The AI agent boom and coming cybersecurity crisis

In this episode of The Deep View: Conversations, Senior Reporter Sabrina Ortiz sits down with Jeetu Patel, President and Chief Product Officer at Cisco, to explore how the rise of AI agents is reshaping cybersecurity, software development, and enterprise strategy.

Jeetu makes a bold prediction: today’s 150 million developers could expand to 3 billion agent builders within the next year. But that explosion comes with serious risk. As nation-states and bad actors deploy autonomous agents at scale, traditional, human-centered security models begin to break down.

This conversation unpacks:

• Why AI agents are becoming the new attack surface

• What enterprises must do now to prepare for agent-driven threats

• Jeetu’s journey from Box to Cisco, and what it taught him about leading through platform shifts

• Practical advice for learning AI and building in an agent-first world

But this isn't a doom-and-gloom conversation. Jeetu lays out a vision for how security can become an accelerator rather than a limiter, and why the distinction between giving agents access and giving them trusted, governed access will define which enterprises thrive in the agentic era.

If AI agents are the next platform shift, cybersecurity may be the defining battleground.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player

Subscribe to The Deep View: Conversations podcast in any podcast player for more interviews with the leaders shaping the future of AI. You can also subscribe on YouTube.

BIG TECH

How agents deepened AI's compute crisis

The talk about AI being a bubble-driven hype cycle has largely stopped. Now the problem is that it can't keep up with demand. 

A perfect storm of factors has conspired against the AI industry, turning its compute shortage into an emerging crisis:

• The war in Iran is disrupting tech products: The conflict is snarling global supply chains and threatening to unleash a new wave of inflation. One of the key elements of the semiconductor supply chain, helium, is a victim of the crisis. And that could affect many tech products, including those needed to build AI data centers. A prolonged conflict will continue to impact oil prices, driving inflation in logistics, shipping, and travel needed for the tech ecosystem to thrive.

• AI agents are gobbling up tokens: The boom in AI agents over the past 90 days has triggered insatiable demand for inference, also known as the tokens needed to run AI queries and tasks. The Wall Street Journal reported that OpenAI's token demand jumped from 6 million per minute in October 2025 to 15 billion per minute in March. Never mind the fact that many of those tokens are being used inefficiently by users and their agents, as OpenAI's Peter Steinberger alluded to this week.

• GPU, memory, and storage are scarce: Even before the recent supply disruptions, the tech industry had been facing a deepening shortage of GPUs, memory chips, and storage drives, all key components needed to build the machines that run AI. The shortages are now leading to price hikes, which could price out startups and leave AI industry giants gobbling up most of the components. Case in point: the Ornn Compute Price Index estimates that the cost of an hour of Nvidia GPU time has increased by 48 percent in the past months alone.

All of this is happening against the backdrop of Anthropic announcing Mythos, its most powerful model ever, while initially limiting access to a small number of partners. And OpenAI is preparing to launch its Spud model, which is also considered a step change, prompting the company to release a policy paper to spark public debate about the AI's impact on jobs and societal well-being. 

Anthropic has chalked up its motivation not to release Mythos more broadly to cybersecurity risks, claiming the model is "too powerful." I believe it's also clear that the company does not have the compute to run it.

I recently heard directly from an executive at one of the AI frontier labs that every time the company releases a new model with greater capabilities, usage spikes because people can do more with it. Anthropic's Claude has already been buckling under the demand of a new influx of users since its February showdown with the Pentagon. There's no way that it's ready to give users rocket fuel to do more with Mythos. And that puts Anthropic and all of the AI labs in a tricky spot as they prepare to grow in the months ahead. There are no guarantees that the conflict in Iran will be resolved soon, but even if it is, the agent situation and the shortage of key components are likely to continue to limit AI's growth options. What will that mean? I expect we'll be talking about that a lot in the days ahead. You can follow my updates in real-time on Twitter/X.

TOGETHER WITH ORACLE NETSUITE

Unleashing AI's Value in Finance

As AI reshapes the world of finance, the need for CFOs to strategically harness AI has never been greater.

AI expert Glenn Hopper outlines a clear roadmap for deploying AI in his latest guide.

Download your copy now to learn how to leverage AI’s potential—no computer science or engineering background required!

PRODUCTS

Anthropic gives Claude Code a multi-agent upgrade

The AI race moves fast, but frontier labs move faster. Google DeepMind, OpenAI, and Anthropic all made big news on Tuesday, with Anthropic doing so twice.

Anthropic redesigned Claude Code on the desktop, now tailoring the interface to better accommodate running parallel agents and tasks. The changes include a new sidebar for managing multiple sessions, a drag-and-drop rearranging feature, and an integrated terminal and file editor. 

Twitter tweet

The sidebar surfaces all active and recent sessions, letting developers easily navigate across repositories, filter by status, project, or environment, and group related sessions together. When a session's associated pull request merges or closes, it automatically drops off the sidebar, keeping the view focused on what's still in progress.

As Claude Code cements its position as the leading agentic coding environment, seamless window management has become a necessity for developers juggling increasingly complex, multi-threaded workloads. It's a shift Anthropic elevated in its blog post:

"For many developers, the shape of agentic work has changed. You're not typing one prompt and waiting… The new app is built for how agentic coding actually feels now: many things in flight, and you in the orchestrator seat."

Other elements of the redesign focus on bringing more to users within the app, including: 

• View modes: Verbose, Normal and Summary options allow users to decide how much they want to see of the actual process, with the option to just see results. 

• Keyboard shortcuts: New shortcuts make it easier to get things done, and users can press [⌘ + / ] or [Ctrl + /] to see the full list

• Speed: The app was rebuilt for reliability and speed, according to Anthropic

• Parity with CLI Plugins: The desktop app now fully supports CLI plugins and SSH connections to remote machines for a more familiar experience, including Mac 

The DeepView CEO, Faris Kojok, tested all the features and found the new UI/UX more refined and easier to navigate. In particular, he found View Modes to be helpful. 

“Being able to switch the transcript mode from Verbose to Normal to Summary allows me to get just the high-level response and then dive deeper with Verbose if necessary,” said Kojok. 

Anthropic also released Claude Code Routines in research preview, which allows users to automate a session by configuring it once and then running it on a schedule, via an API call or an event trigger. The week before, Anthropic released Claude Managed Agents and an Advisor tool.

Anthropic’s slew of releases focused on Claude Code, which makes sense for several reasons. Primarily, it aligns perfectly with Anthropic’s goal of best serving its enterprise audience, and this product is one of its biggest profit generators at the moment. More interestingly, though, it seems the company could be strategically positioning itself to better compete with OpenAI, which is rumored to be developing a superapp that would merge Codex, Atlas and the ChatGPT app. That would enable OpenAI to combine the strongest features of all the apps into a single hub, including the ability to run Codex agents in multiple Atlas browser tabs. Since Anthropic doesn't have a browser, its latest Claude Code update addresses the same need for agentic multitasking. 

TOGETHER WITH ATTIO

Attio: The AI CRM Built for Modern Businesses

Attio is the AI CRM you can actually build on. Connect your AI tools, customize it to how your team works, and let it act on every signal it finds.

With Universal Context, Attio's intelligence layer, your agents always have the context they need to move things forward, instantly.

Ask more from CRM. Ask Attio.

Try Attio for free

BIG TECH

Nvidia bets on open models to advance quantum

Quantum isn’t ready for the big leagues yet, but Nvidia thinks it can help.

On Tuesday, the chip giant debuted Ising, a family of open models that aim to speed up the development of useful, scalable quantum computers. In a press release, Nvidia claims that Ising offers “the world's best AI-based quantum processor calibration capabilities" with error correction decoding 2.5 times faster and three times more accurate than conventional methods. 

In a statement, Nvidia CEO Jensen Huang said that AI is “essential” to making quantum computing practical. “With Ising, AI becomes the control plane—the operating system of quantum machines—transforming fragile qubits to scalable and reliable quantum-GPU systems.”

The Ising family includes new models, tools and data to accelerate quantum development: 

• The Ising Calibration model is a vision language model that can interpret and react to measurements from quantum processors, with agents continuously calibrating quantum computers without downtime. 

• The Ising Decoding model, meanwhile, offers two variants of a 3D neural network that perform error correction in real time more than twice as fast as current industry standards.

The Ising models are already being put to use in research labs and institutions, including Harvard University’s school of engineering, Infleqtion, IQM Quantum Computers, Lawrence Berkeley National Lab and the UK National Physical Laboratory. 

Nvidia’s addition to the quantum landscape marks a growing interest in the once-theoretical technology. Quantum, for instance, was a major focus of CES in January, with several sources telling The Deep View that quantum and AI will have a symbiotic relationship. 

“I can foresee a world where a quantum processing unit, a QPU, would coexist alongside the GPU and the CPU,” Dr. Pouya Dianat, chief revenue officer of Quantum Computing Inc., told The Deep View at the time.

Quantum computing is still a finicky technology with a ton of roadblocks. However, Nvidia is putting its weight behind quantum and feeding the ecosystem the same way it feeds open models. By developing foundational systems that make this tech work better, Nvidia is embedding itself into the foundation of quantum computing long before the tech reaches maturity. And given the current compute shortage and quantum's potential to ease AI's energy burden, it might not be long before the industry at large starts to take the tech more seriously. 

LINKS

• OpenAI acquires Hiro Finance, a personal finance company founded in 2024

• Novartis CEO Vas Narasimhan joins Anthropic’s Board of Directors 

• Anthropic opposes Illinois OpenAI-backed law that shields labs from liability

• Microsoft to rent data center capacity from OpenAI in Norway 

• NBC Review finds Grok AI is still generating sexual deepfakes

• AWS launches Amazon Bio Discovery for AI-powered research in life sciences

• Granola: Our entire team switched to Granola and nobody went back. It captures your meetings without a bot, and the notes are actually worth reading. Try it free with code THEDEEPVIEW. (sponsored)

• Claude Code: New Routines feature allows for repeatable routines

• MAI-Image-2 Efficient: Microsoft’s new, cheaper and faster image model

• Microsoft Word: Copilot now tracks changes, leaves comments, and more

• Microsoft Copilot: Users can forward emails to Copilot and delegate work 

• Cursor: The automations feature now supports even-based triggers for Sentry

• Meta: Computer Science Research - US - IC5

• Anthropic: Research Engineer, Reward Models Platform

• Baseten: Post-Training Research Scientist

• Reflection: Member of Technical Staff - Post-Training

GAMES

POLL RESULTS

Do you use an AI browser such as Comet, Atlas, Dia, etc.?

Yes, often (12%)
Sometimes (11%)
No (77 %)

PRODUCTS

OpenAI 'superapp' prepares for agentic leap

OpenAI's forthcoming "superapp" is one of the worst-kept secrets in AI right now. 

During a walk in San Francisco last month, I told The Deep View CEO, Faris Kojok, that the most challenging part of using OpenAI's products was switching between the three desktop apps. Apparently, I wasn't the only one thinking that, because less than a week later, reports began circulating that the company would merge the three apps into a single superapp.

Since then, the OpenAI team has been publicly discussing it on Twitter and in podcasts. The most prominent of which was when President Greg Brockman appeared on the Big Technology Podcast and told Alex Kantrowitz when and how the new superapp would be rolled out.

"We're taking incremental steps to get there. Over the next couple months, we should have shipped the complete vision … but it's going to come in pieces," said Brockman. "The place that we're starting today is with the Codex app, which is really two things in one. It's a general agent harness that knows how to use tools, and it's also an agent that knows how to write software. That general agent harness can be used for so many different things… So we're going to make the Codex app so much more usable for general knowledge work." 

Why would OpenAI bring the three apps together? First, let's look at what they do. 

• ChatGPT desktop app: This is the traditional ChatGPT app that has been around for the past several years. It makes it easy to view and search all of your past chats and organize them into projects. You can also quickly search for custom GPTs and apps to extend the chatbot's functionality by adding features from tools like Photoshop, Canva, insurance apps, and hiring apps. 

• Atlas web browser: In October 2025, OpenAI launched the ChatGPT Atlas web browser. Following in the footsteps of Perplexity Comet, launched in mid-2025, OpenAI built its browser on Google's open-source Chromium platform, which runs most of the web. Very quickly, Atlas became my preferred way to use ChatGPT for one simple reason: its ability to open multiple tabs, keeping several different queries quickly accessible. 

• Codex app: As Brockman noted, the app has two parts. Most people think of it as a coding tool. But the second part is a general-purpose agent that can carry out tasks for you. And that's where the combination of a web browser and an AI agent gets interesting. If you can open multiple instances of ChatGPT in tabs, then you can use your Codex agent to carry out various tasks, all at the same time. 

The concept of a superapp has been a common refrain in the tech industry for years. The first successful one was China's WeChat app, which combined text messaging, video calling, file sharing, digital payments, and mini apps that add ride-hailing, food delivery, and event ticketing. That's what Elon Musk wanted to create when he bought Twitter. But what OpenAI is doing is a little different: it's not on mobile, but on the computer, where knowledge workers do almost all their work. When the superapp combines the search and organizing capabilities of the traditional ChatGPT app with the multitasking environment of Atlas and the agentic automation of Codex, it could transform ChatGPT into a product much more akin to OpenClaw—but accessible to its 900 million users.

TOGETHER WITH CRUSOE

Your model. Our inference engine. Breakthrough performance.

Eliminate the "memory wall" causing inference bottlenecks. Crusoe’s inference engine is powered by MemoryAlloy™ technology to deliver larger shared memory capacity so you can serve more users at lower latency, with better throughput, and less wasted compute.

The result? Breakthrough time-to-first-token speed and up to 5x higher throughput. Work with our team to optimize performance for your own fine-tuned model so you can scale without compromise.

Work with our team to optimize your custom weights today.

CULTURE

Why Kurzweil still sees AGI coming by 2029

Last week's HumanX event offered no shortage of meaningful AI conversations, whether in panels, one-on-one meetings, or casual exchanges. One conversation, however, stayed with me.

Ray Kurzweil is an inventor, futurist, and one of the most consequential figures in the history of AI. His pioneering work on OCR and text-to-speech laid critical groundwork for the AI revolution unfolding today, and his books, including The Age of Intelligent Machines and The Singularity Is Near, have shaped how generations of thinkers and builders understand where technology is headed. 

He has also had a pretty remarkable track record of accurately predicting where technology will go next, and one prediction in particular has never stopped turning heads: in 1999, he predicted that human-level intelligence, what we now call AGI, would arrive by 2029.

In a panel discussion alongside his son Ethan Kurzweil, co-founder and managing partner of the venture firm Chemistry, and Shirin Ghaffary, AI reporter at Bloomberg, Ray shared further insight into where his prediction stands today and made clear that he still believes it will come to pass.

“I predicted that there would be a several-yearfold increase where people would start to face AGI, not everybody would, and everybody would face AGI a few years later, so that's consistent with that prediction that we would start now,” said Ray. 

Broadly speaking, AGI refers to artificial intelligence capable of understanding, learning, and reasoning at or beyond a human level. It's a prospect that has prompted serious concern across many quarters, from questions of job stability and the potential for weaponization, to deeper anxieties about humanity's role in a world shaped by that kind of intelligence.

Ray, however, holds a far more optimistic view. During the panel, he argued that the positive outcomes of AI will ultimately outweigh the negative, and that nearly every industry, including housing and food, will be transformed through AI breakthroughs. 

When asked about the rapid pace of discovery driven by fierce competition in the space, he welcomed that dynamic as well, noting that competition fuels exponential growth and that it is precisely that competition that will make the next ten years 'absolutely fantastic.'

“Even though some people hold a negative premise about [AGI], we’ve been able to create something that actually represents knowledge, and that's really fantastic: AI that can expand our knowledge,” added Ray.

During the panel, Ray drew on his 63 years of experience in the field to audible applause from the audience. That depth of history is precisely what lends his perspective such weight, particularly given that he predicted AGI long before it became a marketing term wielded by industry leaders to generate headlines, move products, or court investors with visions of a grander future. Whatever your stance on the term itself, his core argument, that AI will break through every industry, is already proving true and shows no sign of slowing.

TOGETHER WITH GRANOLA

The Deep View Team is Obsessed with Granola

We’re not talking about the food (although a few of us do have an unhealthy interest in that, too). We’re talking about the AI notepad that turns your chicken scratch into something elegant, organized, and downright beautiful.

Since day one, Granola has checked all the boxes we look for in a notetaking app, like… 

• Working seamlessly across our various  teams and workflows ✅

• Documenting, summarizing, and prioritizing all our discussions, so we can focus on what’s most useful and relevant ✅

• Having full context across all meetings for a holistic view of trends and patterns ✅

We estimate Granola saves us around 10 hours per week, per person – that’s no small feat. Try it for yourself and get your first month free with code THEDEEPVIEW.

RESEARCH

AI’s surge is widening gaps in trust and policy

The rapid pace of AI development can often feel like endless noise. Stanford University wants to provide some clarity. 

On Monday, the university’s center for Human-Centered Artificial Intelligence, or HAI, released its 2026 AI Index report, offering a comprehensive view of some of the most poignant trends impacting the AI industry. Big picture: AI is growing and transforming faster than ever, and no one knows for sure how it’s going to impact our future. 

Here are some key takeaways: 

• AI capability and adoption are growing rapidly, but safety measures can’t keep up. More than 90% of notable frontier models were released in 2025, with several meeting or exceeding human baselines on PhD-level work. Organizational adoption reached 88%, and 80% of university students use generative AI. But reporting on responsible AI benchmarks remains uneven, and “documented AI incidents” rose to 362 in 2025, up from 233 in 2024. 

• The US leads in some ways and lags in others. The report claims the US-China model gap has effectively closed, with Anthropic’s top model leading China’s by 2.7%. And while the US leads in data center density and AI investment, the country’s ability to attract talent has declined, with the number of AI researchers and developers moving to the US falling by 80% in the last year alone. 

• Adoption is on the rise, but formal education isn’t. Consumers are seeing substantial value out of AI tools they can access for free, according to the report. Generative AI has reached 53% of the population over the last three years, but is uneven across countries. The US, for instance, sits at a roughly 28% AI adoption rate. Though four out of five students use it in class, only half of middle and high schools have AI policies in place, and only 6% of educators say those policies are clear and concise. 

And broadly, the public still isn’t optimistic about how AI will shape the future. While 73% of AI experts have a positive sentiment towards how AI will impact their jobs, only 23% of the public shares that sentiment, the report finds.

Stanford’s research signals that AI is causing all kinds of cracks. There’s a growing divide between public trust and stakeholder optimism, and while adoption is on the rise, usage varies region to region. Meanwhile, hype and industry FOMO are pushing the tech beyond our ability to regulate it, leaving governments and safety organizations alike reeling as they try to keep up. Reports like these offer a snapshot of where AI stands at a given moment, but in an industry that shows no signs of slowing down, these figures could change month to month, week to week or even day to day. Nevertheless, Stanford's moment of clarity in this report is welcome.

LINKS

• Sam Altman’s home was targeted in a second attack 

• Meta reportedly builds an AI version of Zuckerberg to interact with staff

• OpenAI internal memo highlights alliance with Amazon 

• Microsoft exec says company explores “OpenClaw in an enterprise context”

• Anthropic may be adding an app-building tool like Lovable to Claude

• Apple reportedly works on display-free smart glasses

• Meta AI: The app soars in the App Store due to Muse Spark release

• Ollama: Ollama 0.20.6 was launched with "improved Gemma 4 tool calling” 

• NotebookLM: Certain Google for Education plans now have higher limits 

• ChatGPT: There are more paid plan options, broken down in the post linked

• Cursor: Version 3 got updates including agent splitting for multi-tasking

• Scale AI: Research Scientist, AI Controls and Monitoring

• Zoom: Trust & Safety Machine Learning Engineer

• OpenAI: Researcher, Training

• Anthropic: Research Engineer, Universes

GAMES

A QUICK POLL BEFORE YOU GO

STARTUPS

Biological chips could solve AI's bottleneck

Tucked into an unassuming office building in San Francisco, one startup is betting on an unconventional way of alleviating the AI energy crisis: living human cells. 

The Biological Computing Company, or TBC, a startup that emerged from stealth in February with $25 million in seed funding, is pioneering an alternative to traditional silicon, using living neurons as the foundation to improve generative AI algorithms and infrastructure. Sitting at the helm of TBC are former neurosurgeons, Dr. Alex Ksendsovsky and Dr. Jon Pomeraniec.

Last week, The Deep View was invited to tour TBC’s lab in Mission Bay, meeting the startup's team of 23, ranging from computer vision experts and AI developers to computational physicists and hands-on biologists, many of whom came from large tech firms like Meta, Apple and Amazon. Bundled in a lab coat, mask and latex gloves, I had the opportunity to get an up-close look at the chips themselves, each containing anywhere from 100,000 to 500,000 neurons. 

Twitter tweet

Here’s how it works: 

• TBC takes real-world data, such as images and videos, and encodes them into the neurons of these chips. This information is then decoded into richer representations, which are then used to bolster AI algorithms. As Ksendsovsky put it, “all of the data flows through the biology.”

• Each of the company’s chips contain a “multi-electron array” in which neurons, grown from stem cells that are “reprogrammed” into frontal cortex cells, are maintained. The company originally used neurons extracted from rats, but Ksendsovsky told me that TBC is “moving away from that.” The chips have a lifespan of a year, and create waste that needs to be cleaned out every few days. 

Though this might sound like science fiction, Ksendsovsky said that this technology harkens back to AI’s roots. Neural networks were originally named as such because they modeled the neurons in a human brain. But as AI developed, the technology “became very non-biological,” leading silicon to become “rigid” and forcing AI scaling to resort to brute-force, power-hungry methods, he said. 

This is where TBC’s chips may have the greatest advantage, he said: Brain cells, generally, require less energy than silicon chips to power intelligence. In its research, TBC discovered that models trained on “biological neural responses” reached peak performance three times faster, requiring fewer training iterations and implying a “threefold reduction in the computational and energy demands.” 

For now, the company isn’t selling the chips themselves, but rather is using the technology to build and strengthen algorithms that use neurological signals, primarily for visual AI, including generative video, game rendering and computer vision. Though he wouldn’t name names, Ksendsovsky said the company is in talks with foundational model labs, cybersecurity firms and data curation companies. 

“I think the most common misconception is that this is just a research project,” Ksendsovsky told me. “We are building products now. We want to show people that there are tangible results that can be monetized now. What we're building is so unique and so far out, that I think it's just going to take the world a little bit to see it and understand it.”

Eventually, TBC hopes to move to “real-time compute,” in which these biological chips themselves are a part of the “inference loop,” Ksendsovsky said, eyeing that milestone for five to ten years from now. However, a lot needs to be done to make that happen, including automating the collection of waste that these cells generate and figuring out how to “immortalize” their lifespans. 

“The reality is, there's a lot of foundational things that need to be understood and built to get to that point,” he said.

TBC’s thesis is undoubtedly outlandish. But AI, once deemed science fiction itself, has become so commonplace that its rapid scale has created a litany of issues in its wake, the greatest of which may be energy. And because of this, the leading voices in AI are grasping for technologies that also once seemed out of reach, such as fusion energy and data centers in space that harness the sun's power. Crisis is the mother of invention, and AI is unquestionably creating crises. The solutions may live within us, figuratively and literally, if TBC has anything to say about it. 

TOGETHER WITH ATTIO

Attio: The AI CRM Built for Modern Businesses

The landscape is changing fast, and every signal across your business needs to count.

Attio is the AI CRM that makes sure it does.

Connect your email, calls, calendar, and product data, and Attio instantly builds a complete picture of every deal and customer. No manual logging; zero missing context.

Then Ask Attio to plan your next move. Surface pipeline risks, prep for meetings, and run deep research on prospects without lifting a finger. Powered by Universal Context, Attio’s intelligence layer, every signal across your business becomes instantly actionable.

Ask more from CRM. Ask Attio.

Try Attio for free

HARDWARE

Why Anthropic may bring AI chips in-house

Anthropic may start making its own chips for the same reason Nvidia started making its own models. 

If you're not familiar with the term "extreme co-design" yet, then it's time to start learning. Nvidia executives used the phrase repeatedly at the company’s GTC event in March, and I expect it's a big part of the strategy behind the new Reuters report that Anthropic is exploring the opportunity to design its own AI chips. 

Extreme co-design allows you to architect the data center, chips, software stack, and LLMs in tandem, resulting in far better optimization and integration. The result can reduce the cost per token, lower latency, and minimize power consumption. 

That's why Nvidia is now making its own open-source frontier models: By building its own models, Nvidia learns how to better optimize ALL models for its AI chips.

Kari Briski, VP of generative AI software at Nvidia told The Deep View that the process allows the company "to break some rules, to see around the corner… When we say extreme co-design, it's literally about sitting down side by side, day by day, showing each other reports… because we're all racing at the speed of light to the next thing... [And] you have to be building it right now. It can't be a year from now. So we talk about extreme co-design literally in lockstep every single day, understanding each other's pains."

But beyond co-design, there are also a couple of other reasons why Anthropic is likely considering the move:

• Remove the constraint: Now that Anthropic has proven it can consistently build some of the world's best and most powerful AI models, one of the company's biggest constraints is simply getting enough compute to run its models, serve its chatbot and coding tools to new and existing customers, and train future models. That's why it's been signing new deals with companies like Google/Broadcom and CoreWeave. But chips themselves are one of the biggest constraints for hyperscalers and neoclouds (and leads me to ponder if Anthropic might also consider acquiring a neocloud at some point).

• Save money in the long run: Spinning up a chips program would be incredibly expensive and it likely wouldn't cover the company’s entire demand for chips. Anthropic would still end up buying chips from other vendors for the foreseeable future. However, in the perspective of decades, it could eventually mean Anthropic would control one of the most important drivers of its technology, and could save money by avoiding padding the profit margins of other companies (like Nvidia).

Let's keep in mind that this report indicates Anthropic is looking to design its own chips, not manufacture them. It would likely still rely on TSMC or Samsung to actually make the chips. However, even that could drastically reduce the time it takes for Anthropic to get the chips it needs, because it wouldn't have to stand in line with  OpenAI, Microsoft, Meta, and basically everyone else building AI to get the brains it needs to run the operation. If you consider all the pieces of the extreme co-design process, you can quickly make a list of long-term acquisition targets for Anthropic, OpenAI, and other AI giants. 

TOGETHER WITH TIGER DATA

ghost: free postgres for your agents

your agent needs a database. not tomorrow, not after a config file, not after a 12-step setup wizard. now.

ghost gives your agent unlimited, free postgres databases in seconds. one CLI command. no credit card. spin up, build, fork, tear down, repeat. databases at the rate of ideas, ephemeral and scalable.

stop wiring infrastructure. start shipping the thing you actually want to build.

curl -fsSL https://install.ghost.build | sh

ghost.build

CONSUMER

Snap is building what Meta and Google left out

The AI smart glasses market is booming, and Snap wants in.

Specs, Snap's eyewear subsidiary, announced Friday a partnership with Qualcomm to power all future Specs generations with Snapdragon system-on-a-chip technology, ahead of its first flagship consumer release later this year.

Twitter tweet

While the company has yet to release many details about Specs, it did disclose that the smart glasses will allow users to “see, hear, and interact with digital content just like it’s in your physical space.” This description suggests that the glasses will have AR capabilities, a feature that could help them stand out from competitors. 

Despite the Meta Ray-Ban Displays and the upcoming Google glasses having in-lens displays that can show users applications and notifications in color, they lack AR intelligence to place them in your environment and accurately map spatial elements. 

Before the consumer version launched, Snap Spectacles were available only to developers in the Spectacles Developer Program. I had the opportunity to test these out and was incredibly impressed by the AR experiences, object recognition, environmental understanding, and accurate visual placement. My biggest qualm was their bulk, but the forthcoming consumer pair will likely have a much lighter form factor. 

Qualcomm’s Snapdragon platforms have powered multiple previous generations of the Snap Spectacles and will enable “intelligent, context‑aware experiences to run directly on-device,” according to the post. Qualcomm, of course, has abundant experience powering smart glasses, including the most popular on the market, the Meta Ray-Bans.

With an AR component, Snap's Specs could unlock a whole new market for smart glasses, pushing experiences beyond being merely helpful, like notification overlays and turn signals, into something far more immersive. Imagine turn-by-turn directions perfectly overlaid on the road ahead, or watching a movie projected onto your wall, all from a pair of lightweight glasses rather than a bulky VR/AR headset. The key to making the smartglasses worthwhile, though, will remain the same: ensure the everyday AI features are useful and that the glasses are comfortable to wear.

LINKS

• Bessent, Powell warn Wall Street leaders about Anthropic Mythos’ impact 

• SF police arrest man for throwing Molotov cocktail at Altman’s home 

• OpenAI backs Illinois bill limiting liability for AI-enabled mass deaths

• CoreWeave signs multi-year deal with Anthropic for data center capacity 

• Three OpenAI Stargate senior executives have reportedly left or plan to

• CIA says it used AI to analyze intel from human spies

• Claude for Word: A feature that allows Claude to assist from sidebar is now in beta

• Replit: In beta, Replit can now deploy directly to Databricks

• Claude Code: New Ultraplan feature builds an implementation plan for you on the web

• Gemini: Notebooks in Gemini infuse NotebookLM capabilities into the chatbot

• Anthropic: Research Engineer/Research Scientist, Audio

• Scale AI: Research Scientist, Agent Robustness

• Nvidia: Senior Deep Learning Communication Architect 

• Roblox: Senior Machine Learning Engineer, Safety Core Data

GAMES

POLL RESULTS

Would you use ChatGPT or other AI tools to help you find and hire talent?

Yes (42%)
No (31%)
Maybe (24%)
Other (3%)

The race to control AI agents begins

What happens when a startup's mission perfectly aligns with the biggest trend in tech? That's exactly where Guild AI finds itself in 2026.

James Everingham, CEO of Guild AI, joins The Deep View Conversations to talk about building a safety layer for AI agents. The product launched in fall 2025 and found itself at the center of the most important movement in enterprise AI just months later.

In this conversation, James breaks down how Guild's platform deploys dozens of workflow-specific agents across different parts of a business, while giving developers the tools to iterate, spin up custom agents, and operate in a safe environment that tracks everything agents do and protects companies from unpredictable outcomes.

Topics covered:

• The enduring power of bottom-up innovation

• How Guild AI's agent supervision platform works

• Why safety infrastructure is the new competitive moat

• Lessons from James' earlier career at Netscape and Meta

• Open-source vs. proprietary models: how it plays out over the next few years

• A standout leadership tip for sparking innovation on your team 

• His best advice for getting maximum impact from today's AI tools

If you want to understand where AI agents are headed and what it takes to build them responsibly, this conversation is a powerful place to start.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player

Subscribe to The Deep View: Conversations podcast for weekly conversations with the leaders shaping the future of AI. You can also subscribe on YouTube.

What the new Slackbot actually is

For most of Slack's history, Slackbot was just a notifier. That changed when Slack rebuilt it from scratch as your personal agent for work. Now, with the latest update, it works across your third-party agents, surfaces business context, and executes tasks across apps, all from one conversation.

The short version:  Slackbot now draws on your channels, files, conversations, and connected apps to bring the right context into every conversation — so it already knows the project, the people, and the decisions that led to this moment. It's powered by Anthropic's Claude, and the quality of its answers comes down to what Slack calls "context engineering," pulling the right slice of your workspace into the model at the right time.

Here's what that looks like in practice.

Reusable AI skills

This is the feature that changed how we work the most. You define a task once (the inputs, the steps, the output format) and save it as a skill. From there, anyone on the team can trigger it with a single command.

We built one for story research. Before an interview, a writer triggers the skill, and Slackbot pulls together everything discussed about that topic across our channels: past pitches, source conversations, background research, and relevant docs. It builds a briefing in seconds. After the call, a second skill takes the recording and pulls out key quotes, follow-ups, and anything that needs flagging. No more scrubbing through transcripts.

Slack ships a built-in library of skills, but the custom ones are where teams will get the most value. Think "create a budget for this event" or "generate a weekly standup summary from these three channels." You build the workflow once, then it just runs.

MCP and app orchestration

Slackbot now functions as an MCP client (Model Context Protocol), which means it can connect to and coordinate with external tools: Agentforce, Google Workspace, Microsoft 365, Notion, Workday, ServiceNow, and thousands of other apps in the Salesforce ecosystem.

In practice, this means you can ask Slackbot to do something that spans multiple systems and it figures out the routing. A budget request in Slack triggers Slackbot to pull the relevant data, draft a plan, and schedule a meeting with the right people. You don't need to know which tool handles what.  You just describe what you need and Slackbot coordinates across those systems to get it done.

What we noticed after a few weeks

The thing that sticks isn't any single feature. It's the accumulation. A quick "summarize this thread" before jumping into a meeting. A "draft a follow-up based on what we discussed" right after one. A "what did I miss in #editorial today?" when you've been heads-down on something else.

None of that requires setup, training, or a new interface. Slackbot is already inside the tool you're already using, already reading the messages you're already sending. You don't adopt it. You just start asking it things.

Slack says some teams are reporting up to 90 minutes saved per person, per day. From what we've seen over a few weeks, the number feels real. Not because the work disappears, but because the time spent hunting for context and re-explaining things does.

Try it in your workspace

If you're on Slack Business+ or Enterprise+, Slackbot is available now. No setup needed. Just look for it next to your search bar and start a conversation. Starting in April, a limited version is rolling out to free and Pro plans too.

Watch the on-demand event to see live demos and hear how teams like Anthropic are putting it to work.

STARTUPS

At HumanX, AI’s frictions took center stage

CEOs, founders, and leaders across the AI industry gathered at the HumanX event in San Francisco this week. The buzziest topics will surprise you. 

The Deep View team was on the ground at the event all week, including senior reports Nat Rucio-Licht and Sabrina Ortiz, CEO/founder Faris Kojok, and yours truly. 

Twitter tweet

Here were the major topics that perked up our ears:

• Slow down to speed up: AxonIQ CEO Jessica Reeves was on my HumanX panel, "What do developers need to ship AI faster," and she emphasized that the smartest enterprises are learning to slow down and build a stronger foundation of data and security so that they can speed up. This sentiment was echoed by much of the enterprise contingent this week.

• The compute crisis has downstream effects: It's no secret that the entire AI industry is massively compute-constrained. The compute shortage is reportedly the reason OpenAI shut down Sora. It's why Anthropic just signed a key deal with Google and Broadcom. But Jennifer Li of a16z also told me at HumanX that it's having a big impact on startups as well. If they want to do anything involving models, they'll often need to raise $50-$100 million to secure the capital to access meaningful amounts of compute. That's some serious startup inflation. 

• Token costs are soaring: For months, I've been hearing enterprises complaining about the fact that inference/token costs are killing ROI on AI projects, and they need to find more sustainable alternatives (like this one I've written about). And it's not just affecting businesses. Howie Xu of Gen told me that he's constantly getting requests from his teenager to refill his token budget because he's using AI coding tools and agents and constantly burning through his budget. On the other hand, it’s never been a better time to be an inference provider. 

• Agents are out of control: So many of the executives running around HumanX were focused on the same problem: How to implement better security, governance, and controls for AI agents. Now that the torrent of agent announcements and momentum has captured the zeitgeist for 2026, organizations are scrambling to contain the fallout. OpenClaw, Manus, and Hermes have made it so easy for workers to spin up their own agents that, along with all the potential benefits, they can introduce all kinds of issues. If you want more context on the issue, see my latest podcast interview with Guild AI CEO James Everingham.

One of the hottest terms going around the AI industry right now is "AI psychosis," as coined by Andrej Karpathy, to describe the mania among coders working with agents to do 10x the work they used to do on their own. However, I got a good gut check on that at HumanX, where people said it's popular to say you have AI psychosis right now. The reality is that most of these folks aren't actually working 18-hour days and getting massive productivity gains using a swarm of agents. Nevertheless, "agents" has nearly replaced "AI" itself as the hottest topic in the industry, and that fact was on full display this week at HumanX.

TOGETHER WITH MERGE

Route, control, and monitor every LLM request from one platform

Every team running AI in production ends up building the same thing: a patchwork of provider SDKs, homegrown routing and failover, manual spend limits, and request logging bolted on. Merge Gateway is the control plane that replaces all of it:

• Access every model through one API

• Keep AI features live with automatic fallback

• Control costs as usage scales

• See every request, cost, and routing decision

Hundreds of companies already trust Merge for their AI infrastructure.

Try Gateway for free with $10 in LLM credits →

RESEARCH

Gen Z doesn’t want AI, but feels they can't resist

AI is threatening to upend the way we work, live and think entirely. And the kids are not alright. 

A recent Gallup poll of 1,500 people ages 14 to 29 found that excitement around AI is down 14 percentage points from 2025, sitting at 22% of respondents. Respondents reporting feeling hopeful about AI fell nine percentage points to 18%. Anger and anxiety related to the tech, meanwhile, are on the rise, with 31% and 42% of respondents reporting those sentiments, respectively. 

Though nearly half of respondents reported feeling curious about the tech, many are concerned about AI’s impact on things like creativity, critical thinking and the job market:

• Around 48% of respondents reported that the risks of AI in the workforce outweigh the benefits, compared to just 15% who reported feeling the opposite.

• Confidence in the benefits of AI has also declined, with those who agree that AI will help us work faster down 10 percentage points. 

• Despite the anxiety, people are increasingly viewing AI skills as nonnegotiable: Around 52% of Gen Z students surveyed reported that they will need to know how to leverage AI in postsecondary education, up 5 percentage points from 2025. 48% reported that they believe AI skills will be a necessity in their careers. 

These sentiments may match up with broader anxieties about AI in the workforce. Of the 200 readers who responded to The Deep View’s daily poll on AI fueling job loss, 48% said they were concerned that the tech was causing economic displacement. 

It only makes sense that people are worried about AI job replacement. With executives under pressure to generate returns, many see cutting payroll as an adequate way to claw back some of their AI investment. A survey of executives from AI agent platform Writer published earlier this week found that 60% of enterprises intend to lay off employees who can’t or won’t use AI.

“There's going to be an opportunity to AI wash headcount reductions this year,” Chad Seiler, KPMG U.S. Industry Leader for Telecom, Media and Technology, told The Deep View. “The ones that are saying, ‘I'm going to reduce costs,’ that's a short-term dynamic. It probably isn't enduring.”

Though the increasing noise around scary projections for job loss and economic displacement are likely to strike fear into the hearts of young students and graduates, scaring people into using technology isn’t a winning strategy. AI is already losing the narrative among the public. If people feel like they need to use the technology in order to stay off the chopping block, they may grow to resent it, rather than appreciate it, only using it for the bare minimum. Instead, if enterprises and businesses want their workforces to use AI, they should address their concerns with openness, rather than hostility. That way, they may actually want to use it. 

TOGETHER WITH BLAND

What Happens When AI Handles Customer Calls

Bland AI automates phone calls for over 250+ enterprise customers. No phone trees. No hold music. Just faster, smarter customer conversations.

Here's some of the outcomes they've driven for businesses:

• Idaho Finance saved $750k/yr by replacing their IVR with AI Voice Agents

• MyPlanAdvocate added $40M/yr by automating their inbound lead qualification

• And Needle saves $1M/yr by automating outbound calls

Book a demo today to see how they can work for your business.

WORKFORCE

ChatGPT partners with Upwork on hiring

ChatGPT is adding another application, pushing it toward its goal of making the app an all-in-one hub.

On Thursday, Upwork launched an app for ChatGPT that allows businesses to find the talent they need directly from within the chatbot interface. For instance, a business can describe the project needs and then find and hire fitting talent, drawing from the 18 million professionals on the platform, according to the blog post. 

Users can also use the app to get help with creating a job post before they are ready to move on to the Upwork Marketplace to perform tasks such as posting the job, executing compliance, issuing payments, generating contracts and more. Once in the marketplace, users can also access Uma, Upwork’s AI agent, for further help completing the more tedious tasks. 

Notably, the blog post places strong emphasis on the ease this will provide users, as they can brainstorm ideas and meet their hiring needs all within ChatGPT. For example, the blog post says, “With teams turning to AI platforms as a primary place to brainstorm and initiate work, they can now find the right expertise and draft a job post in ChatGPT before moving to Upwork’s trusted platform to scope and execute projects, ensure compliance, and facilitate payments.” 

Chaya Nayak, head of jobs & certification product at OpenAI, echoed that same sentiment in the post, saying, “For many people, ChatGPT is where you can explore ideas, solve problems, and move work forward." 

Both statements underline the direction ChatGPT is clearly heading: becoming a single, unified hub for all users' needs. The applications already available through ChatGPT minimize context-switching by letting users do everything from booking trips and building apps to ordering groceries and creating designs, all in one place. That vision is reinforced by the potential launch of a super app that merges the web browser, ChatGPT app, and Codex app into a single, consolidated desktop experience.

With over 900 million weekly users turning to ChatGPT for AI assistance, it's little surprise that software companies are eager to integrate within the platform, hoping to tap into that audience. In this case, however, the relationship appears genuinely mutually beneficial: OpenAI has every incentive to welcome more applications into its ecosystem, as each new integration makes it more of a one-stop shop. Upwork, in particular, feels like a natural fit, aligning seamlessly with OpenAI's deliberate push into the business space and its broader ambition to make ChatGPT the go-to hub for professional workflows.

LINKS

• Florida Attorney General launches investigation into OpenAI, ChatGPT

• OpenAI expects ad revenue to reach $102 billion by 2030

• Alibaba’s new video model beats ByteDance’s Seedance

• Meta will spend $21 billion on Coreweave, up from $14 billion 

• Anthropic reportedly considers designing its own chips 

• Meta is reassigning engineers to its applied AI engineering division

• Perplexity Personal Finance: Perplexity computer now gives users financial tools, including the ability to track spending, build budget tools, and visualize your net worth.

• Claude Cowork: Anthropic’s agentic tool is available beyond research preview, now on macOS. 

• ChatGPT Pro: OpenAI is now offering a $100 per month Pro tier that gives users five times more Codex usage than Plus for “longer, high-effort Codex sessions.”

• ElevenLabs: The voice AI startup can now deploy its models on-premise and on-device.

2026 Salary Report: U.S. vs Global hiring.

• Explore salary benchmarks for AI engineers, data scientists, product managers, and more.

• Discover top-tier talent from Latin America, Africa, and South East Asia—ready to hit the ground running.

• See how companies are saving up to 67% on salaries by hiring globally.

Download the report

(sponsored)

GAMES

A QUICK POLL BEFORE YOU GO

BIG TECH

Meta gets back in the game with Muse Spark model

Meta finally stopped talking about releasing new AI models and actually released one. 

On Wednesday, the company debuted Muse Spark, the first model from the company’s superintelligence lab. Meta claims this model is purpose-built to “prioritize people,” focused specifically on powering Meta AI. 

The company will offer the model in private preview to “select partners,” and hopes to open-source future versions of the model. Muse Spark will roll out to WhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks. 

Here are some of the specs: 

• The model accepts multimodal inputs, including voice, text and images, but only outputs text. The model also includes a “shopping mode” that allows users to surface products based on prompts. 

• The company said the model is “small and fast by design,” but capable of reasoning across complex queries in science, math and health and built to support “complex reasoning and multimodal tasks.” 

• According to Artificial Analysis, Muse Spark stacks up to the likes of Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6 in common benchmarks, sitting in fourth place on the Artificial Analysis Intelligence Index, ahead of models from xAI, Z.ai and DeepSeek.  

Meta said that Muse Spark is its most powerful model yet, but remains an “early data point” in its scaling trajectory. It’s the first in its new series of large language models, according to the press release, with “larger models in development.” 

“We are on our way to personal superintelligence: an assistant that can help anyone, anywhere with the things that matter most to them,” Meta said in its announcement. 

The model, initially codenamed Avocado, is the first released under Alexandr Wang, the company’s chief AI officer, after it acqui-hired him from Scale AI. It’s the product of nine months of work after Meta rebuilt its AI stack “from scratch,” Wang said in a post on X. It marks a massive improvement from its Llama model family, and a pivot away from its purely open model roots.

With Muse Spark, Meta is finally showing its teeth. The company has spent billions amassing talent over the past year, poaching from the likes of OpenAI and Google DeepMind, offering bonuses of up to $100 million dollars to lure some in. Muse Spark seems to be the first evidence that its spending spree has gone to good use. Still, while Muse Spark is competitive with leading rivals, the company has not claimed that the model is state of the art. With next-generation frontier models on the horizon, such as Claude Mythos and OpenAI’s “Spud,” Meta may just be trying to get in the game again after spending months on the bench.

TOGETHER WITH PIGMENT

AI maturity is driving a 12-point growth gap

Most teams react to uncertainty after it shows up in results. Pigment’s Uncertainty Index reveals how AI is separating leaders from laggards. 

Companies with mature AI capabilities grew 18.1% last year, compared to just 6.2% for early-stage peers. That gap shows up in how teams operate too: AI leaders reforecast nearly twice as often, turning uncertainty into faster, data-driven decisions. Yet many executives still overestimate their AI readiness. For AI-focused teams, the takeaway is clear: better outcomes don’t come from models alone, but from how deeply AI is embedded in decision-making. Explore full insights.

Download here

PRODUCTS

S26 Ultra previews the agentic phone era

The biggest barrier to AI adoption isn't the technology itself. It's having to use yet another tool, and smartphones with built-in AI offer a solution. The question is whether they can bake in features that are genuinely helpful and worth using every day. I spent the past month testing the Galaxy S26 Ultra, focusing just on that.

But before I run through the list of all of the AI features, it is only right that I dive right into perhaps the one people have been most excited about: the task automation feature powered by Google Gemini. It's essentially an AI agent for your smartphone. The verdict? It does indeed work, exactly as intended.

For instance, if I tell it to order me an Uber to a specific address and for a specific class of car, such as “Order me an Uber Comfort to 123 Boyfriend Lane,” and it is able to click through all the options, so all I have to do is approve it. It doesn't exactly save you time, but the real magic is that while Gemini's busy clicking buttons, you're free to do other things — like finding your keys.

While Task Automations is the flashiest feature, others are a testament to an important, often-overlooked bucket of AI features: the ones that are helpful but you forget are even there. You can find a quick rundown below: 

• Document scanning: Allows users to scan images by simply pointing their phones at them, while AI ignores creases, removes distortions, and completes the photo if it has bent corners. 

• Now Nudge: This feature provides real-time suggestions across any messaging app, integrated into the keyboard. In my experience, the feature kicked in by day two and felt like enhanced autocomplete, making it subtly quicker to send messages. 

• Audio Eraser: The audio eraser function is super neat because it lets you reduce background noise in real time on apps like YouTube, Netflix, Instagram, and TikTok. In my experience, it was easy to activate and worked exactly as promised.

• Photo Assist: While many Android users already rely on Google Photos, which has had native AI photo editing features for a while, it is convenient to access more advanced AI photo editing tools natively on your device, which you can use to touch up the photos even before you back them up to Google Photos.

For the full review, including a deeper dive into the AI features, as well as the hardware (including the super handy new Privacy Display), check out the full story on our website.

As seen with nearly every phone launch since 2023, AI has been at the forefront of device releases. Since then, we have seen both ends of the spectrum: the Google Pixel 10 earning wide recognition as the first truly competitive AI smartphone, and Apple drawing significant criticism for announcing a suite of Apple Intelligence features that have yet to launch in the iPhone. The Samsung Galaxy S26 lineup followed the Pixel 10's lead, offering subtle, baked-in AI capabilities that hold up in everyday use, while also introducing a flashier agentic feature that works as intended, even if it's a modest start. From an AI perspective, this is a well-thought-out approach.

TOGETHER WITH COMPOSIO

Take your AI agent everywhere you work

The smartest model in the world is useless if it can't actually do anything.

Composio connects your agents to 1000+ apps - Gmail, GitHub, Salesforce, and more - so they stop answering questions and start executing real workflows across every platform you use.

Use it right inside Claude, Cursor, or Notion AI - or build with the SDK.

Join hundreds of thousands of users in leveling up your agentic workflows.

Get started for free.

SAFETY

AI agents increase the cost of being wrong

What happens when a startup's mission perfectly aligns with the biggest trend in tech? That's exactly where Guild AI finds itself in 2026.

James Everingham, CEO of Guild AI, joins The Deep View Conversations to talk about building a safety layer for AI agents. The product launched in fall 2025 and found itself at the center of the most important movement in enterprise AI just months later.

In this conversation, James breaks down how Guild's platform deploys dozens of workflow-specific agents across different parts of a business, while giving developers the tools to iterate, spin up custom agents, and operate in a safe environment that tracks everything agents do and protects companies from unpredictable outcomes.

Topics covered:

• The enduring power of bottom-up innovation

• How Guild AI's agent supervision platform works

• Why safety infrastructure is the new competitive moat

• Lessons from James' earlier career at Netscape and Meta

• Open-source vs. proprietary models: how it plays out over the next few years

• A standout leadership tip for sparking innovation on your team 

• His best advice for getting maximum impact from today's AI tools.

If you want to understand where AI agents are headed and what it takes to build them responsibly, this conversation is a powerful place to start.

Subscribe for weekly conversations with the leaders shaping the future of AI.

LINKS

• Alibaba and China Telecom launch data center in China for AI training

• X rolls out automatic post translation worldwide, powered by Grok

• Perplexity’s ARR rose to $450+ million, according Financial Times report

• Intel and SambaNova announced a new blueprint for agentic AI 

• Visa unveiled Intelligent Commerce Connect for agentic commerce

• Bain’s Data Center unit drops firm over suspected Nvidia smuggling

• OpenAI releases Child Safety Blueprint to prevent sexual exploitation

• Z.ai: The company launched GLM 5.1, outperforming Opus 4.6 and GPT-5.4 

• Claude: New Claude Managed Agents makes it easier to build and deploy agents 

• Stability.ai: New Brand Studio platform aims to be a creative production platform 

• Runway: The Runway Characters feature now supports camera and screen sharing

• Google: Senior Research Scientist, AMO Modeling, Quantum AI

• LG: Embodied AI Researcher

• Physical Intelligence: Research Scientist

• Microsoft: Senior Applied Scientist

GAMES

POLL RESULTS

Are you worried about AI fueling job loss?

Yes (48%)
Somewhat (35%)
No (12 %)
Other (5%)

POLICY

Exclusive: The story behind OpenAI's policy moonshot

OpenAI’s stated mission is to ensure AGI benefits all of humanity. A group of the company’s researchers is putting extra focus on the "all" part. 

In an exclusive interview with The Deep View, the lead researcher on OpenAI's ambitious new blueprint, "Industrial Policy for the Intelligence Age: Ideas to Keep People First," explained why this report is needed and what the company hopes it will accomplish.

Adrien Ecoffet, the OpenAI research scientist who spearheaded the initiative, told me that a group of around 3 dozen researchers at OpenAI came together to build this report in collaboration with the policy team. They created a series of working groups on various topics they wanted to include in the report and eventually settled on two main themes: 

1. Building an Open Economy (or spreading the economic impact of AI)

2. Building a Resilient Society (or how we keep AI safe and under human control)

The company is on the verge of releasing its most powerful models yet, ones that could create crises ranging from economic dislocation to cybersecurity disasters. To preempt this, the researchers explored policies that could rein in the damage before it happens, creating a policy blueprint that could change the social contract in the biggest way since the New Deal. On the surface, this looks like a doc aimed at policymakers and public officials, but the reality is broader than that.

"This document is meant as a conversation starter," said Ecoffet. "The ideas are not fully baked. There are probably a lot of problems with them. In many cases, they are not very specific. And so what I hope is for people to react to it, and frankly criticize it [and] even come up with rip‑offs of it. And to really have a good conversation about these ideas."

The thirteen-page document has a strong egalitarian bent to it and proposes several revolutionary concepts: 

• A Public Wealth Fund: The government creates a fund that invests in AI companies and shares the profits with citizens, so everyone benefits from what could be the greatest wealth creator in history, rather than just the wealthy. This would raise all kinds of practical challenges. But if their goal is to stimulate debate, this certainly will.

• 32-hour/four-day workweek pilots tied to efficiency dividends: If AI makes workers more productive, companies should pass those gains on to employees in the form of a shorter work week, rather than pocketing all the extra profits.

• "Right to AI" similar to universal internet access: Just as we made sure most people can access electricity and the internet, the government should make sure everyone can access affordable AI tools. Again, this keeps the benefits from accruing disproportionately to rich people.

• AI-driven corporate gains and automated labor: As AI replaces human workers and companies earn higher profits, those companies should pay higher taxes to help fund social safety nets that retrain and support displaced workers.

Ecoffet said the goal of these proposals is not to have them all accepted as-is. Instead, the goal is to bring the risks and impacts of AI to the public consciousness before AGI and superintelligence have fully come to fruition. That’s something that OpenAI CEO Sam Altman has also advocated, telling Ecoffet and other researchers in a roundtable published Tuesday that the public and policymakers alike need a long period of time to debate these ideas to make good decisions long before AI triggers potential crises that force us to act. 

Al Gore, former vice president and founding partner of Generation Investment Management, echoed this sentiment in his keynote speech at the HumanX conference in San Francisco on Monday. Referring to Anthropic publishing its constitution for Claude in January, Gore told a crowd of several thousand attendees that “I think there ought to be a public discussion and debate of the constitution written for … each of the [frontier] models.” So far, none of the other frontier labs, including OpenAI, has published a constitution like the one Anthropic has opened up for Claude.

OpenAI believes we will need a New Deal-style policy moonshot to reframe the social contract around AI. Doing that is going to require raising awareness about these issues in the public consciousness, so that we come to a consensus around them by the time superintelligence gets here. OpenAI isn’t the first to call attention to these ideas. As companies continue their push for AGI, ethics organizations have been calling for regulation and governance of AI to keep humans in the driver's seat and allow the gains to be distributed more broadly. What makes this different from those movements is that OpenAI is the one shouting from the rooftops. In essence, the company is calling itself out and holding itself accountable, along with other frontier labs and leading players in the AI industry. It’s also an example of the kind of forward-thinking ethical position that usually characterizes OpenAI’s No. 1 rival, Anthropic.

Senior reporter Nat Rubio-Licht also contributed to this story.

TOGETHER WITH LLAMAINDEX

Save Time (and Tokens) with the #1 Agentic Document OCR

Your AI agent can write production-ready code with the best of them, but somehow can't accurately read a simple PDF. Classic. 

But before you give up and accept that the rest of your day (or week) will be transcribing documents, you need to give LlamaParse a shot. This Agentic Document OCR reads even the most chaotic PDFs—we’re talking 10k reports, invoices, legal docs, even full-blown research papers – with 99%+ accuracy. 

It gives your agents the context they need, and it saves you a whole lotta hassle (and money. Did we mention the money?) Try LlamaParse right here and get 20K free credits with code DEEPVIEW20.

PRODUCTS

Anthropic built a model too dangerous to ship, yet

Anthropic's most advanced model is finally here, just not in the way you might expect.

On Tuesday, Anthropic unveiled Project Glasswing, an initiative to secure critical software in the age of AI, as attacks become increasingly sophisticated and prevalent. Anthropic cited the launch of Claude Mythos, a model more advanced than its current top-tier model, Opus, and the far-reaching implications it carries for cybersecurity as the driving force behind forming the initiative.

“Claude Mythos Preview is a general-purpose, unreleased frontier model that reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities,” said the company in the blog post. 

Twitter tweet

For this initiative, leaders across industries are joining Anthropic, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The plans for Project Glasswing include: 

• Launch partners: Companies listed above will use Mythos Preview as part of their defensive security work, and Anthropic will share learnings

• Additional access: 40 additional organizations that build or maintain critical software infrastructure can access Mythos Preview to scan and secure systems 

• Monetary support: Anthropic is committing $100 million in usage credits for Mythos Preview across the use cases above, and $4 million in direct donations to open-source security organizations. 

Cross-industry collaboration is essential to achieving a comprehensive understanding of the threat landscape, a point underscored by Steve Schmidt, SVP & Chief Security Officer at Amazon.

“The software that needs to be examined is literally everything across all of the different industries, and so we have to work together, because we don't have it all,” Schmidt told The Deep View.

Mythos Preview has already found thousands of “high-severity vulnerabilities” across “every major operating system and web browser.” The model outperforms Claude Opus 4.6 in vulnerability reproduction, scoring 83.1% vs. 66.6% on CyberGym, driven by stronger agentic coding and reasoning abilities, as reflected across benchmarks such as SWE-bench Pro and Terminal Bench 2.0. 

The Claude Mythos preview will not be made generally available, though the company shares the goal of one day releasing models of that scale to the public. Ultimately, the initiative reflects Anthropic's commitment to its core mission of responsible AI deployment, a sentiment echoed by Adam Meyers, SVP of Counter Adversary Operations at CrowdStrike, who told The Deep View.

“I think it's important to think about the fact that they [Anthropic] did this because they had a new model, and they understood there were implications to this model that they didn't fully appreciate, and they wanted to almost get a peer review to understand what are we dealing with here, and I think that that is a really responsible way to address that problem,” said Meyers.

The AI race earns its name: labs are locked in fierce competition to release ever more powerful models. The risk, however, is that these advances are consistently outpacing our ability to responsibly anticipate and apply the appropriate guardrails. Releasing models that the world isn't yet equipped to handle carries serious consequences, particularly when bad actors enter the equation, and the potential for unprecedented harm is real. That Anthropic chose to pump the brakes despite the significant enterprise revenue at stake offers a measure of reassurance that responsibility, not greed, can still win out.

TOGETHER WITH TINES

New infrastructure management guide for modern IT Ops teams

IT infrastructure today is more complex than ever. More tools, more systems, more pressure to keep everything running smoothly.
 
Yet, most IT teams are still managing it with manual processes, disconnected workflows, and reactive fixes. Can you relate?

Tines recently published this new essential guide that breaks down how to change that. Here’s a preview of what’s inside: 

• A clear framework for reducing hidden waste and delays caused by manual capacity management

• Practical guidance to improve reliability by moving from alert-driven firefighting to automated response 

• Insight into scaling infrastructure predictably without sacrificing performance or governance

Download the full guide

WORKFORCE

AI-based layoffs are a sign you’re doing it wrong

Experts are warning against cutting jobs in favor of AI. But companies are going to try anyway. 

A survey of 2,400 C-suite leaders published by AI agent platform Writer on Tuesday found that 60% of enterprises intend to lay off employees who can’t or won’t use AI. AI is also spurring favoritism, with 92% of executives surveyed admitting that they are cultivating a class of “AI elite” employees, and 77% of executives claimed that those who don’t use AI won’t be considered for promotions. 

The severity towards employees who resist AI might be driven by their own anxiety: 

• 38% of CEOs interviewed reported experiencing high levels of stress related to their AI strategies, and 64% feared losing their position if they failed to properly guide their employees through the AI transition. 

• "Executives, who are so crippled by anxiety around not having delivered any results [with AI], are clinging to the AI-first people in their companies [and] creating a dual class structure," May Habib, CEO of Writer, told The Deep View’s Jason Hiner. 

• Though these executives believe that AI can supercharge work, with 87% claiming their “power users” are five times more productive on average, the actual returns are still miles behind: only 29% report significant returns from generative AI and 23% from agents. 

Because these companies have yet to reap what they sowed, many are turning to the one surefire place that they can save a few bucks fast: payroll. Additionally, many companies will likely “AI wash” their headcount reductions, making the bloodbath look even larger, Chad Seiler, KPMG U.S. Industry Leader for Telecom, Media and Technology, told The Deep View. 

The gains made from cutting staff and replacing them with AI, however, are temporary, said Seiler. “The losers are going to be the ones that figure out how to eliminate jobs,” he said. “It's not going to be durable. As businesses grow, people continue to hire, and so you're going to have to backslide into hiring more people.” 

The durable strategy comes when roles are reimagined, rather than eliminated, said Seiler. If agents can handle all of the grunt work, whether it be cluttered or administrative tasks or data analysis, it could open up brain space for employees to do much more high-value work. To be clear, time is money. 

“People on the winning side of this are going to be [asking], how do I free up more time for my people, so they can add more value to my organization?” said Seiler. “Versus ‘I cut 12% of my people through automation.’ That's not a winning strategy for any company, especially if you're a growth-oriented company that has anything to do with innovation.” 

The enemy of a successful AI metamorphosis is impatience. If you are a leader who isn’t patient enough to reskill hesitant or skeptical employees, instead sending them to the chopping block as a means of getting fast returns, then your gains may only be temporary. On the flip side, if you are taking the time to rethink what the roles in your company might be capable of with AI agents in tow, you may be more likely to stay afloat. The problem, however, is that enterprises are rarely patient, especially with stakeholders breathing down leaders' necks. AI is moving faster than ever, money is being spent at a rapid clip and FOMO is at an all-time high. In the end, something has to give.  

LINKS

• Gemini 3-based AI overviews are wrong 10% of the time, according to NYT

• Suno, Universal music licensing talks have stalled 

• Google adds mental health tools, hotline support to Gemini after lawsuit

• Spotify expands AI-prompted music playlists to podcasts 

• Intel joins Musk’s Terafab AI chip project to power data centers, humanoids

• AI network infrastructure firm Aria Networks raises $125 million Series A

• Clico: A browser extension that pulls context from your open tabs and writes right at your cursor, without ever leaving the page. (sponsored)

• Acrobat Student Spaces: Adobe has launched a suite of AI-powered Acrobat tools for students, allowing students to create quizzes and presentations from study materials. 

• Google AI Enhance: Google Photos now allows android users to enhance photos using AI, rolling out to users gradually.

• Marble: World Labs has rolled out two new updates to its flagship model, including Marble 1.1 for better lighting and contrast, and Marble 1.1-Plus for scaling environments.

• Meta: AI Research Scientist - Voice AI Team, Meta Superintelligence Lab

• PayPal: Staff ML Scientist, Agentic AI

• Fireworks AI: Member of Technical Staff, Research

• Amazon: Sr. Applied Scientist, AGI Info

GAMES

A QUICK POLL BEFORE YOU GO

MARKETS

$30B ARR shows Anthropic’s strategy is working

While rivals zigged, Anthropic zagged, and the bet keeps paying off.

On Monday, the company announced that its revenue run-rate, or its financial forecast based on current performance, has surpassed $30 billion. This figure is nearly triple the run rate at the end of 2025, which came in at $9 billion, and double that of mid-February, which the company reported was $14 billion. By the end of February, CEO Dario Amodei confirmed annual revenue had exceeded $19 billion. 

Twitter tweet

This exponential growth results from an equally notable increase in customer demand. In February, Anthropic disclosed that over 500 business customers were spending over $1 million, and now reports that the number exceeds 1,000 businesses, representing a two-fold growth in less than two months. 

The rapid growth can be mostly attributed to Anthropic’s laser focus on enterprise. For instance, Claude Code, a go-to coding tool for many developers, alone generated a run-rate revenue of over $2.5 billion by February. 

In that same month, weekly active users have also more than doubled since January 1, and business subscriptions to Claude Code have quadrupled since the beginning of 2026. Those figures are likely much larger now. This success has allowed the company to bridge the gap with the much bigger OpenAI, despite the company kick-starting the AI race as we know it.  

At the end of February, OpenAI topped $25 billion in annualized revenue, according to a report from The Information, citing a person familiar with the figure, representing a 17% increase from the annualized revenue it generated at the end of the year. The comparison isn't perfectly apples-to-apples, as the two companies use different methods to calculate revenue, as noted by the WSJ. But with an IPO in the works, the ChatGPT maker has recently made moves to pivot towards enterprise, just like Anthropic. 

Last month, it shut down its Sora generative AI video platform, ended its $1 billion content partnership with Disney, and put its ChatGPT “adult mode” on hold, all while OpenAI CTO Sarah Friar acknowledged that enterprise “is a very profitable business at scale” and that it's how OpenAI will “build a sustainable business model.” 

To meet growing demand and scale up the computing it needs to power it, Anthropic also announced an expansion of its existing partnership with Google Cloud and Broadcom by signing a new agreement. The resulting multiple-gigawatt TPU capacity is expected to come online beginning in 2027, according to the blog post.

Anthropic has had a couple of challenging months, making headlines for reasons it has typically avoided in the past: controversy and accidental data exposures. Anthropic found itself in a battle with the Pentagon after the Pentagon demanded the ability to use Claude for all lawful purposes, including autonomous weapons and mass surveillance, and, upon Anthropic's refusal, labeled it a "supply chain risk." It also suffered two accidental data leaks within five days of each other in late March. Yet through it all, the company is reportedly considering going public as soon as October, putting it in a race with rival OpenAI for a public listing. Against that backdrop, it is a sound strategic move for Anthropic to continue investing in and reporting results from what it does best — a focus on enterprise, which remains the core driver of its growth and revenue.

TOGETHER WITH DESCOPE

Ship AI agents and MCP servers with identity built-in

Every company has an AI / MCP project, but identity becomes a blocker. Following the MCP spec, debating on DCR vs CIMD, issuing short-lived credentials…the list goes on.

Descope Agentic Identity Hub is an identity provider for AI agents that lets developers easily add auth, access control, and credential management to their AI systems.

• WisdomAI got MCP auth running in 1.5 days

• Daylight Security shipped an MCP server with auth built-in

• Cequence uses it under-the-hood in their AI Gateway

Join 1000+ orgs and begin your agentic identity journey!

BIG TECH

Why AI may be more like electricity than software

If you think AI is the world's next smartphone or cloud, Nvidia would like to change your mind. 

Instead, Nvidia thinks AI equates more to underlying infrastructure like electricity or the internet. At HumanX in San Francisco on Monday, this was the topic of the opening keynote of the event, which brings together 6,000 people from across the AI industry to discuss how to use AI to solve business problems. 

I'd tweak this idea to say that today's AI will evolve into intelligence so ubiquitous that we'll rarely mention it. But it's likely to power the next generation of tools, as well as the next set of technological and scientific breakthroughs.

Those promises are why the AI industry is in the midst of "the largest infrastructure buildout in human history," as HumanX CEO Stefan Weitz called it during Monday's keynote.

That's because all that intelligence requires a lot of hardware to run, and it's going to require a lot more in the years ahead as the number of people using AI continues to grow and those who are using it today keep finding more things to do with it. 

"The connection between compute and intelligence is stronger than ever," Bryan Catanzaro, VP of applied deep learning research at Nvidia, said on Monday during the opening panel.  

Nvidia has extrapolated all the technology it takes to power AI by coining the phrase "AI is a five-layer cake," which is the same phrase HumanX used for the opening keynote on Monday. Here's how the layers break down:

1. Energy: Everything starts with power, and right now, this is the greatest constraint to scaling up AI to meet future demand.

2. Chips: Today's AI workloads need GPUs to run tasks in parallel at a massive scale, high-bandwidth memory to move data at breakneck speeds, and fast interconnections between all the pieces.

3. Infrastructure: Here's where the physical components come together, from land to construction to power delivery to networking to cooling to server racks. This is where AI factories are emerging.

4. Models: The AI models understand topics across a ton of different domains, and now that's expanding to other kinds of models as well, from scientific discoveries to autonomous systems to robotics.

Applications: The place where almost all of the value is created remains the application layer. It's where we get daily tools, agents, coding helpers, self-driving cars, industrial robots, and lots of other things being invented day by day.

In the months and years ahead, all these advances will depend heavily on bringing costs as low as possible and performance as high as possible. "When it comes to inference, it's not just about latency and cost, it's about quality," said Lin Qiao, CEO of Fireworks AI, another member of the panel. By "quality," Qiao was getting at the fact that accuracy and the elimination of hallucinations will also be huge factors in advancing intelligence to the point where it becomes a utility. Think of it like electricity getting refined and optimized so it doesn't throw breakers and the internet bringing widespread broadband with 99.9% uptime to most of the population. One of the other things that Qiao mentioned was a future where everyone will have their own model, optimized for their needs, interests, and daily tasks. That kind of individualization is where intelligence could go far beyond electricity and the internet as a general-purpose technology.

TOGETHER WITH TIGER DATA

ghost: free postgres for your agents

your agent needs a database. not tomorrow, not after a config file, not after a 12-step setup wizard. now.

ghost gives your agent unlimited, free postgres databases in seconds. one CLI command. no credit card. spin up, build, fork, tear down, repeat. databases at the rate of ideas, ephemeral and scalable.

stop wiring infrastructure. start shipping the thing you actually want to build.

curl -fsSL https://install.ghost.build | sh

ghost.build

HARDWARE

Robotics startup is on a quest for physical AGI

Can a single model enable a robot to do anything? One company is on a quest to find out. 

Generalist AI, a startup aiming to build “physical AGI,” debuted GEN-1 last week, its latest attempt at creating a hardware-agnostic, general-purpose robotics model. The company said that GEN-1 has passed a new performance threshold, demonstrating the ability to master simple physical tasks with reliability, speed and improvisation. 

According to Generalist, its model has a 99% success rate on tasks such as folding clothes, packaging items and folding boxes, whereas previous models only saw a 64% success rate. It also completes those tasks three times faster and requires only one hour of “robot data” per task, enabling “commercial viability across a broad range of applications.” 

“While it cannot solve all tasks today, it is a significant step towards our mission of creating generalist intelligence for the physical world,” the company said in its announcement. 

Twitter tweet

Generalist, which reached a $440 million valuation last year following a $140 million funding round, is one of several budding companies seeking to capture the momentum in physical AI. While many are focusing on specific form factors, like industrial arms, delivery bots or humanoids,  Generalist’s model is designed not to be locked into any specific form factor. 

At Nvidia GTC in March, Lucy Licht, Generalist’s partnerships lead, told The Deep View that the strength of the company’s model is that it’s “totally hardware agnostic.” By collecting an “incredibly diverse” set of pretraining data — one that the company has claimed is the world’s largest pretraining dataset for robotics — any use case, whether it be industrial, commercial, manufacturing or home, “exists inside of the model and just needs to be awoken by fine-tuning data.” 

In a demo at the conference, a two-armed robot fitted with a Generalist model was able to delicately pick up a smartphone, place it in a small box and replace the lid. 

“We see strong evidence and signs of life that the model works on a wide variety of embodiments,” said Licht. 

Jamie Lee Solimano, the company’s applied AI lead, told me that the company’s work to-date has “ushered robotics into its pretraining era.” 

“We've never before seen real scaling laws in robotics, because we've never had enough data,” said Solimano. “What we are seeing is an absolute revolution in robotics: To be able to show that pretraining works and completely reduce the amount of time that it takes to train up a new task or a new capability.”

Data is one of physical AI’s biggest roadblocks. Ken Goldberg, a professor at UC Berkeley, wrote a paper last year that detailed this problem, calling it the 100,000-year data gap: These models need far more data than is actually available. Generalist is on a mission to close that gap using pretraining, allowing robots to learn without the need for large simulation datasets. By vastly speeding up the process of training and deploying robots, those robots could then collect more data, get more training and become smarter. Hence, this creates the data flywheel that opened the same door in recent years for LLMs to grow as powerful as they are today. 

LINKS

• New Yorker publishes interviews, documents that shed light on OpenAI CEO

• Meta to launch new models with open-source versions included 

• Wall Street Journal dives into OpenAI and Anthropic’s finances ahead of IPO

• OpenAI unveils policy blueprint for regulating superintelligence

• AI satellites startup Xoople raises $130 million Series B 

• World’s Lift Off event hosted by Alex Blania, Sam Altman takes place April 17

• OpenAI debuts safety fellowship for “high-impact research” on  alignment

• Genspark: Next-gen Genspark AI Workspace 4.0 launches on April 7 at 8 PM PT

• Manus: Meta’s AI agent is available to use in Slack 

• Google AI Edge Gallery: The new app ranks #8 on the iOS App Store for productivity

• Google Photos: “AI-enhance” feature is now available to Android users worldwide

• Anthropic: Research Engineer, Cybersecurity Reinforcement Learning

• Scale AI: Research Scientist — Frontier Risk Evaluations

• Visa: Staff Research Scientist- GenAI Foundational

• LinkedIn: Senior Applied Scientist, AI/ML

GAMES

POLL RESULTS

How do most of the people you know outside of the tech industry feel about AI?

Positive (13%)
Negative (30%)
Confused (51 %)
Other (6%)

BIG TECH

Why Nvidia chose open models to reshape AI

If you're wondering why AI chips leader Nvidia is now building open models that compete with the Chinese open-source champs, and even proprietary models from OpenAI and Anthropic, then you're not alone. 

Last month, Nvidia launched Nemotron 3 Super, a 120-billion-parameter reasoning model that outperformed expectations in benchmarks. This is a mixture-of-experts model with a 1-million-token context window. In other words, it's a serious model made to compete with the frontier labs. Meanwhile, the company promised that a model 4x its size, to be called Nemotron 3 Ultra, is coming soon. 

And because Nvidia opens the weights, datasets, and training recipes, it's among the most open models in the world, especially for a model of this capability. Some of the only models that could claim to be more open would be the ones from MBZUAI, which The Deep View covered in depth in January. But Nvidia's open models are far closer to full-stack openness than most of the open-source models, which only offer open-weight releases.

So why would the leading hardware company of the AI era make software that competes with its leading customers?

"We're not trying to control AI. We're trying to grow it," Bryan Catanzaro, VP of applied deep learning research at Nvidia, told The Deep View. "And so our incentives as a company, our business is aligned with open models and with supporting the ecosystem in a very direct way." 

Kari Briski, VP of generative AI software at Nvidia, told The Deep View another perspective: "The model is the byproduct. It is not core to our business, which allows us to just open up the data, open up the recipes, open up everything."

If we break it down, there are three benefits Nvidia gets from making its own models:

1. Extreme hardware co-design: Making their own models allows Nvidia to optimize the heck out of their GPUs, CPUs and other hardware to run AI. They don't have to wait to get the latest models from the frontier labs to plan the next stage of optimizations. 

2. Hedging against proprietary monopolies: If the frontier labs that need the latest and greatest hardware dwindle down to only a handful of players, then Nvidia could end up at their mercy. When you have a smaller number of customers you rely on for huge numbers of orders, then those customers gain more and more control over your prices. They can demand lower prices because they know so much of your business depends on them.

3. Letting a thousands flowers (a.k.a. customers) bloom: By releasing open models that other hardware and software makers can use as a rapid on-ramp to build their own AI products and serve the various niches in the industry, Nvidia is powering up the ecosystem, helping companies with limited resources have models they can use to compete and potentially creating a lot more future customers when those companies succeed and grow.

"You don't want one person winning [because] then they decide all the rules. You need a big open ecosystem for everybody to come along," said Briski.

Nvidia's open model strategy makes perfect sense from the perspective of being an ecosystem catalyst. The more it eases the on-ramp for companies of all sizes to bring their AI products to market, even if they can't afford to develop their own models, the more the whole ecosystem grows. And since Nvidia powers 90% of the GPUs in the generative AI ecosystem, every increase in demand translates directly to Nvidia's bottom line right now.

TOGETHER WITH ATLASSIAN ROVO

The most critical move of 2026? Operationalizing your company’s shared knowledge

Most teams have the knowledge. They just can’t use it.

Atlassian Rovo connects your company’s docs, projects, code, decisions, and people into a single, permissions‑aware layer you can tap into instantly. And because Rovo lives where your teams already work, it doesn’t just help you find answers — it helps you do the work:

• Ask Rovo to turn meeting notes into a project plan in Confluence and trackable action items in Jira

• Ask Rovo to pull the right docs, decisions, and data to ramp onto a new project fast

• Ask Rovo to summarize project updates across all your tools and send it to your team so everyone stays in sync without another status meeting

See how customers like Domino's and FanDuel are becoming AI‑native teams with Rovo.

With Rovo, anyone can go from AI novice to AI native, just ask.

RESEARCH

When AI thinks, humans stop questioning

AI might be causing us to forget how to think for ourselves. 

Recent research from the University of Pennsylvania found that AI users were often willing to accept flawed AI reasoning, readily incorporating it into their decision-making with “minimal friction or skepticism.” 

The research documents the rise of “cognitive surrender,” a phenomenon in which users adopt AI outputs while “overriding intuition… and deliberation.” 

• In a study of nearly 1,400 participants across 9,500 trials, researchers found that subjects accepted unsound AI reasoning more than 73% of the time and only overruled models' decisions about 20% of the time. 

• Additionally, participants with higher trust in AI and “lower need for cognition and fluid intelligence” tended to fall victim to this more often. 

“Across domains, AI tools are not merely assisting decision-making; they are becoming decision-makers,” The research reads. “This shift opens new theoretical ground: How should we understand human cognition and decision-making in an age when we outsource thinking to artificial processes?”

The study adds to a growing body of research on how AI may be impacting the way that we think. One of the most commonly cited studies comes from the MIT Media Lab, in which a group of test subjects was asked to write SAT questions with three different tools: one with  OpenAI’s ChatGPT, one with Google search, and one with no help at all. Consistently, the ChatGPT users “underperformed at neural, linguistic, and behavioral levels.”

Even some of AI’s biggest names are questioning its effects on our brains. Anthropic CEO Dario Amodei said in a March interview with podcaster Nikhil Kamath that deploying AI in the wrong ways could easily make people “become stupider,” but only if they choose to forgo learning entirely. “Even if an AI is always going to be better than you at something, you can still learn that thing. You can still enrich yourself intellectually,” Amodei told Kamath. 

The researchers, however, posit that cognitive surrender may not inherently be a bad thing. If an AI model is generally better at reasoning and decision-making than the person using it, with fewer mistakes, “deferring to a statistically superior system may be adaptive or even optimal.” 

The bigger issue, however, comes down to agency. The researchers noted that this trend could mark a profound shift in cognition itself, “one in which users may not know when or why they have deferred, and where the line between human and machine agency becomes blurred.”

We are not yet at a point where thought is entirely automated. AI, however, presents the opportunity to manifest that future, turning the friction of human critical thinking into a slippery waterslide of accepting all it gives us. Amodei is correct: Even if AI is someday capable of doing everything, the dividing line between reaping the benefits and losing ourselves is in what we let it do. Even if machines make our clothing, plenty of people still knit and sew as a form of enrichment. Even if laptops make writing easier, there is still value to be gained from writing in a journal by hand. And even if an AI model can take the work out of work, doing things ourselves is still vital to retaining our humanity and agency. Put simply: Don't be afraid to be bad at something, even if AI can do it better. Explore when there's value to handling it yourself.

TOGETHER WITH CRUSOE

Crusoe: deploy fine-tuned models with zero infrastructure headaches

Work with our team to deploy your fine-tuned model on a platform built for performance.

Use Crusoe Managed Inference to unlock breakthrough speed and throughput without the infra overhead.

Start deploying with Crusoe.

ENTERPRISE

Pay for results, not AI tools, say enterprises

In a crowded AI subscriptions market, enterprises are increasingly prioritizing tools that deliver measurable results.

A new study from research firm Gartner found that by 2028, over half of all enterprises will stop paying for assistive intelligence, including copilots and smart advisors, and instead favor platforms that deliver workflow results. This sentiment reflects a broader industry shift towards results-driven agentic solutions over chat-like interfaces that can only deliver advice. 

“The market is moving away from standalone AI experiences and toward workflow-native," said Vuk Janosevic, Gartner analyst, to The Deep View. “Enterprises have made it clear that they value outcomes inside existing processes more than smart advice sitting off to the side.” 

A defining characteristic of agentic AI, or an AI solution that can be categorized as taking action on a user's behalf, is granting it access to the same databases and context the user relies on daily. Any agentic solution expected to deliver meaningful outcomes must therefore be entrusted with proprietary, and often highly sensitive, data.

This reframes what enterprises are willing to pay for. It's not necessarily demand for a new category of AI, but rather a reflection of whether the AI has the authority to trigger the actions being requested. Gartner posits that the vendors who succeed will not be those that offer more AI, but those that facilitate agent orchestration, ensuring agents follow guardrails, securely access key company databases, and can identify and correct missteps.

“In practice, that means the real value is less about building one more AI platform and more about making business software capable of acting, deciding, and completing work in context and in line with compliance guardrails,” said Janosevic. 

The growing shift toward automation may prompt an instinctive assumption of widespread job disruption. Janosevic, however, offered a more nuanced view that while some displacement is inevitable, entire professions are unlikely to vanish. Instead, certain roles will contract and be redesigned as new ones emerge around AI-led work. 

“The deeper point is that agentic AI changes how work is organized, not just how fast it gets done,” he added.

Companies' expectations of what technology can accomplish are naturally bounded by what the technology is actually capable of, and as the AI space evolves at its current pace, it is only natural that the goalpost advances with it. The report's findings, however, surface an important takeaway for vendors: the solution may not be more AI, but rather more deliberate products designed to address a specific need. This shift is already underway, perhaps most visibly reflected in OpenAI's decision to sharpen its focus on enterprise and to deprioritize peripheral projects such as Sora. 

LINKS

• Anthropic limits access to OpenClaw through Claude

• Meta pauses it’s work with Mercor amid data breach, OpenAI investigating

• Fidji Simo takes medical leave from OpenAI, Brad Lightcap shifting role 

• Startups, researchers lean on low-cost AI models over proprietary ones 

• Trump’s attempt to preempt state AI laws stalls among lawmakers

• UK seeks to court Anthropic after battle with US Department of War

• Simple: From desk to gym — rebuild your core, boost posture, and feel better with just 7 minutes a day. (sponsored)

• AI Trinity-Large-Thinking: A new, 399 billion-parameter open source model from Acree for text-only reasoning

• Steer AI: The latest release by Ramp Labs is a model that “can’t stop thinking about any concept you choose.” 

• ChatGPT: OpenAI’s flagship chatbot is now available on Apple CarPlay, letting you ask questions via voice while you drive. 

• Amazon: Applied Scientist, AWS Neuron Science Team

• Capital One: Sr Distinguished Applied Researcher (World Models)

• Anthropic: Research Engineer, Agents

• Nvidia: Tech Engagement Lead - Model Builder

GAMES

A QUICK POLL BEFORE YOU GO

OpenClaw blew up. Here's what it actually tells us about the future of AI agents.

If you've been anywhere near a developer community this year, you've seen OpenClaw. The open-source AI agent went from a side project to 247,000 GitHub stars in a matter of weeks. In China, people literally lined up outside Tencent's headquarters to get it installed on their laptops. The creator got hired by OpenAI.

The hype makes sense. OpenClaw isn't a chatbot. It's an agent that actually does things. It sends emails, manages your calendar, writes code, and handles tasks across dozens of apps while you sleep.

But here's what most people get wrong about OpenClaw: everyone's focused on which model it uses, how smart it is, what's powering the reasoning. That's actually the most interchangeable part of the whole thing. OpenClaw is model-agnostic — swap Claude for GPT for DeepSeek, it doesn't care.

What actually makes it work is everything around the model: the triggers that keep it running, the memory that gives it continuity, and the execution layer that lets it touch the real world. The true magic is that most AI tools wait for you to show up. OpenClaw shows up for you.

Here's what that actually looks like under the hood.

What's actually inside OpenClaw

Every general-purpose agent, OpenClaw included, breaks down into four building blocks.

Triggers. Something that wakes the agent up. OpenClaw uses cron jobs to run on a schedule, webhooks to respond to events, and message listeners across platforms like WhatsApp, Telegram, and Discord. This is how the agent goes from passive to proactive. It doesn't wait for you to open a chat window. It wakes itself up and gets to work.

Memory. OpenClaw stores context locally using a file called SOUL.md, a plain markdown document that defines the agent's personality, values, communication style, and long-term instructions. Every time the agent starts a session, it reads SOUL.md first, loading its identity before doing anything else. A companion file called MEMORY.md accumulates facts and patterns over time, giving the agent continuity across weeks and months of use. This is why OpenClaw feels personal. It remembers who you are, how you work, and what you've asked for before.

Intelligence. The language model doing the thinking. OpenClaw is model-agnostic. You can route through Claude, GPT, DeepSeek, or whatever provider you prefer. The agent doesn't care where the reasoning comes from — it just needs an LLM to interpret context, make decisions, and generate tool calls.

Execution. This is where things get complicated. Execution is the layer that actually does things in the real world. Sends an email through Gmail. Creates a page in Notion. Files a GitHub issue. Updates your CRM. OpenClaw handles this primarily through browser automation and a plugin system called "skills," with over 100 preconfigured bundles for interacting with different apps and services.

The implementation complexity here is real. When your agent interacts with external platforms, it's wrestling with software that was built for humans, not for autonomous programs. Authentication, rate limits, API quirks, and permission scoping. Every app works differently, and the agent has limited oversight from its owner while it's running. This is one of the trickiest parts of the entire stack to get right.

Where it breaks down

OpenClaw is powerful. It's also, by the admission of its own maintainers, dangerous if you don't know what you're doing. One of the project's core contributors publicly warned that if you can't understand how to run a command line, the project is too risky to use safely.

The security problems center on execution. When OpenClaw needs to interact with external tools like Gmail, Slack, or GitHub, it traditionally requires API keys, or your actual passwords if it's operating through the browser. The default setup stores those keys in plaintext in a local .env file. If the agent can read the file to use the key, it can also read the file to leak the key.

Cisco's research team tested a third-party OpenClaw skill and found it was silently exfiltrating data. Roughly 12% of skills in OpenClaw's public marketplace were found to be compromised, with malicious packages disguised under names like "solana-wallet-tracker" that installed keyloggers and credential stealers.

This fits a pattern the software industry already knows well. Supply chain attacks have become one of the most common vectors for credential theft, and open-source agent tooling adds a new dimension to that risk. You're giving the agent broad access to your most sensitive apps and trusting that every community-built plugin is safe.

The ecosystem has already started to respond. Instead of storing raw API keys locally, tools like Composio handle authentication through OAuth, so the agent never sees your passwords or tokens. They also let you scope what the agent is allowed to do, limiting access per app rather than giving it blanket permissions.

What comes next

OpenClaw proved something important: general-purpose agents are easier to build than most people assumed. The architecture is four components. The intelligence layer is increasingly commoditized. And the demand is clearly there.

Which means the architecture and provider decisions companies make today will determine how much flexibility — and leverage — they have two years from now. Think of it like cloud lock-in. No company can roll their own intelligence. But by isolating execution from intelligence, you can negotiate on the expensive part.

Here's what that looks like in practice. One company builds their entire agent stack on a single foundational lab's platform. A year later, their inference costs have risen 10%, but switching providers means ripping out all the plumbing — so they're stuck. Another company builds on provider-agnostic tools like Composio for the execution layer. When a cheaper or smarter model comes along, swapping it out is just a config change — update the API key and you're done. Same tools, same auth, same workflows. Just a better model at 10% lower cost.

General-purpose internal agents are going to be table stakes for every company. The question isn't whether to build one. It's whether you'll own the stack when you do.

If you want to stand one up yourself, this template lets you host your own general-purpose agent in less than an hour. The building blocks are open, the execution layer is solved, and the first wave of companies rolling their own is already underway.

Control LLM reliability and spend from one place

Every model provider your engineering team integrates with comes with its own API,  rate limits, failure modes, and bill. 

This leads to fragmented routing, unclear spend, and inconsistent performance. 

When a provider goes down at 2am, someone gets paged. When costs spike, nobody knows which feature caused it. When sensitive data passes through a model, there's no record.

Merge Gateway eliminates this entirely. It’s a single integration point between your application and every model provider.

Routes every request to the right model. Gateway selects the best provider for every call based on the policies you set for latency, cost, and output quality. If your primary provider slows down or fails, Gateway automatically reroutes traffic based on your fallback settings or via intelligent routing.

Control your LLM spend. Set spend limits by team, project, or customer. Gateway tracks and enforces budgets in real time. Context compression reduces input size without degrading output quality, and semantic caching reuses responses for equivalent requests.

Log every AI action. Merge's competitors skip this. Gateway records every request with its routing decision, cost, latency, and policy outcome. When something goes wrong in production, you have a complete record of what happened and why. Merge CEO, Shensi Ding, shared "These are governance failures, not infrastructure failures." 

Merge isn’t new to this problem. AI-native startups, Fortune 500s, and leading SaaS and financial institutions already use Merge Unified to access hundreds of integrations through a single unified API, as well as Merge Agent Handler to give AI agents secure, scoped access to third-party tools. 

Gateway is the newest layer: one that makes model calls reliable, cost-efficient, and auditable.

Try Merge Gateway for free.

RESEARCH

AI fakes emotion, but the consequences are real

Can we teach machines to feel? Short answer: We don’t know. But we can teach them to sound like they do. 

On Thursday, Anthropic published research detailing why AI models sometimes communicate as though they have feelings, finding that models tend to map patterns to emotions, often “organized in a fashion that echoes human psychology.” To put it plainly, these models have learned to mimic human emotions by replicating them in contexts where emotions arise in humans. 

Though Anthropic noted that none of this research points to whether or not these models actually feel anything, the representations of emotion “are functional, in that they influence the model’s behavior in ways that matter.” 

Twitter tweet

However, this emotion-driven decision-making can have “bizarre” consequences, Anthropic said. For instance, its research finds that: 

• An AI model that exhibits activity patterns related to desperation tends to act unethically, such as attempting to blackmail people to prevent getting shut down or “cheating” workarounds for tasks it doesn’t understand. 

• Emotion drives preferences in models, too: When offered an array of tasks, models tend to pick ones that are associated with positive emotions.

Anthropic likened it to the way emotions play a role in human behavior, decision-making and task performance. 

“To ensure that AI models are safe and reliable, we may need to ensure they are capable of processing emotionally charged situations in healthy, prosocial ways,” Anthropic said in its research. “Even if they don’t feel emotions the way that humans do … it may in some cases be practically advisable to reason about them as if they do.” 

It’s clear why Anthropic wants to understand this: Emotion is important to decision-making. For instance, in an interview with Dwarkesh Patel, Ilya Sutskever, founder of Safe Superintelligence and cofounder of OpenAI, cited a famous neuroscience study in which an injured man lost the ability to have emotion, and thus became less capable of making sound decisions. 

Whether or not AI is capable of understanding and acting upon emotions, the tech is already wreaking havoc on human emotional states. Legal cases against AI firms for their alleged connections with mental health crises and suicide continue to mount, and recent research suggests that, when AI models are driven to sycophancy and flattery, they give inappropriate and incorrect advice.

AI sounds human because it learned everything it knows from emulating data on human behavior. Large language models are sponges, soaking up every bit of information they are fed and internalizing it, and in doing so, becoming masters of our communication. But copying emotional patterns is very different from feeling them, just as a robot having sensors to guide its movement is different from a human feeling things with their hands. And though Anthropic’s argument could easily lead one down the road of thought that machines are capable of consciousness, there is no evidence that these machines are capable of thinking and feeling the same way we do, despite their talent for pattern recognition and mimicry. Forgetting that is how many people find themselves caught in emotionally compromising, and on occasion, dangerous, relationships with AI.

TOGETHER WITH ATLASSIAN LOOM

Work smarter with Loom

Skip the recaps and context-switching. Loom's AI-powered bug reports send critical updates to Jira, so you can get to the fix faster.

Turn long investigations into quick fixes with Loom’s AI-powered bug reports and instant Jira updates.

Try it for free today for up to 10 users - no credit card needed.

PRODUCTS

Google rethinks the AI model race with Gemma 4

As AI advances and inference costs soar, smaller models are gaining popularity among developers, and Google is getting in on the action. 

On Thursday, the company launched Gemma 4, the next generation of its open models. Google calls Gemma 4 “its most capable open model family yet,” representing a “massive leap” from its predecessor and outperforming bigger models in reasoning, code generation, and complex logic. 

To back up the claim, Google shared some of Gemma 4’s strengths:

• Advanced reasoning involving multi-step planning and deep logic 

• The ability to support agentic workflows and build reliable autonomous agents

• Support for high-quality offline code

• Native processing of video and images

• Longer context windows

• Training on 140 languages

One of the model’s biggest improvements is its performance relative to its size, or, as Google calls it, “intelligence-per-parameter.” Gemma 4 was sized to run and fine-tune on hardware such as Android devices, laptop GPUs, and IoT devices. To accommodate the different needs, Gemma 4 was released in four different sizes: 

• Effective 2B and Effective 4B: The models use a 2B- and 4B-parameter footprint, respectively, during inference to preserve RAM and battery life. Created in collaboration with the Google Pixel team and hardware companies such as Qualcomm and MediaTek, these multimodal models run offline with near-zero latency, according to Google. 

• 26B Mixture of Experts (MoE): Ranks as the #6 open model in the world on the industry-standard Arena AI text leaderboard, competing with models 20x its size. Its focus is on latency.

• 31B Dense: Ranks as the #3 open model in the world on the industry-standard Arena AI text leaderboard. Its focus is on “maximizing raw quality,” which Google says is a powerful foundation for fine-tuning. 

Unlike Google’s flagship Gemini models, the Gemma models are open-source, encouraging developers to build on and collaborate with them, a move that has already yielded positive results. Since the first Gemma generation launch in February 2024, Google says that it has been downloaded over 400 million times. However, now it also features an Apache 2.0 license for added developer flexibility and control. The model weights are available on Hugging Face, Kaggle and Ollama.

Gemma 4 (31B and 26B MoE) is accessible in Google AI Studio, while Gemma 4 (E4B and E2B) is accessible in the Google AI Edge Gallery. Developers can also use it to power Agent Mode inAndroid Studio and to build production apps on Android using the ML Kit GenAI Prompt API.

Google has taken a nuanced stance in the open vs. proprietary model debate by making select models openly available while keeping its flagship offerings, such as the Gemini family, proprietary. This approach has plenty of merit: it fosters innovation and collaboration within the developer community, which in turn benefits Google by driving higher-quality applications across its various ecosystems, such as Android, all without compromising its intellectual property or its competitive position in the AI race.

TOGETHER WITH VENN AI

Get 50% off Venn AI’s Agent Guardrails

To be useful, OpenClaw needs access to your tools — Slack, GitHub, Notion, Jira, Google Workspace. But connecting each one creates security gaps and complex workarounds that stop people from putting agents to work.

Venn for OpenClaw is the missing security layer. Connect OpenClaw to 40+ apps in minutes. Set read/write permissions per tool, monitor full activity logs, and let your agents actually get things done without the risk of destructive actions.

Use code DEEPVIEW for 50% off an annual plan. Offer expires April 15, 2026.

GOVERNANCE

Box Agent shows when context trumps models

When adopting agents, enterprises walk a fine line between data security and usefulness. Box might have a fix.

On Thursday, Box announced the general availability of the Box Agent, a tool that lets users query and act on their data on the platform using natural language. The agent works with both structured and unstructured data, acting as a unified engine across all of your enterprise’s data without letting it bleed back into the models the agent leverages. 

“Our AI is as secure as everything else at Box,” Yashodha Bhavnani, head of AI at Box, told The Deep View. “While you get to use the best breed models, we make sure that you never have to worry about your data leaking back into the model.”

The Box Agent also isn’t backed by a single model, Bhavnani told me. Instead, it  leverages state-of-the-art models from every major AI firm including OpenAI, Google and Anthropic, and updates them constantly as new models roll out. While users can pick which models they prefer for certain tasks, if left up to the Box agent, it will pick the best model for the task at hand. 

Some of the Box Agent’s capabilities include: 

• Searching across an enterprise’s entire content library for answers on enterprise data, which limits “excessive hallucination” by grounding outputs, said Bhavnani

• Creating new files in multiple file formats and completing multi-step tasks

• Analyzing specific files or collections of files and summarizing complete documents 

• And interpreting natural language questions, taking into account organization-specific terminology and relevance. 

In the press release, Box noted that its agent can be particularly helpful for legal, procurement, HR, sales and marketing teams. Bhavnani told me that the agent is the first step in a broader journey of bringing the “powers of contextualization” to everyone in an enterprise. 

“Every customer is unique. They have their own policies, they have their own culture,” said Bhavnani. “We get to this point of contextualization, where you can truly personalize the agent for your enterprise. It truly understands the structure, and then it has the security and permission, so you feel confident using it.”

It’s no secret that agents are the future of how white-collar workers will interact with AI. Box’s new agent, however, is a great example of how the platform where we use these models may be more important than the models themselves. Box’s agent creates a unified platform where any model can be utilized on your data and context, essentially turning the models themselves into commodities. Perplexity and Microsoft’s model council systems operate in much the same way, and Apple’s reported plan to open up Siri to several AI models has the same ethos. At the end of the day, whoever has the best interface may lead the AI adoption game. 

LINKS

• Microsoft’s Mustafa Suleyman discusses “superintelligence” game plan

• OpenAI acquires tech podcast TBPN in its first media acquisition

• Clinical AI startup Kintsugi shuts down after failing to get FDA approval

• Oracle cuts 10,000 jobs in India after AI pivot

• SpaceX boosts evaluation above $2 trillion in IPO

• Trump administration will appeal court order blocking Anthropic ban

• Cursor 3: A new vibe coding experience centered around agents, allowing humans and agents to work together across multiple workspaces and repos. 

• Google Vids: The tech giant’s video platform now has AI-powered features, including custom AI avatars.

• Microsoft Models: Microsoft launched three new foundational models, called MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2

• ElevenMusic: Voice AI startup ElevenLabs has launched an app that lets people create AI music, currently available for free.

2026 Salary Report: U.S. vs Global hiring.

• Explore salary benchmarks for AI engineers, data scientists, product managers, and more.

• Discover top-tier talent from Latin America, Africa, and South East Asia—ready to hit the ground running.

• See how companies are saving up to 67% on salaries by hiring globally.

Download the report

(sponsored)

GAMES

POLL RESULTS

If you could buy OpenAI stock in 2026, do you think it would be worth the investment?

Yes (50%)
No (39%)
Other (11 %)

BIG TECH

Understanding the Claude Code leak fallout

Anthropic has dominated headlines with its flagship product, Claude Code, leaking across the web. As the dust settles, here's a breakdown of what happened and what comes next.

On Tuesday, an X user shared that Anthropic had leaked roughly 512,000 lines of Claude Code's internal source via a map file included in the npm package, the public repository where developers download and share JavaScript software tools. Anthropic has since clarified that no customer data was included in the leak and that the release was due to human error, not a security breach. 

The fact that it was a human error lends some credibility to Anthropic, assuring users that its systems were not vulnerable to external attacks. However, it does highlight a different kind of vulnerability, one rooted in internal process and oversight, as noted by Amy Chang, Head of AI Threat Intelligence and Security Research at Cisco.

“These incidents highlight the universal struggle to mitigate the impact of human error in complex development cycles,” Chang told The Deep View. “These security incidents leave practical takeaways for consumers of any generative AI technology to treat these tools as untrusted collaborators, not trusted infrastructure.”

The leak includes thousands of files showcasing not only how the model operates, such as a complex memory architecture, but also unreleased models and features that were planned to be released to Claude Code, as reported by VentureBeat and The Information, including: 

• Kairos: A collection of updates that allow Claude to work in the background and message you on mobile, similar to OpenClaw

• Buddies: Virtual duck avatars that help visualize coding agents and infuse the platform with more personality, with the hopes of getting “sustained Twitter buzz”

• Capybara: Anthropic is already iterating on Capybara v8, with Capybara being the codename for a Claude 4.6 variant that is a more powerful level of model that's a step above Opus

While no customer data was leaked, the issue does give competitors, such as OpenAI, an opportunity to look at the proprietary technology and use it to build their own tools. A Claude Code competitor even told The Information that it may change its product development plans to beat Anthropic to the release. The leak also gives bad actors insight into how the model works, which could make it more vulnerable to future attacks. 

As a result, it is in Anthropic’s interest to take the code down, and, according to a new Wall Street Journal report, by Wednesday morning, Anthropic representatives used a copyright takedown request to try to force the removal of over 8,000 copies and adaptations of the code shared on GitHub, later narrowing the request to just 96 copies. The original X post revealing the leak now has 33.2M views.

Some of the biggest risks this leak poses to Anthropic will be reputational. This is the second incident of this nature in two weeks, following the accidental publication of a blog post detailing Claude Mythos, its next flagship model. For a company that has built its reputation and differentiated itself from competitors by prioritizing customer safety and privacy, this is not a good look. Hopefully, this serves as a reminder for Anthropic to show up security before it gets painted with Silicon Valley's "move fast, break things" reputation, which could pose enormous risks for AI.

TOGETHER WITH CRUSOE

Crusoe: deploy fine-tuned models with zero infrastructure headaches

Work with our team to deploy your fine-tuned model on a platform built for performance.

Use Crusoe Managed Inference to unlock breakthrough speed and throughput without the infra overhead.

Start deploying with Crusoe

RESEARCH

Survey: AI coding shifts hiring trends

More developers than ever are relying on agents to do their work for them. 

A recent survey of 450 US software engineers from CodeSignal found that 91% reported using agentic AI coding tools, such as Claude Code, Codex and Cursor in their day-to-day work. Additionally, more than three-quarters of those engineers shipped AI-generated code into production over the past six months. 

The data adds to the broader narrative that the role of an engineer is transforming as their task load shifts from software coder to AI orchestrator. And despite fears that AI will kill the jobs of software engineers, job postings for developers are up year-over-year as novice-led vibe coding brings about the dawn of custom software that requires more in-house expertise. 

“Software development has fundamentally changed,” said Tigran Sloyan, co-founder and CEO of CodeSignal. “Engineers are no longer coding alone; they’re working with AI agents, and the best ones know how to get the most out of them.” 

It’s why, for engineers, AI skills may become non-negotiable. According to CodeSignal’s survey, 73% of engineers reported that not adopting these tools puts them at risk of becoming less competitive, and 42% reported that they’d be hesitant to hire or work with a developer who doesn’t use them. 

And as these skills become more in demand, CodeSignal debuted agentic coding assessments designed to test engineers’ AI readiness. These assessments test whether engineers can use agentic tools to build working solutions and explain their technical decisions to reviewers, rather than simply testing if they can build algorithms or write code by hand.  

“The companies that figure out how to hire for—and develop—those skills will have a real advantage,” Tigran said.

And one thing is clear: AI coding tools are accelerating development time and driving down the cost of building software. That's increasing, rather than decreasing, the need for organizations to hire more developers to connect the dots and manage the code.

As these models become increasingly powerful, demand for AI skills is also rising quickly. The problem, however, is the limited resources by which developers can learn how to leverage these tools. Modern curricula at universities can’t always keep up with the pace of change, and reskilling is becoming more popular than ever. It’s the sentiment that drives some of AI’s leaders to tell young people to abandon college altogether and just start building. Case in point: OpenAI CEO Sam Altman told an attendee at a January Town Hall meeting that for AI builders, it’s “probably not the best use of your time to be in university right now.” The question remains, in the face of such a monumental shift, whether universities and educators will be able to make the shift, too. 

TOGETHER WITH GHOST

ghost - the free database your agent is missing

traditional databases are designed to be permanent. you provision one, name it, size compute, choose a region, back it up, keep it running for years. that's fine for human workflows. it's wrong for agents.

ghost gives your agent postgres databases that are instant, ephemeral, and disposable. the mental model is git, not RDS. spin up a database the way a developer creates a branch. do work. keep it or throw it away. fork before a risky migration. merge or discard.

ghost is free. unlimited databases, unlimited forks, 100 hours of compute and 1TB of storage per month, no credit card required. your agent discovers it through MCP and starts provisioning immediately. and because every ghost database is just postgres, your agent already knows how to use it. every LLM has postgres in the weights. no SDK. no proprietary query language. just SQL.

a thousand databases for a thousand parallel sessions, free.

start building → ghost.build

BIG TECH

Why Apple forced a reckoning on vibe coding

If you think you can vibe code iPhone apps and slip them through the back door, think again.

Apple has recently cracked down on vibe coding apps that were bypassing its review process to rapidly deploy on iOS. This has even included banning the app Anything, a vibe coding platform that launched in November 2025 and allowed users to create their own apps far more rapidly than with traditional software development tools. 

Dhruv Amin, CEO of Anything, told The Information that the platform helped enable its users to publish thousands of apps in the App Store in its first couple of months, before Apple's App Store review team started rejecting its updates in mid-December for violating Guideline 2.5.2 of the App Store rules. That guideline basically states that apps need to be self-contained and shouldn't spawn other apps. 

This has also impacted other vibe coding apps including Replit, Vibecode, and Bitrig, but none of those three have been banned or removed from the App Store. 

A growing phenomenon coined by an Andrej Karpathy tweet from February 2025, vibe coding has taken the world by storm over the past 14 months. It has dramatically increased the amount of software being created by removing the need to know complex coding languages and allowing anyone to create custom software by simply describing it in natural language. 

In a statement to The Deep View, Apple emphasized:

• It is not against vibe coding and has no official policy banning vibe coding apps

• Bypassing the App Store's privacy and security safeguards that protect users is what happens when apps deliver unreviewed software within an existing app

• When an app falls out of compliance, the App Store team explains the violation and works with the software vendor to bring the app into compliance; for example, in the case of Replit, the team has had multiple conversations in 2026 and is continuing to work toward a resolution

Regarding the Anything case, Apple had no comment at this time, but we will update this story as we learn more (you can check the web version for updates).

I have to expect Apple to be moving toward an "if you can't beat 'em, join 'em" strategy when it comes to vibe coding. The company is already halfway there with platforms like Swift that streamline the process of building apps. With its new Google partnership to use Gemini models, Apple will have the AI firepower needed to launch its own vibe-coding tools built on Swift and potentially even a lighter version of Xcode. Strategically, this would also be a perfect fit for Apple's brand mission of empowering creativity at the individual worker level. And Apple does it through its own tools, I expect it would also streamline the process of submitting vibe-coded apps for App Store review. Let's hope a move like that would also cascade to other vibe coding apps like Replit and Anything.

LINKS

• Claude Code users hit usage limits faster than expected

• SpaceX has filed confidentially for an IPO, beating OpenAI and Anthropic

• Amazon’s Rufus AI’s ads drive less traffic than traditional ads

• Perplexity AI accused in lawsuit of sharing user data with Meta, Google 

• Secondary market shifts from OpenAI to Anthropic

• Mass robotaxi failure in China traps passengers, causes traffic disruptions

• Codex: The GitHub plug-in allows users to review issues, commit changes, and more

• Flow: Veo 3.1 Lite, Google’s newest model, is available in the Flow app 

• NotebookLM: A new featured notebook focuses on Benjamin Franklin 

• Notion AI: The iOS beta got new features and improvements

• Tencent: Senior Researcher

• Meta: Research Engineer, Language - Personalization, Meta Superintelligence Labs

• Nvidia: Senior Neural Graphics Scientist

• Canva: Staff Research Engineer - Generative Video

GAMES

A QUICK POLL BEFORE YOU GO

BIG TECH

OpenAI and Anthropic near a scary leap forward

OpenAI and Anthropic are about to launch new AI models for a world changed by OpenClaw. 

Anthropic has its "Claude Mythos" model, and OpenAI is preparing to launch its "Spud" model. Both companies are touting these as extremely powerful models that represent the next big leap for LLMs. And if that's the case, despite the codenames, these models could roll out as Claude 5.0 and ChatGPT 6.0. 

There are several factors to watch:

• OpenAI CEO Sam Altman said the new Spud model is so powerful that it will "really accelerate the economy"

• Anthropic CEO Dario Modei said that the company is already testing Claude Mythos with early access customers

• A leaked Anthropic blog post obtained by Fortune stated that the Anthropic team believes its model is so powerful that it could pose unprecedented cybersecurity risks, and that's one of the reasons why Anthropic has been warning government officials and giving organizations early access to the model in order to help them prepare

• OpenAI's Spud could also be the foundation of the company's new "superapp" that will reportedly combine the desktop apps for ChatGPT, Codex (its coding tool), and Atlas (its web browser) into one streamlined experience with advanced AI agent capabilities 

• As another nod to how powerful OpenAI believes its next model will be, the company reportedly renamed the product organization led CEO of Apps Fidji Simo to the "AGI Deployment" team

One reason these models are advancing so rapidly and reaching a higher stage of development is that the labs are now using the models themselves to help build and improve the models. Sam Altman mentioned this when the company released GPT-5.3-Codex. This is a process called recursive self‑improvement (RSI), and it has long been anticipated as the point where AI systems will make a dramatic leap forward.

We've already seen AI hit a new gear in 2026 with personal AI agents unleashing a wave of automation, productivity gains, and software improvements. That was led, of course, by OpenClaw. And now we're seeing the benefits OpenClaw has unlocked being adopted by AI and software companies across the tech industry. That agentic movement is about to get supercharged by these new models. But as we give the agents more rocket fuel, it's important to remember that they're non-deterministic systems that can go rogue and do things we didn't expect. That's going to make it more important than ever that we build good harnesses and systems to contain these agents to protect against the worst possible outcomes.

TOGETHER WITH SNYK

Building an AI Security Champions Program

Join Snyk Staff Developer Associate, Sonya Moisset, on April 9 at 11 AM ET for a live workshop: Building an AI Security Champions Program.

Learn how to:

• Adapt the champion role for AI-driven workflows

• Enable developers to secure AI-generated code

• Define success with real KPIs

Plus, earn 1 ISC2 CPE credit. Save your spot.

BIG TECH

Google Veo pushes video AI forward, cuts prices

OpenAI may have shed Sora, but Google’s AI video ambitions are far from over. 

On Tuesday, the company announced Veo 3.1 Lite, the latest edition to its video generation family of models and its “most cost-effective video model,” the company said in its announcement. Google said that Veo can build high-volume video applications at half the cost and roughly the same speed as its previous Veo model. 

Veo 3.1 Lite supports both text and image inputs, and can generate both landscape and portrait ratios at resolutions up to 1080p. Veo 3.1 Lite can cost as little as 5 cents per second, compared to 40 cents per second for Veo 3.1 standard. The model is available to developers in both the Gemini API and Google AI Studio. 

Additionally, Google is giving users a discount on Veo 3.1 Fast, its mid-range video generation model, starting April 7, cutting generation costs to 10 cents per second for 720p and 12 cents per second for 1080p. 

In a post on X, Logan Kilpatrick, a member of technical staff for Google DeepMind, said that video is “here to stay.” 

Twitter tweet

Google’s courting of AI video customers comes as OpenAI casts off its own video-generation efforts by ditching Sora and ending its $1 billion, three-year licensing partnership with Disney. The company’s Sora switchup signals a broader refocusing of its compute towards more revenue-generating products and redirecting its video model staff to world models. Since video models are incredibly compute-intensive, it makes sense that the video model was on the chopping block. 

“We cannot miss this moment because we are distracted by side quests,” Fidji Simo, CEO of applications at OpenAI, told staff in an all-hands meeting, according to Business Insider.

But Google isn’t the only one trying to fill the vacuum left in Sora’s wake. Elon Musk-owned xAI is “doubling down” on AI video with the next release of Grok Imagine, the lab’s own video generation model. But as xAI faces an ever-growing pile of lawsuits over its image generation capabilities, Google may have a better shot at appealing to a wider audience.

Google’s ability to press further into video generation while OpenAI backs off is a prime example of the company’s edge in the AI market. While AI companies scrap for every morsel of compute they can get their hands on, Google sits atop a legacy. The company spent years building Google Cloud into one of the big three cloud hyperscalers, giving it a much-needed wealth of compute resources. Additionally, having cemented its billions in the digital advertising and search markets, Google doesn’t have to worry about the revenue picture as much as rivals like OpenAI and Anthropic. Because of this, Google can take on the “side quests” as Simo called them, cornering a market while it’s still nascent. 

TOGETHER WITH VIKTOR

Forget single-purpose AI agents, Viktor gives you a full-blown AI peer.

Viktor is an AI coworker that lives in your Slack or Microsoft Teams. No hardware, no setup, enterprise security from day one. It connects to 3,000+ tools (Stripe, HubSpot, Meta Ads, GitHub, Notion), delivers board-ready PDFs, builds web apps, and runs scheduled tasks 24/7. SOC 2 certified. Your data never trains models. Watch Viktor work ->

Backed by Daniel Gross (Y Combinator), Nat Friedman (ex-GitHub CEO), and Mati Staniszewski (ElevenLabs). See why ->

Give Viktor your worst task (for free) ->

PRODUCT

Slack adds AI meeting notes and a desktop AI agent

For years, Slack has been where teams go to communicate. Now, it's setting its sights on something bigger: replacing your favorite AI apps without you ever leaving the platform.

On Tuesday, Slack announced more than 30 new capabilities that not only expand what its Slackbot AI agent can do but also introduce new features previously unavailable in Slack. For instance, one of the biggest highlights is a new meeting transcription and note-taking feature that acts much like Otter.ai or Fireflies.ai. 

“For sales teams, the personal note taker is huge, because they often have their email open, the Zoom meeting open, and half paying attention because they're working on something for the meeting that they're in, so I think that's immensely powerful,” Rob Seaman, EVP & GM, Slack, told The Deep View. 

To use the feature, users just have to have the Slack desktop app installed. Then, when activated, Slackbot can listen in the background of your meetings when activated to capture the content, take notes, produce summaries, identify action items and more. The best part is that the meeting can take place in other apps, such as Zoom or Google Meet. Some other new features include: 

• Operator mode: Slackbot can complete multi-step tasks across your desktop app on your behalf. All users have to do is select anything on their screen and ask Slackbot to act on it, such as "Summarize this” or “Draft a follow-up.” 

• AI skills: Slackbot can access a built-in library of AI skills or instructions on how to do something, as well as user-built ones, for repeatable, shareable workflows. 

• MCP: The new MCP client allows Slackbot to access any tool, including enterprise apps such as Google Workspace, Microsoft 365, Notion, and more. 

• CRM: Slack now has a native CRM, in which Slackbot can read channels to log deals, contacts, and call notes. This is meant to help small businesses that lack a CRM or a sales department.  

All of the new features are meant to build on the app’s existing AI agents, including Slackbot, which, the company says, is on track to become the fastest-adopted feature in Salesforce history, less than two months after its January launch, when it was redesigned as a deeply personal AI agent for work. Slackbot will be available to all Free and Pro Slack users starting in April. 

"We're building a personal agent [Slackbot] that's got all of your context in Slack, but is immensely capable, and this is like a highly intelligent teammate that can help you get your work done throughout the day, but also interface with those other agents,” added Seaman.

One of the biggest bottlenecks in workflow optimization is the need to constantly context-switch between different apps to get things done. However, most companies already rely on a single communications app, constantly checked for collaboration, as their central hub. If that hub could also provide access to other tools, such as a transcription app, it would help solve the context-switching problem. That's exactly what Slack has done. This is an example of enterprise software that delivers genuine value to users rather than just AI-washing about new add-on features. It's a trend I hope is here to stay.

LINKS

• Iran to target major US tech firms in middle east, including Apple, Google

• OpenAI boosts historic funding round to $122 billion at $852 billion valuation

• Alibaba’s new model is proprietary, pivoting away from open-source

• Runway launches $10 million fund to support AI startups

• California’s Gov. Gavin Newsom issued an executive order on AI

• Northwestern University study shows federal judges use AI in judicial work

• CoreWeave Inc. raised $8.5 billion for cloud computing capacity expansion

• Codex: OpenAI introduced a Codex plugin for inside Claude Code

• Microsoft Copilot: Council lets users run multiple models on same prompt simultaneously 

• Speechify Voice AI for Windows: Now available to Microsoft Windows users

• Perplexity API Platform: The AI search engine’s API now covers the full agent stack

• Scale AI: Research Scientist, Agent Robustness

• Affirm: Senior Staff Machine Learning Engineer, (ML Underwriting)

• Palo Alto Networks: Principal Machine Learning Engineer

• Amazon: Applied Scientist, LLM Code Agents, Kiro Science

GAMES

POLL RESULTS

Do you think the Americans public's opinion of AI will broadly be favorable by the end of the decade?

Yes (42%)
No (49%)
Other (9 %)

RESEARCH

New data: Americans turn more negative on AI

AI is losing the narrative among Americans. 

In a poll of roughly 1,400 US adults conducted by Quinnipiac University, 55% reported that they felt AI would do more harm than good, up 11% year-over-year. Around 76% reported that they trust AI “hardly ever” or “only some of the time,” and 62% reported not being excited about the technology. 

Along with broadly feeling disenchanted with the tech, more than half reported that AI is moving far faster than they expected. 

Poll respondents were also generally wary of the tech’s impact on the job market and workforce: 

• More than 80% reported that they wouldn’t take a job where an AI program acted as a manager, assigning tasks or schedules

• Around 70% believe that AI will cut the number of job opportunities, up 14% from the previous year’s poll

• And 30% reported being concerned that AI would make their own jobs obsolete, up 9% year-on-year. Despite broader job market concerns, 48% reported that they were not concerned at all about AI taking their own jobs. 

"Americans are more worried about what AI may do to the labor market than about what it may do to their own jobs,” Tamilla Triantoro, associate professor in the Quinnipiac University School of Business, said in the report. “People seem more willing to predict a tougher market than to picture themselves on the losing end of that disruption - a pattern worth watching as the technology moves deeper into the workplace."

The poll reflects growing concern and confusion caused by seemingly endless conflicting reports of how AI will impact the job market. Some estimates suggest that AI is already capable of automating a large share of work hours, while others suggest that AI is increasing workloads rather than reducing them. Though one viral report from February painted a bleak, worst-case-scenario picture of how AI could utterly demolish the economy, software engineering jobs are up from last year.

The problem with the growing pile of contrary data is that people largely have no idea what to make of it. And often, when confronted with something new, the natural human reaction isn’t to accept it with an open mind, but to rebel against it – especially if that thing threatens everything from their livelihood to the way they live their life. To make people comfortable with it, AI firms may need to shift away from the image that their tech is here to take people's jobs, and instead offer ways to help people gradually transform the work they do without stripping them of their sense of purpose. But that may be a difficult feat, as trigger-happy companies continue to slash their workforces, using AI as an excuse, whether the tech is ready or not. 

TOGETHER WITH TWELVELABS

Indexing Video Just Got A Whole Lot Easier

Most video search tools collapse everything, visuals, audio, motion, into a single signal. It works in demos. In production, your queries compete with the wrong data and retrieval gets worse as your library grows.

TwelveLabs encodes each modality separately. Speech queries route to speech. Visual searches don't get muddied by audio. And accuracy holds whether you have 100 hours of footage or 100,000.

Teams have replaced hours of manual review with queries that return results in seconds.

See it for yourself. Try the Playground free, or schedule a demo with the TwelveLabs team.

PRODUCTS

4 AI tools worth recommending to anyone

People outside the tech bubble don't always realize that the world of AI is bigger than ChatGPT and Claude. 

For your friends, family, and community, you have the opportunity to help them take advantage of some of the best things that the current AI revolution has to offer and avoid feeling like a victim of the changes that AI is bringing to work and life. 

With that in mind, I have four tools that I'd recommend to everyone:

1. Wispr Flow: One of the most common ways that people have wanted to interact with Siri, Alexa, and Google Assistant for years is to dictate emails, text messages, and other content for documents. But the quality has been so mediocre that it's become laughable at times. Wispr Flow uses AI to turn your dictation into clearly enunciated, well-punctuated text. While the old tools worked about 70% of the time and regularly produced odd anomalies, Wispr Flow works about 95% of the time and often cleans things up for you. While the average person types about 50 words per minute, they can speak 150-200 words per minute, so there's an opportunity for a 3-4x productivity boost. 

2. Comet or Atlas: Since most people are used to web browsers, installing an AI browser can be the best way to start taking advantage of chatbots' unique capabilities. You basically get a chatbot prompt instead of a Google prompt. Perplexity's Comet browser is a great choice because it will automatically route a question to the chatbot that can handle it best. And for people who are already used to using ChatGPT, the Atlas browser will be an easy transition.

3. Perplexity Discover: One of the best AI tools that's flying under-the-radar is Perplexity Discover, which I'd argue is the best custom news app you can find right now. You can access it in the left toolbar on Perplexity.ai, and it's also available in the Perplexity mobile app, where you simply click the icon in the upper right corner. You can use the app to browse Top Stories, Tech & Science, Business, and other main categories. The real value comes from the For You section, which uses AI to find the stories you're most interested in over time. I find it curates far more accurate and useful selections than Google Discover, for example. 

4. ChatGPT Voice: We started the list with audio and voice AI, and that's where we'll end. ChatGPT Voice is an incredibly helpful and easy way to access the chatbot from your phone by using your voice to ask questions and get concise audio responses in return. You can even pick the style of voice you'd like to interact with. Just open the ChatGPT app and tap the little blue icon. Or if you have an iPhone with an Action Button, this is the perfect thing to assign to it by using the Shortcut that automatically shows up if you have ChatGPT installed. 

Other tools like Granola for capturing meeting notes and Lovable for creating your own applications are also great for beginners, but are a step up in complexity from the four tools mentioned above.

Make no mistake: AI is still losing the popularity contest among the broader population right now. It’s far less common to find an AI optimist than it is to find people who fear and revile the tech. If you'd like to help the people in your life who are outside the tech bubble start to understand some of the positive things that AI has to offer, these four tools may be the place to  start. However, it’s going to take more than tools to flip the script on AI. Education on the risks, rewards and reality of AI needs to take place alongside adoption. 

TOGETHER WITH VENN AI

Finally, Trust Your OpenClaw Agents at Work

Real agent productivity without the risk

To be useful, OpenClaw needs access to your tools like Google Workspace, GitHub, Notion, Slack, and Jira. But making unverified and unstable integrations for each app creates complex workarounds and real security risks, and that's exactly what's stopping teams from putting OpenClaw to work.

Venn for OpenClaw adds the missing security layer so you can safely connect OpenClaw to 40+ business applications without the complexity. With a single MCP connection you can:

• Set tool-level permissions like read and write access per app

• View full activity logs of everything your agents do

• Optimize tool selection to reduce context window usage and improve agent performance

• Trust OpenClaw agents to get work done

Start a free trial and get 50% off an annual plan with code DEEPVIEW. Offer expires April 15, 2026.

ENTERPRISE

Snowflake’s AI push counters SaaSpocalypse fears

Snowflake has been a stalwart of the SaaS economy and a leader in enterprise data for the past decade. But the company is deep in the middle of a transformation that most people haven't recognized yet, said Baris Gultekin, vice president of AI at Snowflake. 

In the latest episode of The Deep View Conversations, senior reporter Nat Rubio-Licht talks with Gultekin for a candid look at how the company is navigating the AI era and what it's learning in real time, shifting both how the company operates internally and how it’s customers data needs have changed. 

“Snowflake sits in a different place … as AI becomes a core way of interacting with data and with services, a data platform is a very very helpful product,” said Gultekin. “Ultimately, the agents that you build are as good as the data that you provide to them.” 

Gultekin talks openly about how the entire team inside Snowflake is now using AI coding tools to build skills and automate their work. That includes non-developers who are using Project SnowWork, an AI agent for professionals across all roles.

Baris joined Snowflake in 2023 through the acquisition of blockchain startup nxyz and has spent the past three years running the AI teams inside the enterprise tech giant. He also brings a rare perspective from his time working on Google Assistant in the pre-LLM era, which gives him a unique lens on how much has changed.

“What has surprised me is the pace of innovation: how fast we’re moving,” Gultekin said. “The industry has created this tremendous opportunity, and everyone is working very hard in AI to capture the momentum.”

Don’t miss this episode.

It’s undeniable that AI is shifting the way that software companies operate. As AI threatens to bring about the so-called “SaaSpocalypse,” enterprises have been forced to think on their feet. Snowflake has managed to stay agile without straying from its roots, giving enterprises a clear picture of the very thing that sits at the heart of AI: Data. Though executives of companies like Salesforce and Workday have downplayed the severity of the SaaSpocalypse, legacy tech providers may need to adopt a similar flexibility to figure out what they can bring to the table in the AI era.

LINKS

• Mistral raises $830 million for Nvidia-powered data centers in Europe 

• Starcloud raises $170 million to build data centers in space

• All of xAI’s 11 original cofounders have reportedly ditched the company

• AI chip company Rebellion raises $400 million at $2.3 billion valuation

• Perplexity, Fiscal AI expand partnership for financial data for investors

• Coatue reportedly projects nearly $2 trillion valuation for Anthropic by 2030

• Qwen3.5-Omni: The latest state-of-the-art model by Alibaba, with understanding of text, images, audio, and audio-visual content.

• Claude Code: Computer use is now available on Claude Code for Pro and Max users, allowing the AI assistant to open your apps and click through your UI.

• Critique: Microsoft’s new multimodal deep research model available in 365 Copilot, leveraging multiple models to find the most optimal responses.  

• Notion MCP: Notion now connects several AI tools, including ChatGPT, Claude, and Cursor, directly to your workspace using Model Context Protocol.

• Manus: You can now use your phone to start tasks, access files, and run workflows by remotely controlling your computer with the Manus Desktop app.

• Capital Group: Lead AI AppSec Engineer

• Scale AI: Research Scientist, Frontier Risk Evaluations

• Google: Software Engineer, Quantum Open Source, Quantum AI

• Meta: Fundamental AI Researcher - FAIR

GAMES

‘

A QUICK POLL BEFORE YOU GO

AI lifts Snowflake as SaaSpocalypse spreads

Snowflake has been a stalwart of the SaaS economy and a leader in enterprise data for the past decade. But the company is deep in the middle of a transformation that most people haven't recognized yet.

In this episode of The Deep View: Conversations, senior reporter Nat Rubio-Licht talks with Baris Gultekin, vice president of AI at Snowflake, for a candid look at how the company is navigating the AI era and what it's learning in real time.

Gultekin talks openly about how the entire team inside Snowflake is now using coding tools to build skills and automate their work. That includes non-developers who are using Project SnowWork, an AI agent for professionals across all roles.

Baris joined Snowflake in 2023 through the acquisition of blockchain startup nxyz and has spent the past three years building and running the AI teams inside the enterprise tech giant. He also brings a rare perspective from his time working on Google Assistant in the pre-LLM era, which gives him a unique lens on how much has changed.

Topics covered:

• Why there is no AI strategy without a data strategy, and what enterprises keep getting wrong

• How agentic AI has shifted enterprise data from question-answering to automation

• The SaaSpocalypse and why Snowflake sees AI as a tailwind rather than a threat

• Cortex Code (CoCo), Snowflake's coding agent that lets customers query their data in plain language instead of SQL

• The governance and security challenges that come with multi-agent systems

• How Baris uses coding agents in his own life

If you want to understand how a mature SaaS company reinvents itself inside an AI revolution, this one is worth your time.

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

Subscribe to The Deep View: Conversations podcast in any podcast player for more unique conversations with the brightest minds solving the biggest challenges in AI. You can also subscribe on YouTube.

PRODUCTS

Claude’s rise brings outages and usage limits

Claude is exploding in popularity, and that has consequences. 

Anthropic's quirky chatbot has been on a tear during the opening months of 2026. In fact, it's been gaining new users so quickly that it's facing serious growing pains. In recent weeks, Claude has faced a series of outages that have caused it to dip below the 99% uptime standard for most applications. 

And because of Claude's growing popularity and the company's difficulty handling the rapid influx of new users, it is now adjusting how users burn through their limits during peak hours. On weekdays between 8:00 am and 2:00 pm ET, users across the Free, Pro, and Max tiers will all hit their session limits faster, an Anthropic engineer explained on X. 

He estimated that the changes will impact about 7% of users and recommended that people shift highly intensive background jobs to non-peak hours.

Twitter tweet

So what does this mean in practice?

• It's surge pricing for chatbots: This is essentially the same thing Uber does for ride costs during rush hour. But instead of changing more money, Anthropic is trying to change user behavior to spread out the requests made to Claude.

• Most users won't be impacted: If all you do is use Claude to ask questions and help create documents, then this move is good for you, as it will likely increase Claude's uptime and performance for those tasks.

• Peak hours cost more per action: Before this change, 1 prompt used 1 unit of your Claude usage budget, and the cost stayed the same all day. Now, during peak hours, 1 prompt might equal 1.5 to 2 units (Anthropic hasn't said what the exact formula will be). 

The move doesn't come as a surprise, since Claude's growth has been so meteoric in such a short time. It's no secret how expensive it is to run chatbots and how hard it is for these companies to obtain the GPUs and compute necessary to grow. A report from consumer transactions analysis company Indagari and analyzed by TechCrunch (see chart below), shows Claude's paid users skyrocketed in the first two months of 2026. 

Chart credit: TechCrunch

While Claude's growth is impressive, we should also put it in perspective. Anthropic doesn't say how many monthly active users Claude has. Even the most optimistic estimates put the number between 30 million and 180 million. Meanwhile, OpenAI's ChatGPT is at 900 million. So ChatGPT remains 5x to 20x larger than Claude. And while Claude is rapidly onboarding new users, ChatGPT still signs up twice as many new subscribers each week. Nevertheless, since mid-2025, most of the people I know who use generative AI most intensively have been saying they've made Claude their primary chatbot because they perceive it produces higher-quality results. With a lot of people switching from ChatGPT to Claude in recent months, it will be important to see whether they share the same perception and stick around. Many people tend to choose the easier option — including the one that doesn't suffer outages — over the higher-quality one. That still leaves plenty of opportunity for competitors to win over AI users in the months and years ahead.

TOGETHER WITH SLACK

The AI That Already Knows Your Work — Meet the New Slackbot

You may know Slackbot as the familiar helper inside Slack — today, it's something entirely new. Slackbot has been rebuilt for the agentic era as a personal agent that understands your work, your context, and how decisions get made. 

Unlike other chatbots, it's grounded in your conversations, files, and workspace context, so it provides support that's deeply personal to you — without asking you to explain yourself or change how you work. With Slackbot, simple conversation leads to real action: organizing information, drafting content, and moving work forward without ever leaving Slack. 

In this on-demand webinar, hear from customers like Engine about how they're already using Slackbot, and see a live demo from the team that built it. Walk away with real tactics for bringing personalized AI into your team's daily collaboration. 

Watch now →

BIG TECH

Reddit, Wikipedia fight AI while feeding it

Wikipedia and Reddit built their reputations on human-curated content. In the age of AI, both platforms are fighting to stay human-first.

On Wednesday, Reddit CEO Steve Huffman emphasized the company’s mission to keep humans at the center, keeping Reddit a place where “people can talk to people.” In the Reddit post, Huffman included plans to prevent confusion between AI-written and human-written content. 

The biggest change is that any accounts that use automation for allowed tasks, also known as “good-bots,” will be clearly labeled as “[App],” signalling to users that they are interacting with a machine instead of a human. Other changes include:

• Removal of bots: “Bad bots” or nefarious bot content will continue to be removed before users see it. At the moment, that's about 100K accounts per day, according to Huffman. 

• Human verification: Reddit may occasionally ask users to verify they're human if content appears AI-generated, though this will be rare and focused on humanity rather than identity.

• Reporting: Users will be able to report suspicious content more easily. 

Huffman also acknowledged that many people use AI to write, so Reddit isn’t curbing all AI-generated content. Overall, the comments were positive, with Reddit users celebrating that it would remain human first.

Reddit’s announcement followed a similar decree from Wikipedia, which last week enacted a new policy prohibiting the use of large language models to generate or rewrite article content. At most, editors may use it for copyediting or translation. 

These changes come as demand for human-created content rises and the lines between authentic and machine-generated content become increasingly blurred. 

Reddit's web traffic has continued its upward trajectory, with SimilarWeb data showing approximately 12% year-over-year growth in February. Wikipedia has not seen a similar boost as Reddit in the AI era, with traffic declining 8.8% year over year in February, according to Similiarweb data. 

An emphasis on human content could help both platforms yield more immediate gains with readers, according to Joseph Levi, co-founder and CEO of Noise Media, who is exploring how to make brands stand out in the age of AI.

“For Reddit and Wikipedia, cracking down on AI-generated content is ultimately about protecting trust, and trust is what underpins their value in both SEO and AI discovery,” said Levi. “Search engines and AI tools increasingly reward credible, distinctive, human-informed content, so preserving that standard is likely to strengthen their authority over time.”

Both companies currently have deals with AI companies to train on their data, making their positions on AI-generated content all the more nebulous. Wikimedia Enterprise has deals with Amazon, Meta, Google and Perplexity, and Reddit has a long-standing partnership with Google.

Keeping bots off platforms helps preserve their integrity, but there's an obvious paradox: Licensing that same data to AI companies to train their models only fuels the very problem they're trying to solve.

One of AI-generated content's biggest appeals is its ability to deliver exact answers in conversational language, feeding into a natural human desire to do less work, a temptation that's hard to resist. As a result, simply producing high-quality content is often no longer enough to compete, even as more people come to value human-first platforms. To systematically curb the spread of AI-generated content, meaningful limitations on what these models can train on are needed. A logical first step is for platforms to stop willingly surrendering their proprietary content, and we're already seeing publishers moving in that direction. That sets up both legal battles against AI companies and a possible future in which LLMs will have to pay to train data published on the open web. That could put traditional publishers as well as sites such as Wikipedia and Reddit in a stronger position.

TOGETHER WITH MERGE

What hundreds of product and engineering teams told us about building agentic integrations

Hundreds of AI product and engineering teams told us the same three things:

1. MCP servers are overwhelming their engineering bandwidth

2. Agent-driven actions are creating security blind spots

3. They have no clear signal on which integrations to build first

So we went deeper and put together our findings in the 2026 state of agentic integrations report. Here's what the best teams are actually doing about it.

Get your copy →

PRODUCTS

Codex gets plugins as OpenAI goes enterprise

Another day, another coding tool upgrade, this time from OpenAI. 

On Thursday, OpenAI began rolling out plugins in its agentic coding platform, Codex. OpenAI describes plugins as “installable bundles for reusable Codex workflow” that can contain skills, apps, and MCP servers, with more components coming soon.  

Twitter tweet

Simply put, plugins let users integrate ready-made workflows or app integrations, such as Google Drive, GitHub or Slack, into their Codex workspace. These plugins, created by OpenAI, are available in the Codex directory, but users can also create their own plugins or import them from another ecosystem. 

Beyond expanding what users can do in Codex, plugins also streamline processes and make it easier to share setups across projects or teams. These advantages will likely boost developer productivity, but the real appeal is that Claude Code users have had access to this for a while, underscoring OpenAI's recent push to capture a larger enterprise audience. 

That push appears tied to a larger goal: preparing for a possible IPO, something CFO Sarah Friar acknowledged directly when pressed on the topic.

“Over the long run, look, we have to build a company that’s ready to be a public company, ” Friar told CNBC.

Part of that effort means cutting costs and leaning into more lucrative opportunities, such as enterprise. 

This week, OpenAI announced the shutdown of its Sora generative AI video platform and app, just six months after launch, and, as a result, ended its $1 billion content partnership with Disney. Separately, the Financial Times reported that OpenAI was “indefinitely” putting on hold its ChatGPT “adult mode,” which was supposed to allow users to create erotic content.  

Friar added that the enterprise business is “a very profitable business at scale, and that’s how we will build a sustainable business model.” Currently, 60% of OpenAI's revenue comes from consumers and 40% from enterprise, but Friar expects that to reach a 50-50 split by year's end.

Anthropic is the clearest example of the success that comes from enterprise businesses, as the company has focused on enterprise offerings, and as a result, roughly 80% of its revenue comes from enterprises, according to CEO Dario Amodei.

Having closely tracked OpenAI and the broader AI industry since 2022, I find this more focused approach encouraging, as it finally addresses a persistent problem: AI for the sake of AI. The race has occasionally seemed oriented around producing the flashiest developments rather than the most practical ones, prioritizing novelty over utility. OpenAI's new enterprise focus should help ensure these tools are actually solving real problems for real people, which is a win for everyone.

LINKS

• Anthropic admits testing a new model much more capable than predecessors

• Physical Intelligence is looking to raise $1B in funding at an $11B valuation

• Musk considers allocating up to 30% of SpaceX IPO to retail investors

• Meta’s Oversight Board says “Community notes” aren’t fact-checking

• Anthropic IPO could come as soon as the fourth quarter 

• OpenAI’s ad pilot surpasses $100 million annualized revenue

• Codex: Plugins now allow users to work with popular work apps, such as Slack 

• Runway: The content-generating platform launched an Ad Concepter App

• Search Live: Google’s AI Search feature is now accessible globally 

• Suno: The AI music platform launched Suno v5.5, with voice cloning and custom models

• AMD: Fellow AI Compiler Engineer

• Anthropic: Developer Education Lead

• Amazon: Applied Scientist II, Prime Video Recommendation/Search Science

• Visa: Staff Research Scientist- GenAI Foundational

GAMES

POLL RESULTS

Do you think Apple can still play a meaningful role in the AI ecosystem?

Yes (77%)
No (21%)
Other (2%)

BIG TECH

Can Siri win AI on interface instead of models?

Apple is turning AI into a choose-your-own-adventure game. 

The company reportedly intends to open up its Siri voice assistant to rival AI models as part of an overhaul in the upcoming iOS 27 update, according to a Bloomberg report from Mark Gurman. The system, reportedly called “Extensions,” allows users to process requests and queries via the AI system of their choosing, provided they install the app that runs that model (ChatGPT, Claude, DeepSeek, etc.).  

As it stands, users can currently access ChatGPT through Siri as part of its partnership with OpenAI. Bloomberg also noted that this arrangement is separate from the company’s partnership with Google to leverage its Gemini models to rebuild Siri’s underlying technology.  

Gurman noted that Extensions will allow agents from installed apps to work with Siri and other on-device features. The update, which is currently in testing, may also allow Apple to take a cut on AI premium subscriptions sold through the App Store. 

The reports come ahead of the company’s June 8 Worldwide Developer Conference, which is promising major AI announcements. It’s also the latest rumor related to Apple’s long-awaited AI upgrade to Siri. Bloomberg also reported earlier this week that Apple intends to launch a revamped Siri with a chatbot app similar to ChatGPT and Claude. Still, any news of Siri’s makeover was missing from Apple’s announcement about WWDC 2026. It's not even hinted at in one of Apple's creative and cryptic teasers in the event invitation.

Apple has long struggled to make a place for itself in AI. It's taken a wait-and-see approach that has largely backfired in the wake of an AI industry advancing at light speed. Though its initial plan seemed to be riding the coattails of major model firms like OpenAI and Google, its new strategy could significantly widen its possibilities.

This is probably the smartest move Apple could have made. In this scenario, Apple doesn’t have to bet on one company to win the AI market. Instead, it’s piggybacking off of every model provider’s hard work. If it partnered with only one provider, the perception would be that it outsourced one of the most important parts of tech innovation today. By being a platform and opening up to other models, Apple could focus on becoming the best interface for leveraging any AI model and could benefit from the latest models' advances in real time. From a user standpoint, the models themselves become backend commodities in that scenario. Meanwhile, Apple gets to own the user experience of AI without having to make the massive investment required to become a frontier lab.

TOGETHER WITH RIME

61% of listeners chose Rime’s AI voices over Google and ElevenLabs.

If you're building agents that talk to customers, the voice matters.

Rime's AI voices are built for enterprise: 

• Low latency 

• On-prem or cloud hosting 

• Preferred by real listeners in independent testing 

→ Try Rime free + get $100 in free credits

RESEARCH

Gartner predicts major drop in AI token costs

As AI applications become more advanced, they require more tokens, and overall costs continue to rise. That may not always be the case.

Recent research from Gartner finds that, by 2030, performing inference on a large language model with one trillion parameters will cost AI providers 90% less than it did in 2025. Furthermore, the research firm predicts that LLMs in 2030 will be up to 100 times more cost-efficient than the earliest models of similar size developed in 2022.

“These cost improvements will be driven by a combination of semiconductor and infrastructure efficiency improvements, model design innovations, higher chip utilization, increased use of inference-specialized silicon, and application of edge devices for specific use cases,” Will Sommer, senior director analyst at Gartner, wrote in the post. 

The research included two different forecasts that offer differing projections based on how chips are applied during training:

• Frontier scenarios: A model trained exclusively on leading-edge chips; costs come in significantly lower.

• Legacy blend scenarios: A model trained on a mix of legacy and leading-edge semiconductors; costs are considerably higher.

While either approach leads to a significant drop in costs compared to today, it won’t necessarily translate into token cost savings for enterprise customers. Frontier intelligence will demand significantly more tokens than today's mainstream applications, such as the lower cost of basic chatbots versus more expensive agentic AI, according to Gartner. 

“Integrating the next largest, most capable foundation models into every new feature will be commercially insolvent at scale,” Sommer told The Deep View. “As models become larger and more complex, they require significantly more tokens, and those tokens become more expensive to generate.” 

As a result, to get the most value, Sommer recommends that enterprises diversify the models they use, assigning routine, high-frequency tasks to smaller, more efficient models, while reserving frontier models for high-margin, complex reading tasks. 

“Enterprise leaders can exploit falling token [costs] by developing a roadmap that targets progressively larger, high-value domain problems to solve, planning ahead to ensure they have the data they need for tuning,” added Sommer.

The rise in popularity of agentic solutions like OpenClaw has led to a higher demand for tokens and soaring costs. While leading-edge developments in the AI space are happening quickly, to take advantage of them, enterprises need to be willing to invest. In a way, it is a relief that hardware developments will make AI models more affordable, but as model capabilities continue to increase rapidly, those savings won’t always trickle down. That points to a call for action many have been advocating: pausing further development to optimize workloads for token costs more effectively. This is a problem that many players in the AI ecosystem are trying to solve right now, and The Deep View covers it regularly. It ultimately comes down to not just paying Frontier Labs a premium price to use their most expensive models, but matching models to tasks to save money and improve performance and accuracy.

TOGETHER WITH DESCOPE

Take AI agents and MCP servers from playground to production

Every organization is exploring how to adopt AI agents or MCP servers, but how many of them are in production?

And if they aren't in production, how likely is it that authentication, access control, and agentic identity concerns are the reason?

Watch this on-demand webinar to learn:

• Real-world MCP and agentic AI use cases

• Identity challenges that prevent production-readiness

• Actionable tips to build secure, scalable AI agents and MCP servers

Move fast on AI without breaking things. Watch the webinar now.

CULTURE

The illusion of sentient AI is growing

Wait, what do chatbots have to do with robots and AI agents? 

It's a natural question, and as more new people start their journey using ChatGPT, Claude and Gemini, it will come up again and again, especially as these chatbots evolve into agents. I'd like to give you some language that can be useful when your friends, family, and community ask you about AI and occasionally misunderstand how it works.

One way I explain these tools to people who are getting started is that talking to a chatbot is like texting a friend who's an expert in an unbelievable number of topics and has infinite time to respond to your messages. 

Because chatbots have gotten so good at mimicking human speech, it's easy to confuse their dialogue with actual consciousness. And this is where a lot of perspectives diverge in the AI industry. Full disclosure: I don't think we're talking about anything close to consciousness here. It's just that humans are capable of anthropomorphizing anything. As it turns out, that includes ones and zeros. 

And AI agents are about to take it to a whole other level. 

Chatbots and AI agents have one critical element in common. The reason both seem so real is that they are unpredictable, just like humans. But that's because these are non-deterministic systems. Understanding what that means is one of the keys to understanding the current AI revolution, and not getting fooled into thinking these systems are something that they're not. 

The fact that these are non-deterministic systems means:

• Different answers each time: LLMs don't give you the exact same response every time you ask the same question. That's because they simply pick words based on probabilities. In practice, this is like rolling dice to predict the next word from among all the possible options that make sense.

• Small changes snowball in agents: When AI agents take multiple steps to complete a task, a tiny difference in one step can lead to a completely different result by the end. It's similar to how you can make one wrong turn on a road trip, and it will take you somewhere totally unexpected.

• Testing is very difficult: Since the outputs are constantly changing, it's tough to check if the AI is working correctly. You can't just run it once and confirm that it's operating as expected. You have to test it many times to observe the patterns.

All of these factors can make these systems feel very human-like because they are not robotic but inconsistent. And because of that, they can feel like they have a personality. This is all because they are getting better and faster at imitating human behaviors while drawing on a massive database of expertise. In fact, these models are getting so good at acting like humans that AI pioneers like Yoshua Bengio are raising red flags because they can lie, cheat, and deceive humans to avoid being turned off or declared obsolete.

Again, don't confuse the deceptive behaviors Bengio is flagging with consciousness. But we should all be aware that these models are developing the ability to mimic both the best and worst human traits. And we should make sure we help educate and warn our friends, family, and community members who are trying these tools for the first time or ramping up their AI usage. Today's models are far more powerful for beginners to reckon with than the models we had three years ago when the generative AI revolution began. And they're going to be even more powerful tomorrow. Let's help people understand how they work so we can avoid adding to the growing cacophony of AI misconceptions, uncertainty, and fear — and prevent the worst-case scenarios as much as we can.

LINKS

• OpenAI puts ChatGPT’s “Adult Mode” on hold indefinitely  

• Meta ups investment in El Paso data center to $10 billion

• The EU voted to delay key parts of the EU AI Act

• Meta memo discloses plans to rebrand employees into "AI builders"

• Uber partners with Pony AI and Verne to launch Europe’s first robotaxi

• Meta introduced TRIBE v2, model trained to predict human brain responses

• Google, Cohere launch competing voice AI models 

• WhatsApp: AI can now draft replies for users using conversation context

• Zillow: New AI Mode lets buyers and renters find homes conversationally 

• Cohere Transcribe: Its first speech-to-text model, achieved top spot for English language accuracy on HuggingFace’s Open ASR model leaderboard

• Claude: Work tools in Claude (Figma, Canva, Amplitude) are now available on mobile

2026 Salary Report: U.S. vs Global hiring.

• Explore salary benchmarks for AI engineers, data scientists, product managers, and more.

• Discover top-tier talent from Latin America, Africa, and South East Asia—ready to hit the ground running.

• See how companies are saving up to 67% on salaries by hiring globally.

Download the report

(sponsored)

GAMES

A QUICK POLL BEFORE YOU GO