ... people that come together to make something that is just the glimmer in their eye," said Gala Avary, a partner in General Cinema Dynamics, on the panel. "Suddenly, things that weren't possible, like locations and permitting, all of a sudden you're able to do those things with these AI tools."

As major Hollywood studios take fewer risks, smaller indie films are gaining traction among audiences, as evidenced by the massive popularity of films like "Obsession" and "Backrooms" compared to big-budget, blockbuster Disney and Amazon movies that came out around the same time. It's entirely possible that, in the best-case scenario, AI will enable more of those ideas to come to life. At the same time, money is money, and if these major studios can save some of it by replacing creatives with AI, they will likely do so. Still, Runway's AI film festival left me brimming with even bigger questions than the economics of it all, wondering, as a creative myself, what it means to be an artist when a machine can do the creating for you. The biggest of them: If AI tools make everyone good from the jump, will virtuosity become obsolete? Is it still important to be bad at something in order to appreciate how hard it is to get good at something?

TOGETHER WITH IBM

Guiding AI with project-level rules

AI agents are reshaping development, reportedly generating an average of 48% of code for surveyed organizations. But with that shift comes growing concern: 55% of engineering leaders surveyed are concerned about losing shared understanding of how their codebase evolves, and 39% are worried about shipping with confidence. 

The issue is not that AI produces bad code. It is that each agent makes different decisions on frameworks, testing and patterns. Over time, this can create inconsistency that’s harder to review and maintain. Project-level rules provide a structured way to address this, encoding conventions and standards directly into workflows so AI-generated code remains aligned with how teams build and maintain software.

See how to help keep AI code consistent

WORKFORCE

AI impact on jobs is overstated, says IBM CHRO

Workers are anxious that AI is coming for their jobs. But some HR leaders say those fears may be unwarranted.

AI will eliminate relatively few roles overall, according to Nickle LaMoreaux, chief human resources officer at IBM. While highly administrative and process-driven jobs are likely to disappear, she says the productivity gains from AI will create entirely new roles that don't yet exist.

"As economies grow, we're going to need more jobs," LaMoreaux told The Deep View at New York City Tech Week earlier this month. "AI will take a portion of that, but because we're growing, we still need human jobs to fill that other space."

It's too early to know exactly what those jobs will look like. Instead of focusing on which roles AI might replace, LaMoreaux says companies should think about how employees can use the time freed up by automation. 

That optimistic outlook may offer little comfort to early-career workers.

"If you now have the capacity freed up, can you deploy those employees into new models, new customer needs, new products, or honestly just a long list of work you never get to in your organization?" LaMoreaux asks.

Entry-level roles exposed to AI, including customer service, data entry, and coding, are in "substantial decline," according to recent research from Stanford University's Digital Economy Lab. The study found that knowledge workers with college degrees could face some of the biggest disruptions.

Even so, LaMoreaux says recent graduates can take steps to stand out. Her advice: focus less on responsibilities and more on outcomes. Whether you managed a budget for a volunteer organization or published a research paper, be prepared to explain what you accomplished, how you achieved it, and the impact it had. That principle also applies to AI literacy, a skill employers increasingly want to see in candidates.

Highlighting how you used AI to solve a specific problem, such as building an agent to automate a task, is more compelling than simply listing proficiency with ChatGPT on a résumé.

"AI literacy is important, but I think it's less about knowing a specific tool and more about your application of it," LaMoreaux says. 

Regardless of industry or job title, soft skills will remain difficult to automate. Communication, leadership, judgment, and conflict management, LaMoreaux says, "will not be taken over" by AI.

AI's impact on jobs is becoming a growing source of anxiety for workers. As companies lay off workers and cite AI as a way to cut costs, employees are left trying to figure out what comes next. Some workers are changing careers, while college students are switching majors in search of jobs they believe are less vulnerable to automation. But many of these decisions are being made before AI's long-term impact on work is fully understood. History suggests technological shifts create new jobs. However, AI raises a different question. If AI can perform most of a job's tasks, what work is left for humans? The real challenge may not be whether AI creates new jobs, but whether it creates the kinds of jobs people actually want.

TOGETHER WITH CDATA

Single-source MCP servers fail 1 in 4 prompts

Enterprise AI failure patterns are consistent: schema mapping, date logic, multi-filter conditions, and write validations.  

Without a governed context layer, integrating single-source MCP servers becomes a nonstarter at scale for the exact operations enterprise workflows depend on. 

CData Connect AI resolves all four failure patterns  with governance built into the architecture: role-based access control, per-user authentication, downscoped toolsets, and query-level audit trails.  

See how Connect AI handles what single-source servers can't, now offering 17 new sources including GitLab, Gong, BambooHR, and Zoom, on top of hundreds already supported. 

Start connecting for free ->

CONSUMER

Why AR's comeback now runs through AI

A decade ago, Google Glass, Microsoft HoloLens, and Magic Leap made augmented reality (AR) devices look inevitable. Then Meta bet its future on the metaverse. Neither paid off, but AI may be the catalyst AR needed.

The appeal of AR is simple: wear a device that layers graphics like messages, directions, and information over your environment, no other device needed. At a time when people are actively trying to spend less time on their devices and out in the real world, this should have been a no-brainer. So why hasn't it taken off? It comes down to three factors: utility, form factor, and price.

Convincing people that they can ditch their laptop or phone for something to wear on their face takes a lot of mental reframing. That is especially difficult to do when these devices are often bulky, heavy headsets that, in most cases, require another dangling accessory to be tethered to, such as a computing puck or a battery. Even for the most experimental of users, paying to play can be a bit difficult as the price tags are high, with headsets like the Apple Vision Pro retailing for $3,500. 

But to fix all three of these issues, the first step may be to convince people of the utility of AI. 

There has been an increase in demand for smart glasses in the past year, influenced by people wanting an AI assistant to break out of just a chatbot interface on a device screen and into the real world. According to IDC data, smart glasses without displays surged 167% year-over-year in the first quarter of 2026, with approximately 2.25 million units sold. That has trickled into AR, ER, MR, and VR categories, which the same data says have grown 86% year-over-year.

While smart glasses without displays have been shown to provide AI with more context via cameras and microphones, AR can take this much further. Not only can you give the device context about the world through its cameras, but you can also see the results overlaid directly in your environment. For instance, asking something like 'How would this piece of furniture look in my living room?' would show you the answer anchored in your actual space, in real-life proportions.

The rise of AI has also pushed manufacturers to develop more capable wearable processors, like the Snapdragon AR Elite that Qualcomm unveiled at AWE. As those chips get smaller and more powerful, so too could the headsets that house them, directly tackling the form factor problem.

We saw a glimpse of this at AWE with both the XREAL Aura and the new Snap Specs, which mimic the form factor of regular glasses, albeit chunkier, ditching the traditional headset entirely for AR experiences.

As demand grows and companies iterate, prices for AR glasses will naturally drop, and the hardware will improve. That will mean longer battery life, lighter frames, and all-day wearability, potentially giving AR the mainstream moment it has been waiting for. The Snap Specs offer the biggest meaningful step toward that reality, but it's still several steps away from being a product anyone will buy to replace their phone. Still, the path ahead has never been clearer, thanks to breakthroughs in chips, displays, and general miniaturization of technologies. And, AI is likely to continue to fuel new features to make smart glasses more useful and hardware advances that will accelerate progress toward AR.

LINKS

• Europe 2031, a speculative experiment, predicts EU instability in AI race

• Google DeepMind's John Jumper departs for Anthropic after nine years

• Norway sets a near ban on generative AI use in elementary schools 

• Amazon MGM drops movie on Sam Altman after OpenAI partnership

• Barret Zoph is leaving OpenAI again after five months back

• Anduril explores establishing operations in Israel

• Claude Design: Anthropic's design tool now remembers your context and stays on-brand across projects.  

• ChatGPT: OpenAI's flagship chatbot updated the way users take and upload photos on iOS. 

• Cohere North Mini Code: In an update, the company released 4-bit quant, its first open-source agentic coding model small enough to be run on a Mac. 

• Codex: OpenAI's coding platform can now hand off project threads between local and remote hosts.

• OpenAI: Researcher, Trustworthy AI

• Anthropic: Research Engineer, Safeguards Labs

• Salesforce: Adversarial AI & Research Engineer

• Google DeepMind: Research Engineer, Security and Privacy

GAMES

How Sphinx built an enterprise AI accuracy layer

Accuracy is one of the biggest problems plaguing enterprises as they seek to embed AI within their operations. The problem isn't only on the systems themselves, but rather the data they work from. 

It's the exact problem that Sphinx.ai wants to solve. Sphinx, a seed-stage startup backed by Lightspeed and Bessemer Venture Partners, enforces data accuracy across any AI tool an enterprise uses, giving companies the freedom to leverage whatever model they want without worrying about answers that drift from reality. 

"While AI is quite good at reasoning over things that are in the public domain that it's been trained on, every company has its own idiosyncratic way of laying out information," Rohan Kodialam, CEO and co-founder of Sphinx, told The Deep View. "It's not a question of intelligence, it's really a question of not knowing that tribal knowledge of how your organization works that limits AI's effectiveness."

Last month, the company unveiled Sphinx 2.0, the latest product in its lineup aimed at improving AI accuracy.

• This system essentially creates an institutional knowledge layer through which every AI query is filtered for accuracy. It sits between your AI tools and your data, continuously monitoring and refining outputs to better reflect your business context. 

• It's a stark change from Sphinx's first product, which used an agent to adjust and fix organizational data to prepare it for AI. Now, rather than making your data AI-ready, said Kodialam, Sphinx gets "your AI ready for your data." 

• The system also detects discrepancies and automatically surfaces them for human review by learning how an enterprise measures and tracks specific data. "We will find those instances where what we're seeing is in disagreement with what we already believe to be true," Kodialam said. 

Sphinx 2.0 solves a major pain point that enterprises are grappling with amid AI adoption: the inability to trust or rely on the outputs of these models. With too little data and organizational context, these models tend to hallucinate when queried about business-specific topics, making things up due to gaps in their training data. However, enterprise data is a moving target, with data stockpiles growing at a constant rate, while continually migrating it safely into AI systems is costly. 

Additionally, if enterprises simply connect all their data to AI systems in real time, they lose the observability and governance layers that make it safe. 

"We think there's like a happy medium of not spending a huge amount of money getting your data ready, but preparing your AI to operate in your real ecosystem," Kodialam said.

See how it works

WORKFORCE

Why tech companies jumped the gun on AI layoffs

AI is undoubtedly flipping the job market on its head, but not in the way that you might expect. 

Recent research from PwC found that companies using AI to enhance and augment their employees are being rewarded more than those simply replacing large chunks of their workforce with automation. 

The research finds that AI use in roles that are "professionalized," or used by experts to act as a force multiplier for skills, are seeing greater headcount growth than "democratized" roles, or those in which AI makes a skill easier for a non-expert to perform. 

• Additionally, AI-exposed companies are seeing faster headcount growth than those that are less exposed, 52% compared to 36%, as well as higher wage growth, 24% compared to 17%. 

• And demand for AI skills is skyrocketing, with jobs requiring AI-specific skills growing eight times faster than jobs that don't. 

• Still, the landscape is growing thornier for young workers. PwC said AI-exposed entry-level jobs are now seven times more likely to require "human-intensive" skills traditionally associated with senior-level employees, including strong judgment and leadership skills. These "seniorized" roles grew 35% since 2019, while entry-level jobs dropped off 10%. 

"The companies seeing the greatest returns on AI are using it to amplify human expertise, accelerate innovation and create entirely new sources of value," Joe Atkinson, global chief AI officer of PwC, said in the release. "As a result, they are pulling further ahead on productivity and growth than companies that focus primarily on automation." 

The fear around AI's ability to automate jobs is growing more feverish by the day. It's not helped by the fact that new studies about the tech's impact on the job market continue to provide conflicting reports. The forecasts are creating panic among workers, too, who largely use the technology in their jobs despite resenting it and even lying about their proficiency with AI.

The growing cacophony of contradictory data around AI's impact on jobs isn't the only thing creating a sense of dread about the labor market. There's been a layoff bloodbath in 2026, as some of the biggest proponents of AI, including Meta, Amazon, Microsoft, Block, and others, have slashed their workforces while plunging more cash into their AI transitions. One estimate reported this week found that 40,000 people were laid off in tech last month, the highest single-month total in two years, with AI the most cited reason. Though it's possible that AI was simply a popular scapegoat for layoffs that needed to happen anyway, these cuts set a precedent that job automation is a silver bullet for delivering strong returns to investors. However, PwC's data may indicate that these companies were a bit too trigger-happy in automating their staff away, and companies looking to follow in their footsteps may be better off creating ways for their employees to thrive, experiment, and enhance or expand what they already do.

TOGETHER WITH SALESFORCE

Salesforce put an AI agent inside your CRM, free for 30 days

The small business CRM now does the admin you've been avoiding by itself. Try the whole thing free with no credit card:

• Drafts your follow-up emails in seconds, so nothing slips

• AI summarizes every deal for instant pipeline visibility

• Logs your activity and updates records for you (no more "update the CRM!" pings)

• Answers anything about your customers in plain English so nobody is in the dark

• Up to 10 users, live in minutes

Start your free 30-day trial →

CONSUMER

Snap Specs are a breakthrough step, not a must-buy

I have covered smart glasses for years, tried the new Snap Specs, and have spoken to the CEO, Evan Spiegel. Here are my thoughts on where they fit in the market.

Let's start with their look. Compared to smart glasses like the Meta Ray-Bans, the Even Reality G2s, or any everyday glasses, there is no denying they look significantly more chunky. However, none of the glasses we just mentioned are true AR glasses, in which virtual objects are anchored to your physical environment. That's reserved for full-on headsets. 

So now let's compare the look of Specs to the Apple Vision Pro, the Meta Quest, or the Samsung Galaxy XR. The Specs are much less restrictive and more comfortable than those, especially considering there is no tethered battery pack or puck. 

Even the Meta Ray-Ban Displays, which do have waveguide displays that project color directly into the lens and weigh half as much as Specs, only offer one-dimensional graphics, giving you visual aids but not the same AR experiences that Specs promises, such as a 115-inch cinema screen or a 24-inch monitor placed into your environment.

Now, in terms of weight, at 132 to 136 grams, they are not meant to be worn all day, unlike smart glasses. Rather, they are supposed to be session-based: you use them for work, to watch a movie, or on a plane, and then you take them off. For context, other headsets like the Apple Vision Pro weigh closer to 600 grams and have the same utility. 

That leads me to my next point: price. At $2,195, they are obviously not meant for everyone, as those other headsets are not either. When I spoke to Evan about the price point, he gave me the same response I got from Samsung when I asked about their headset release last fall. This is meant for "early adopters," meaning developers building apps for it, or people who love the idea so much that they're willing to pay a premium.

So, should regular consumers pre-order them? Probably not, unless you are one of those people. In the future, I'm sure they will get lighter, look sleeker, and hopefully become cheaper, but what makes them impressive right now is that, in the current market for what they offer, they are leading the space, if they can deliver on their promises.

For instance, at the same event, Google launched the XREAL Aura glasses, which serve the exact same purpose as the Specs, but they also look a bit strange thanks to their birdbath display tech and bulky battery-and-processor attachment. Also, the price of these has not even been revealed.

Ultimately, it is really cool that Snap has crammed so much power into a footprint even close to that of normal glasses. Are we at a point for mass adoption? Absolutely not, but it's an important first step. For AI aficionados, this could also be a win for you, as there could be a future in which glasses provide AI assistance overlaid onto your actual environment with spatial awareness. That could be very powerful and useful. So for now, it's best to think of Snap Specs not as a more powerful pair of Meta Ray-Bans, but as an impressively smaller and less expensive version of Apple Vision Pro, and an interesting step toward the AR glasses of the future that could replace our smartphones.

TOGETHER WITH DEEL

The AI workforce is here. Is your policy ready?

Governments are racing to regulate AI. Entry-level roles are shrinking. AI hiring has tripled since 2023.

Deel's new policy report breaks down how global AI regulations are reshaping jobs, compliance, and workforce strategy—and what leaders need to plan for next.

Download the report

ENTERPRISE

How Databricks turned openness into an AI moat

If you're clinging to the idea that you can't build a great business on an open ecosystem in the AI era, then Databricks is rapidly making that line of thinking obsolete. 

The company started in 2013 as a platform built around the open-source engine for processing big data, Apache Spark, which the founders created at UC Berkeley. Today, it's a platform where companies organize their data and build AI on top of it. It's valued at $134 billion, and it's headed for a blockbuster IPO soon. Over 60% of Fortune 500 companies use the platform. 

What may be most impressive and surprising is that even as the company has grown its customer base and scaled its revenue to the moon, it has never shed its commitment to open systems. If anything, it's accelerating its commitment to openness. It used the Databricks Data + AI Summit this week to launch an open stack for agentic AI.

That included a flurry of developments:

• OpenSharing: This is a platform that allows organizations to share AI components across clouds and platforms, including AI models, agents, AI skills, and AI applications. The goal is to help organizations avoid lock-in on proprietary platforms. OpenSharing is an OSS project hosted by Linux Foundation. 

• Omnigent: This is a meta-harness, or a harness for harnesses. In other words, it's a powerful agent for managing, monitoring, and controlling your various agents. And Databricks announced that it's making Omnigent open-source. 

• DBRX for open LLMs: DBRX is an open-source LLM from Databricks that it claims outperforms all existing open-source models at launch, a big claim. The goal is to raise the bar for what organizations can build without relying on proprietary AI. Because the full model is open, teams can fine-tune it or use it as a blueprint to build their own custom models.

Jonathan Frankle, chief AI scientist, told The Deep View, "[Databrick founders] Ali [Ghodsi], Matei [Zaharia], Reynold [Xi], and all of our leaders … have worked really hard over the past decade plus to just make sure things are open… It forces us to be on the ball every single day, and it makes our customers feel really comfortable being with us, because they know that they're not trapped here. I think it makes it easier to win business, and it forces us to work really hard to keep business."

In a question and answer session with the press at Databricks Data + AI Summit on Tuesday, Ghodsi said, "Our roots are open-source. Our roots are community. So we want to make sure that we can actually offer that we are independent, we are multi cloud, we work on top of all the clouds, we work on top of all the AI’s. We partner with all of those vendors and we work with open-source. So we think this is an important function in the market, and we hope to provide that."

The bigger question is whether Databricks' open ecosystem strategy is an outlier, or if the company is on to something. Could its approach have broader merit and point the way toward an emerging trend? With all of the big tech companies releasing their own agents and their own frameworks for managing agents, it's all getting very confusing and siloed. More interoperability and standard would allow organizations to take advantage of the best of breed in terms of models, harnesses, and other AI modules. That's especially valuable since new breakthroughs are happening at an incredible pace across the AI industry, and it's unlikely to slow down.

LINKS

• Former Trump adviser Dean Ball to join OpenAI for policy work

• Microsoft makes headway selling AI to Chinese companies

• Google AI engineer Noam Shazeer joins OpenAI

• General Intuition in talks to raise $300 million at $2 billion valuation

• Waymo recalls nearly 4,000 robotaxis after several drive into construction

• Amazon is in talks to sell Trainium AI chips to third-party data centers

• Perplexity Brain: Now in Perplexity Computer, Brain is a continuously learning memory system that plugs every task into a context graph.

• HappyOyster 1.0: An open-ended world model for real-time world creation and interaction.

• Claude Code: The coding platform can now capture work progress as an artifact. 

• Codex Record & Replay: Users can now teach Codex a recurring task, like filing an expense report or submitting a time-off request.

• Meta: Research Scientist, Post-Training (Tech Leadership)

• Google: Senior Research Scientist, Google Research

• Palo Alto Networks: Principal Security Researcher (AI-Assisted Vulnerability Research)

• Adobe: 2026 Intern - Research Scientist/Engineer

GAMES

POLL RESULTS

Should the public have a direct vote on major AI regulations?

Yes (62%)
No (20%)
I’m not sure (12%)
Other (6%)

POLICY

AI’s future may hinge on who writes the rules

Some of the most prominent voices in AI have strong ideas about the way the tech should be governed. 

At the G7 Summit in Évian-les-Bains, France, on Wednesday, Anthropic CEO Dario Amodei and Google DeepMind founder Demis Hassabis called for a US-led coalition to shape standards for AI, according to CNBC and The Financial Times. 

The tech leaders proposed that this coalition be an international effort with the US leading the conversations around AI's risks. They warned of the tech's dangers, including major bioterrorism and cybersecurity risks, if an alliance couldn't be formed to preempt them.

• The proposal was supported by OpenAI's Sam Altman and Canadian Prime Minister Mark Carney, according to media reports. 

• According to the FT, Amodei told G7 leaders they must “resist the temptation to splinter” over the deployment of advanced AI tools. 

The proposal comes days after Anthropic's latest messy dispute with the Trump Administration over the release of its latest models, Mythos 5 and Fable 5. The US government imposed export controls on the models, restricting their use outside the country and by foreign nationals, citing cybersecurity and national security concerns. 

As a result, other government leaders have expressed their concerns: According to the FT, French President Emmanuel Macron said that if the US can "turn off the switch" on powerful models, the impacts could ricochet back on the AI companies leading the market. Indian Prime Minister Narendra Modi also expressed his concern, reportedly saying that democratic nations must have access to leading AI models to safeguard infrastructure.

Macron also warned against the risks of “non-cooperation among democracies" as it relates to AI regulation, calling for stronger regulation to "jointly define common standards."

If major model providers are to be believed, AI will be a fundamentally transformative technology for society. I've written that the power of this technology has concentrated in the hands of an elite few companies, which has created risks regarding who exerts influence over this societal transformation. It's something that the leaders of these companies themselves have even started to recognize, with Altman, Amodei and OpenAI's Greg Brockman all pointing towards the risks of power centralization. Still, despite these leaders calling for international cooperation to temper the risks, the question of how this tech should be regulated is far from settled. If this technology turns out to be as transformative as these companies claim, it has the potential to have a broad impact on the global community. And that raises questions. Should the regulation of something so broadly transformational be placed in the hands of a small group of government and business leaders, or be voted on more democratically by the people it affects? And with technology changing and evolving so rapidly, how will any kind of regulation keep up? 

A QUICK POLL

TOGETHER WITH COLUMBIA BUSINESS SCHOOL EXEC ED

Finance professionals are using AI. Most don't know what they're actually doing with it.

The 8-week AI for Business & Finance Certificate Program from Columbia Business School Exec Ed + Wall Street Prep is built for non-technical roles and helps you cut through buzzwords, apply AI to your work, and lead adoption in your org.

• Apply AI directly to financial analysis, reporting, and decision-making workflows

• Collaborate in LIVE weekly office hours with Columbia faculty

• Earn a Columbia certificate and join a 5,000+ network of finance and business professionals

Use code DEEPVIEW to save $300 on tuition. Enroll today.

GOVERNANCE

Why open ecosystems may beat AI agent lock-in

Nearly all of the big tech and AI companies are racing to solve the problem of AI agents in the enterprise. 

The Deep View has covered the latest attempts from Microsoft, Snowflake, AWS, Anthropic, OpenAI, Perplexity, and others. 

The reason they are all racing to play a role is that AI agents are creating a multitude of problems:

• Data leakage

• Compliance issues

• Privacy challenges

• Security vulnerabilities

• Inaccuracies 

• Cost overruns 

But most of all, agents aren't living up to the promise of automating workloads and creating enough value to justify the massive token costs that agents are running up in most businesses.

And now, the fact that all the vendors are offering their own agent solutions is starting to cause a new set of problems due to lock-in and a lack of cross-platform compatibility.

"The SaaS providers are now each providing you an agent that they want you to use," said Databricks CEO Ali Ghodsi on Tuesday at the keynote of Databricks Data + AI Summit. "They have agent builders that they want you to use. They have their own MCP servers… So it's kind of a quagmire. It's a complete mess." 

Databricks' answer to the problem is, naturally, its own agent, called Genie. But the real answer is the principle that Databricks always seem to circle back to: open ecosystems. 

In an interview with The Deep View, Databricks chief AI scientist Jonathan Frankle said, "The Databricks team has worked really hard over the past decade to make sure things are open. The way I frame it to customers is that I have to win your business every day, because if I don't win your business every day, you can leave… You can just take all this data straight out of Databricks. We don't really have much lock in, and that's kind of the point."

The full Genie ecosystem was announced at Data + AI Summit and includes Genie One for querying data and figuring out tasks, Genie Ontology to give your AI the business context to solve big problems, Genie App Builder for vibe coding business solutions, Genie Agent for reusable, templated agents, and Genie Code for more advanced engineering workflows.

I have to hand it to Databricks for creating solutions that are very polished and easy-to-use while always leaning heavily on an open ecosystem foundation. But the enthusiasm of the community building around Databricks at the Data + AI Summit was what impressed me the most. The customers and partners I talked to loved that they aren't trapped or locked into the Databricks ecosystem. They also talked about Databricks as a partner and collaborator more than a technology vendor that has all the answers to their problems. That's not easy considering a lot of Databricks customers are businesses that have been around for decades and have a ton of legacy systems to manage.

TOGETHER WITH SHOPIFY

Five apps Shopify built within days

As part of the Spring ‘26 Edition, a small group of Shopify designers and developers built five shopping apps in a few days that make everyday activities shoppable, with commerce woven into the experience itself.

One takes a trip itinerary and creates a shopping list. Another suggests products based on your horoscope. All run on Catalog API to query billions of merchant products and Universal Commerce Protocol for the full commerce journey.

They show what's possible for developers, designers, and anyone with an idea.

See what else shipped.

CONSUMER

Evan Spiegel's bet on life after the smartphone

Snap CEO Evan Spiegel joins The Deep View Conversations to discuss SPECS, Snap's long-awaited augmented reality glasses and why he believes they represent a new era of computing.

Senior reporter Sabrina Ortiz interviewed Evan at Augmented World Expo immediately after the SPECS unveiling and Spiegel explained why Snap spent more than a decade building toward this moment, how SPECS differ from AI smart glasses and mixed reality headsets, and why he sees AR glasses as the future beyond smartphones.

Other topics covered include:

• Why Snap calls SPECS a "computer" instead of AI glasses

• How SPECS combine wearability with advanced spatial computing

• The role AI played in making consumer AR glasses viable

• Why shared experiences could become AR's killer app

• The challenge of competing with Apple, Meta, and other tech giants

• Snap's 12-year investment in augmented reality hardware and software

• The importance of developers in building the AR ecosystem

• Why Spiegel believes people are ready for an alternative to smartphones

• How AR glasses could make computing more human

Spiegel argues that after nearly two decades of smartphone dominance, consumers are increasingly looking for a more natural way to interact with technology. Snap's bet is that augmented reality glasses can bring computing into the world around us instead of pulling us away from it.

📺 Watch on YouTube.

🎧 Listen in your favorite podcast player.

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transistor.fm

LINKS

• Shoe company Allbirds rebrands as Smartbird amid AI pivot

• AI medical tool Mira, Google's Amie surpassed doctors at diagnostics

• Estonia will grant digital ID numbers to AI agents to control access

• Uber to expand premium robotaxi service to Houston next year

• AI military startup Twenty raised $100 million at $1 billion valuation

• World models startup Odyssey raises $310 million Series B

• Vercel Eve: An open-source agent framework for building, running, and scaling agents.

• Unreal Engine 5.8: The 3D design platform now has experimental MCP server support. 

• Claude Design: Anthropic's AI design tool has gotten an overhaul, including a fix to how fast it burns through tokens. 

• ChatGPT: OpenAI's flagship chatbot now has a faster, more reliable, and easier to manage scheduled tasks.

• Meta: Tech Lead, AI Research Scientist (Robotics)

• Anthropic: Research Engineer, Knowledge Team

• Tencent: Hunyuan Multimodal Algorithm Researcher (Omni-Modal)​​

• Google: Image Processing Engineer

GAMES