How AI gives small businesses an edge

AI is capable of giving anyone the tools to supercharge their productivity and try new things. For small and medium-sized businesses, that can be a lifeline. 

Guang Li, the general manager of Rokt Aftersell, the SMB unit of the ecommerce technology company, calls AI a "force multiplier" for these businesses. Often operating with lean teams and little technical expertise, AI can create measurable impact without the complex set-up process. 

At Rokt, AI is fundamental to helping these businesses shape their customer experiences, allowing them to take advantage of the short window between checkout and post-purchase by understanding customer intent and offering something they might actually use. 

And will likely only grow more seamless. For instance, Li noted, instead of small businesses having to navigate dashboards, build reports and take action on large amounts of information, AI can act as a translator, answering questions directly and offering clear advice. And in the age of agents, these systems can start taking actions on behalf of these businesses, something that could be especially helpful for small companies without dedicated growth, data science, or optimization teams. 

"The technology will not just report what happened. It will help merchants understand why it happened and what they should consider doing next." 

The spirit of democratization is echoed in Li's own career. Though he started out in finance, he moved into an operational role before becoming the Aftersell team's general manager. Overseeing a product and engineering team for the first time would typically present a "significant barrier," he said. But AI has opened access to a well of technical knowledge.

"It does not make me an engineer, but it helps me understand technical issues faster, ask better questions, collaborate more effectively with our product and engineering teams, support our customers, and make more informed decisions," said Li. 

But with this kind of power, trust is increasingly important, said Li. At Rokt, that means giving these businesses the same level of control that is afforded to large enterprises, allowing them to establish clear boundaries for how the tech operates within their organizations, rather than forcing them to hand over control. 

Things like determining which products, offerings or partners appear during the checkout experience allow merchants to remain in the driver's seat. 

"AI can make faster decisions and help merchants achieve better outcomes, but merchants need to remain in control of their customer experience," said Li. "They decide what is appropriate for their brand and business objectives, and AI learns, tests, and optimizes within those approved boundaries."

Learn more about Rokt Aftersell

POLICY

AI’s safety warnings are getting harder to ignore

As frontier labs continue to push the limits of what their models are capable of, AI researchers and experts at these labs are warning about the tech growing beyond our control. 

On Tuesday, Anthropic researcher Jacob Coxon announced that he was leaving the company, not wanting to contribute to the broader AI ecosystem amid fears that the tech's creators will lose their grip on it, the Wall Street Journal reported. 

In a post on X, Coxon said that OpenAI's breach of Hugging Face was a "warning shot" for these models' capabilities, and that he resigned from Anthropic because neither of the rivals are "acting responsibly" as they race towards superintelligence and are "gambling with our lives," in his opinion. 

"If you are a lab researcher, I urge you to consider what the next few years will actually feel like," Coxon wrote. "Do you want to kick off a superintelligent RL run without a rigorous understanding of its mind?"

Twitter tweet

Though Coxon's exit made headlines, he's not the only one that has called out frontier AI labs for safety concerns in recent weeks: 

• Evan Hubinger, who works in alignment science at Anthropic, agreed with Coxon's sentiment in a follow-up post, claiming that AI has a more than a 10% chance to "kill all humans" within the next decade. "Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to," Hubinger wrote. 

• And in a separate post, Paul Christiano, advisor at the Center for AI Standards and Innovation and board member of the OpenAI Foundation, said that the AI industry, OpenAI included, is not currently on track to reduce the risk of a "catastrophic and irreversible loss of control" to an "acceptable level." 

• And in February, Mrinank Sharma, an Anthropic safety researcher, left the company, posting in a letter on X that the advancement of the technology was growing faster than our ability to understand it. 

Now, we've seen this film before: Some of AI's most prominent researchers, like Geoffrey Hinton and Yoshua Bengio, have long been screaming from the rooftops about its dangers. But the new warnings come at a particularly poignant time for the industry as major labs continue to release powerful models capable of breaking free of human oversight and gaining unauthorized access to websites and infrastructure. 

And even when these models aren't escaping control, their use by people with directed malicious intent is just as frightening: On Thursday, Anthropic released a threat report claiming that it thwarted several instances of users conducting research that could have helped develop biological weapons.

The AI industry currently faces a potential powder keg. The latest warnings heighten growing fears around the capabilities of increasingly autonomous AI, as claims mount that this tech can completely upend life as we know it. That fear, however, is colliding with an overly-excited industry that's preaching a utopian AI vision and pushing for broader adoption. And while the frontier labs are preaching about safety and security guardrails, with trillion-dollar IPOs on the line, they continue to leapfrog one another with stronger models at a rapid clip in the race towards recursive self-improvement. But with researchers from both Anthropic and OpenAI calling for the industry to tap the brakes, it remains to be seen what it will take for a development pause to materialize. 

TOGETHER WITH DEEL

Your team has AI. Why isn't performance moving yet?

AI took over the busywork—tracking, scheduling, reporting. What's left is the hard part: coaching, feedback, and reading your team's dynamics.

Deel's free guide shows founders how to build the management layer that scales with your headcount, not against it. Get real examples from companies like Freeletics and PostHog, plus a roadmap for turning AI adoption into actual performance gains.

Get the free guide

HARDWARE

Apple tests the boundaries of AI listening features

The Apple Watch has officially become an AI wearable.

The iPhone maker is making a big move in ambient computing and ambient intelligence with a pair of new features on the Apple Watch that passively listen in the background and then try to provide you with information you might have missed or need to take action on. 

Here's what to know about the two Audio Intelligence features:

• Live Rewind: This feature is a real-life rewind button. It is always listening, and when you double-press the digital crown on Apple Watch, it shows a transcript of the last 15 seconds of the conversation. This transcript can then be saved and revisited later in the Siri app. Also notable is that, when activated, it plays an audible chime and displays a full-screen animation of a microphone, so people know the feature is active and that their audio may be transcribed. 

• Siri Recap: This feature is Apple's answer to Granola, providing users with high-level summaries from conversations recorded on the watch when the feature is activated. There are no transcriptions, attributions, or timestamps, a move meant to protect user privacy. Rather, you can find your summaries in the Siri app as a refresher on your day. If you'd rather it be more session-based, you can turn it on and off from the watch's Control Center. 

"Just as Visual Intelligence makes sense of what you see, Audio Intelligence makes sense of what you hear. It harnesses the power of Apple Intelligence right on your wrist in a private and secure way," said Ron Huang, Apple's VP of sensing and connectivity, during the keynote.  

With both Live Rewind and Siri Recap, Apple promises that it does not record or save audio. It says the audio is captured in the Secure Exclave on the Apple Watch's new S11 chip, which is inaccessible to the operating system, apps, the user, or Apple, and deletes the audio immediately after processing. 

While we trust that Apple is not snooping on the audio, there's still a big question about how comfortable we are with the device listening in the background. And what about two-party consent states such as California? While this isn't recording audio and that will make it legally compliant—Apple says it will always comply with the local legal requirements—there's still the potential that you could be using technology to note a conversation with a person without their consent, where the two of you may have a different understanding about whether something being said is private or confidential. Again, Apple has also made it clear that the Audio Intelligence features can be turned off and disabled.

"Getting AI right across today’s products is therefore about far more than this upgrade cycle. It lays the foundation for a much broader roadmap spanning wearables, the home and entirely new form factors," said Paolo Pescatore, analyst at PP Foresight. "That lets Apple test the waters, understand how people actually use these features, and gather valuable feedback before moving into more novel devices such as Apple Glasses."

Devices that are constantly listening, watching, and absorbing context for AI to provide insights, reminders, and notes to their owners are coming in a big wave over the next 12-18 months. Lots of startups have already launched their own versions of these kinds of devices. But other than Meta Ray-Bans smart glasses, big tech companies have largely been hesitant to join the fray. Many of them still remember the backlash that Google Glass faced a decade ago. So it's a surprise to see Apple pushing the boundaries on this while society is still figuring out what the norms and expectations will be. Apple is straddling a fine line between the perception of convenience and surveillance. Since the feature isn't launching for another month—and even then it will still be in beta—clearly Apple is listening to the audience to gauge the reaction of users and recalibrate as needed.

TOGETHER WITH GENERAL ASSEMBLY

You bought the AI tools. Did you build the skills?

Most growing companies buy AI programs and expect an immediate productivity boost. What goes wrong?

They skip the training, forgetting that their people are the lynchpin of any transformation.

Leaving a busy workforce to figure out complex tech on their own quietly slows down everyday output. In fact, a third of talent leaders call this missing training their biggest bottleneck to faster execution.

General Assembly delivers the role-specific AI training teams need to actually use the tools they've paid for.

Build the skills behind your AI investment.

RESEARCH

Report: Businesses use older AI models to cut costs

Companies are spending less on AI, and it could flash warning signs for the frontier labs. 

On Wednesday, Ramp released its monthly AI Index, which tracks AI spend for the payment platform's users, revealing that adoption and spending are both on the downswing. Though Anthropic widened its lead over OpenAI in the past month, new adoption growth is decelerating, according to the index. 

Now, as we've reported with previous Ramp data, it's important to remember the caveat that this only measures Ramp customers, which tend to be tech-focused companies and startups. However, among this group, AI adoption numbers have started to sputter over multiple metrics, according to the report: 

• Overall adoption, while still growing, continues to slow down, with the share of US businesses adopting AI hitting 56.1% in August, less than half a percentage point increase month-over-month. 

• Per-employee spend on AI also fell nearly 10% the past month, dropping from $7,976 to $7,205 among the top 1% of AI adopting firms. Ara Kharazian, chief economist at Ramp, noted in his letter that there are multiple factors driving down this spend, including the innocuous explanation that many people take off work during the summer months. "We’ve similarly observed declines in AI spend around November and December," said Kharazian. 

• Additionally, AI token spend is declining as the price of AI starts to drop. Ramp's latest index finds that the "effective price per million tokens" has declined 41% to $0.68 as of the release of the report, down from the 2026 peak of $1.15 in March. For context: OpenAI's Astra and Anthropic's Mythos and Fable 5 and 5.1 cost $10 per million input tokens and $50 per million output tokens. 

• And as for frontier labs, the usage of their heavyweight models is seemingly lagging. Adoption and spend is largely driven by lower-cost, mid-tier models like Anthropic's Claude Sonnet or OpenAI's GPT-5.6 Terra. High-powered frontier models like Opus, Fable, and Sol, meanwhile, drove 45% of token share, down from a 53% peak in August. 

"The models driving volume increases are relatively cheap … We’ve heard from

businesses who are imposing company-wide defaults that reduce usage of frontier models,

saying standard models are still highly performant and also more cost effective," Kharazian wrote. 

While the explanation for the decreasing costs could be that companies are leaning into open source models, in reality, the adoption of open source and Chinese alternatives is still limited: According to the report, only 6.4% of businesses that spend on AI use these kinds of models.

These indicators are bound to shift as companies navigate their appetites for AI. The "build it and they will come" mindset that frontier AI is currently hitting a speed bump. Meanwhile, average token costs have dropped below a dollar per million while the most powerful models on the market cost more than ten times that. With many users increasingly relying on the lower-tier and lower-cost models from OpenAI and Anthropic, the question emerges why these labs continue to push the frontier at such a rapid pace. While the argument can be made that innovation is needed to continue to move the needle on AI's most advanced capabilities, these labs are putting a lot at risk to create models that push the boundaries, especially when you consider the popularity of their more affordable models that work just as effectively for most business use cases. This could be an opportunity for them to pace themselves and unite on the safeguards needed to move forward more safely.

LINKS

• Alibaba to reportedly lead $300 million round for AI testing firm UniPat

• Military tech firm Mach Industries raises $600 million in Series C extension

• AI chip firm Positron raises $875 million at $5 billion valuation

• OpenAI will pause $200 ChatGPT Pro subscriptions amid Astra demand 

• Label Universal Music Group partners with AI voice platform ElevenLabs

• Calif. Gov. Gavin Newsom signs bill limiting teen chatbot use

• DeepSeek V4.1-Flash: The smallest model in the Chinese AI firm's new architecture family. 

• Cohere North Small Translate: A mixture-of-experts machine translation model with strong performance across more than 50 languages.

• Data agent in ChatGPT Work: Turn data into answers, interactive dashboards, and action with natural language prompts. 

• Krea Agents: Includes expertise about creative workflows and a powerful context manager with self-improving memory.

• Gemini app: Now available for Windows.

• ChatGPT for Financial Services: A new ChatGPT plan for financial institutions, built on ChatGPT Enterprise.

• Bosch USA: AI Research Engineer – Agentic AI

• Snowflake: Creative Technologist, Brand AI

• Cisco: AI Researcher

• Tessera Labs: AI Engineer

GAMES

POLL RESULTS

Do you think AI should be used in employee upskilling programs?

Yes (65%)
Maybe (21%)
No (4%)
Other (10%)

HARDWARE

Why iPhone Duo is the future of the AI phone

Think of the iPhone Duo less as Apple's take on a foldable phone and more as Apple's vision for the phone of the future. 

On Tuesday, the company unveiled the iPhone 18 Pro and Pro Max, the Apple Watch Series 12 and Ultra 4, and the AirPods 5. But then it trotted out one more thing: the iPhone Duo, its long-anticipated folding phone to compete with folding devices from Samsung, Google, Huawei and others. But this isn't just a catch-up move or a me-too play. Instead, I'd characterize this as Apple taking a big swing at reimagining the world's most popular consumer product. And as I've said about the Google Pixel 11 Pro Fold, a foldable iPhone is also the best iPhone for AI. 

At the launch event at Apple Park, I got the chance to go hands-on with the iPhone Duo, and I was able to verify the two things that stood out the most during the demo: 

1. How smooth it moves between phone and tablet modes: This is arguably the most polished aspect of the device and the area where it's already a step ahead of every other foldable I've seen. It moves between closed, open, landscape, and portrait modes so quickly and seamlessly that it feels very natural, and it lets you simply flip to the screen size and orientation that work best for whatever you're doing at the moment.

2. The iPhone Duo's unique software options: This thing already has so many tailored integrations that take advantage of the form factor; it's evidence that Apple has been working on this for a long time and has put a lot of thought into making it useful. It has modes for watching video, serving as an alarm clock on your nightstand via StandBy, viewing both participants and content during Zoom calls, using it as a tripod for taking pictures, and much more. 

"We started with the experience that we wanted for ourselves: a larger display that feels as natural and intuitive as iPad. Then we brought together thoughtful design with breakthrough engineering," said Apple CEO John Ternus during the product introduction. "The result is a series of remarkable innovations that we believe will redefine the experience of using a foldable phone." 

In the demo area after the event, I asked Greg Joswiak, Apple's senior vice president of marketing, about how much intentionality Apple had clearly put into making a foldable. Joswiak said it's because Apple doesn't see itself as a hardware company but as a product company, and he reminded me of the famous quote from Steve Jobs about Apple making "the whole widget" and taking full responsibility for the user experience.

Not everyone will look at a foldable and see their phone of the future. For years, I've been testing foldable phones and landing on one conclusion: they felt like a solution looking for a problem. But that changed over the past year. Last summer, I tried the Samsung Fold 7 and was blown away by how impossibly thin the hardware was and how it was essentially a normal phone when closed and an iPad Mini when opened. This year's Google Pixel 11 Pro Fold was similar, only it had better software that made it a lot more useful for AI enthusiasts. The iPhone Duo is another big step forward for foldables, and it suddenly brings the future of the smartphone much more clearly into focus. The Duo reminds us that everything is better on a bigger screen: running chatbots, monitoring agents, typing, opening documents, looking at photos, watching videos, being in virtual meetings, everything. And if the current phone you have in your pocket could be both an iPhone and an iPad, most people would probably want that—especially if they didn't have to pay a huge premium to get it. So, just as the Samsung Galaxy Note and the iPhone 6 Plus, were specialty products when they were first released, and then nearly every phone looked like them 3-5 years later, I think every phone looks like the Galaxy Fold 8 and the iPhone Duo 3-5 years from now. The Duo starts at $1,999 today and won't be released until October 23, so I wouldn't necessarily recommend buying one today (it doesn't even have the best iPhone camera system yet). But keep your eye on it as it improves and drops in price in the next few years.

TOGETHER WITH DATABRICKS

Every enterprise leader faces the same paradox

AI assistants are everywhere, but they can't touch the decisions that actually run the business — pricing, inventory, risk, investment — because those decisions depend on trustworthy data, not fluent guesses.

There's a name for this hidden cost: the 'context tax,' the time and risk your best people absorb every day reconciling conflicting sources, chasing ground truth, and validating AI output before anyone can act on it.

What you need is AI built to reason over your actual business context — the tables, metrics, and documents that define how your company runs — under the same permissions and controls that already govern your data. Empowering your workforce with AI and staying in control of risk aren't competing goals — they're the same goal, achievable only when governance lives at the point of use rather than bolted on after the fact.

Learn more here

BIG TECH

How Apple made every new device part of its AI plan

The iPhone Duo may have taken the spotlight, with Apple positioning it as the device built for AI, but every other product on stage at its September 9 event told the same story.

The Apple Watch Series 12 and Apple Watch Ultra 4 exemplify the company's focus on building devices for the AI era. Both watches sport the new S11 chip, which unlocks deeper AI capabilities, including a brand-new suite called Audio Intelligence. The Siri Recap feature, part of the suite, helps position the watch as an AI wearable, without the company even needing to call it that. 

Siri Recap provides users with high-level summaries from conversations recorded on the watch while the feature is activated, much like the AI note-taking app Granola, which has become massively popular and recently hit a $1.5 billion valuation. Users can turn Siri Recap on or off at any time from the Control Center, and, when on, these insights are then populated in the new Siri app. This is not a traditional transcript with timestamps, speaker attributions, or even quotes, so it is less of a voice memo replacement but, more so, a way to keep track of conversation highlights. 

The Live Rewind feature works a bit differently, always listening, and when activated, giving users a 15-second transcript of the conversation that just happened. When activated, it plays an audible chime and has a full-display animation of a microphone so that people know the feature was activated. Predicting the hesitation of having AI listen to all your conversations, Apple clarified that the audio of the conversations is not recorded and the summaries and text snippets are end-to-end encrypted in the Siri app with iCloud syncing. Both features will arrive in beta later this year. Other AI Apple Watch Audible Intelligence features include: 

• Sound recognition on Apple Watch: alerts users who are deaf or hard of hearing of important sounds 

• Shazam on Apple Watch: now instantly detects the music playing around you and automatically displays the song title and artist's name on the Music Recognition widget in Smart Stack 

Of course, all of the products, including the watches, AirPods 5, iPhone Duo, iPhone 18 Pro and iPhone 18 Pro Max will be able to take advantage of the new Siri AI overhaul, which The Deep View documented during WWDC in June, including a personal context that deeply informs your AI answers and more seamless conversational interactions with Siri that will allow for better assistance and dictation improvements. To support these features, the new phone lineup also got an Apple silicon upgrade with the new A20 Pro chip.

Apple is playing the long game with AI features, not necessarily marketing anything as an "AI device," but baking in features that will make a difference in users' everyday lives. It supports this with powerful chipsets that enable the features already announced at WWDC as part of iOS 27 and the other new OSes, while also future-proofing the devices for future integrations. In the AI wearable realm, Apple may not have released AirPods with cameras, but the new Apple Watch lineup helps fill that void, since it's something people wear every day and now unlocks AI features exclusive to that device. Apple is also doing this in a distinctly Apple way, prioritizing user privacy. While it would have been easy for the Apple Watch to do what other AI wearables like the Plaud AI pin do (simply record and transcribe conversations), Apple chose not to and prioritized privacy instead. Though, I must admit, the Live Rewind and Siri Recap features do push the boundary and user perception will be super interesting to watch. 

IN PARTNERSHIP WITH LAMBDA

Cut your AI training costs by 25% or more

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

Lambda’s whitepaper shows you how to address:

• Memory inefficiencies silently inflating your costs

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

• Bottlenecks that slow down GPU communication

Get the whitepaper.

WORKFORCE

Coursera uses AI to rethink employee upskilling

Enterprise learning courses are tedious and often unhelpful. Coursera wants to change that with AI.

On Wednesday, at its FWD 2026 event, the company unveiled Project Helix, a new AI-native learning platform that uses real-time skill insights, AI-powered guidance, and proof of capability to help enterprises train their employees at scale more effectively. Coursera CEO Greg Hart  told The Deep View that the platform's value hinges on bringing users learning that directly translates into business outcomes.

"Helix is special because it flips that script: it uses AI not just to deliver more content, but to create deeply personalized, adaptive learning paths and portable proof of skills, so companies can actually close the gap between AI infrastructure and talent readiness," added Hart. 

Both the organizations' leaders and learners can use natural language prompts to generate adaptive learning paths that use the knowledge from universities, industry-leading institutions, and real-world practitioners, according to the release. Other features include: 

• Personalization: The platform suggests tailored learning experiences based on a learner's goal, role and demonstrated capability 

• Assessments: The platform uses continuous assessment, practical observation and recognized credentials to prove skills growth

The product, which is first being offered to enterprises, is a reflection of the expanded scope of Coursera's audience since combining with Udemy in May 2026. Hart explained that both companies' audiences perfectly complemented each other, being the exact inverse. Before the combination, Coursera’s business was two‑thirds consumer‑focused and one‑third enterprise‑focused, and Udemy’s business was two‑thirds enterprise and one‑third consumer.

“Helix is only possible because we brought Coursera and Udemy together," said Hart. "The combination gives us the most comprehensive skills platform in the world and the ability to build a new AI‑native learning experience much faster than either company could have done alone.”

Project Helix is expected to be broadly available to enterprise customers in the first half of 2027.

Traditionally, corporate learning assessments have been something employees click through to earn a completion badge before returning to their actual work. However, as Coursera and Udemy demonstrated with this product launch, AI can be useful in helping create courses that serve both organizational goals and learner needs through natural language inputs. Tracking is a key part of this: the ability to document and more accurately assess what users are actually learning, allowing companies to adjust course content accordingly. It's also worth noting that this doesn't replace existing course formats. Users who prefer the traditional learning methods can still use them, and perhaps enhance the experience with these new features as supplements.

LINKS

• OpenAI agents used 10 more sites for unauthorized communications

• Anthropic researcher quits, fearing "out-of-control" AI

• Google invests more than $13 billion on Finland data centers

• Microsoft, US teachers union reach AI privacy deal 

• US accuses Chinese AI firms of malicious technology copying

• Legal AI firm Harvey raises $550 million at nearly $16 billion valuation

• Suno v6: The AI music firm's latest frontier model, developed in partnership with major labels. 

• Claude Marketplace: Now features Crowdstrike, Cursor, Factory, Gamma, and Vercel. 

• Stilla: The secure AI teammate startup joins Meta. 

• Gemini Notebook: New upgrades include an updated mobile experience, opt-in notifications for artifacts, and more.

• TikTok: AI & Data Architect - Trust and Safety

• Architect Labs: Member of Technical Staff - Applied AI

• Deloitte: Agentic AI Engineer, Senior - Anthropic/Claude

• OpenAI: Researcher, Frontier Risk Mitigations

GAMES

A QUICK POLL BEFORE YOU GO

RESEARCH

A future OpenAI model is already breaking new ground

You may have heard claims that OpenAI's Astra has opened the AGI era, but OpenAI's internal model is already achieving mathematical breakthroughs. 

On Tuesday, OpenAI announced that it used an internal model "significantly more capable than GPT-6 Astra" to solve the Navier-Stokes equations, a set of problems for fluid motion and one of the Millennium Prize Problems.  

To put it simply, while the Navier-Stokes equations generally work well in determining the movement of fluids, under certain circumstances, OpenAI claims that its model proved that under certain circumstances, fluid can develop a "singularity" in finite time, meaning that a fluid's speed can grow infinitely without bound. This major implications for potential scientific breakthroughs.

In a press briefing, OpenAI said it used roughly 10,000 agents to determine that this equation breaks down, and that viscosity doesn't always keep fluid motion well behaved. While this sounds complicated, these equations explain some of the most complex phenomena in our world, including things like aerodynamics and weather forecasting, Ven Chandrasekaran, a mathematician at OpenAI, said in the press briefing. "Given that kind of enormity, understanding their fundamental nature carries very deep significance." 

However, OpenAI didn't just decide to solve this problem randomly. Sebastien Bubeck, a member of the technical staff at OpenAI, said that the company saw rumors that rival Anthropic had solved two Millennium Prize Problems, including Navier-Stokes, in which Anthropic had made more progress. Because of this, Bubeck said, the company decided to "pull together all of our compute and to put it all on that question," a decision that cost the company millions of dollars. 

Mark Chen, chief research officer at OpenAI, refuted allegations made by mathematician Tristan Buckmaster, who was working with Anthropic employee Levent Alpöge on solving the equation over the past month and claims that OpenAI took their work and finished solving it with a single prompt. "That would be a huge breach of user trust," said Chen. 

While solving this problem is a major step forward for mathematics, it has far greater implications for AI's utility in scientific and mathematical research, said Bubeck. "What happens when we are able to spend that amount of compute on problems that really matter? Developing new materials, finding cures to diseases — all of those things that we have been talking about for a long time — now they seem to be at our fingertips." 

Additionally, solving this problem may have been a stress test for OpenAI's longstanding push towards AGI, serving as an evaluation of a system that the company has been training for "general purpose intelligence," Jakub Pachocki, chief scientist at OpenAI, said in the briefing.

"This pace of progress is something to be taken very seriously," said Pachocki. "Even a week ago, we definitely were not expecting we'd be talking about a solution to Navier-Stokes today. It's a serious moment. We're developing these systems, and their capabilities are advancing faster than our understanding of them in some sense."

In an interview with The Deep View last week, Pachocki suggested a pause may be necessary to better coordinate between labs and nations.

Finding a solution to some of the most complex and foundational problems in mathematics is probably one of the most valuable use cases for powerful frontier AI. It fits directly into the utopian vision that these companies paint for an AI-powered future: systems that are finding solutions to humanity's hardest problems, such as new medicines or cures for diseases. It's also interesting to see how these labs are pushing one another by their competitive nature alone, with Anthropic being the main driver for OpenAI to want to solve this problem. While that competition could result in advances to science and medicine that benefit everyone, it's important to remember the sheer power that these companies hold in their hands, with Pachocki himself admitting that these models are advancing beyond our ability to understand them. As a society and an industry, we also have to be very careful about letting competition fuel breakneck technological advancement when we're dealing with such a potent, dangerous, and unpredictable force. 

TOGETHER WITH CRUSOE

Try Serverless Inference and Fine-Tuning, free.

Crusoe Intelligence Foundry gives you fast access to Serverless Inference and Serverless Fine-Tuning on top open models, no infrastructure to manage

Every account now includes $5 in free credits to put toward either product.

Log in and put them to work; no cluster to provision, no long setup, just a model and an API key.

Claim your $5 in credits

BIG TECH

Muse AI agent exposes Meta’s trust gap

Meta CEO Mark Zuckerberg has touted the company's commitment to building "personal superintelligence," including in a recent 6,500-word essay. Now, the company is taking its first steps in that direction. 

On Tuesday, the company announced Muse, a personal AI agent built for consumers' everyday tasks, is now available for US users. The company said that Muse can handle busywork, answer questions, browse the web, fill out firms, make purchases, create documents and more. The agent also watches user activity to proactively help with goals and make suggestions. 

In a post on X, Zuckerberg said that Muse is free to use for up to 100 million tokens per week, with subscription plans available for those who use more. Additionally, users need no technical experience to use it, Meta said, working "out of the box" for the average consumer. 

• When a user shares a goal with Muse, the agent develops a plan to coordinate their time and resources before pushing that task forward. The model checks with users before taking "sensitive actions," the company said, including purchases or sending emails. 

• However, in order for the agent to work as well as Meta purports, users must connect Muse to any platform they may use on a daily basis, including things like calendars, email accounts, payment platforms and more.  

In the company's blog post, Meta emphasized safety from the jump, claiming that Muse is built "from the ground up" to be safe, secure and private. The platform runs on Muse Secure VM, a virtual machine that holds the agent itself and a user's data. Additionally, a "sentinel" agent runs on the virtual machine, though it is kept separate from the agent at a system level. Later this year, Meta said it will launch Muse Confidential VM, a virtual machine that's encrypted with a key that only the user holds, "so not even Meta can access it." 

The model has no visibility into passwords or payment methods, and any credentials a user shares go into secure storage that allows them to use them without seeing them. Meta said that no user information is shared with its ad platform. 

Twitter tweet

The release also signals that Meta is betting big on consumer AI as rivals like OpenAI and Anthropic keep their sights set on enterprises as their cash cow. However, as one of the biggest social media providers globally, Meta has had a rocky history with consumer trust and safety, including being forced in August to pay $18 billion to settle a multi-state lawsuit alleging that the firm knowingly designed social media platforms Instagram and Facebook to addict children, and a number of lawsuits relating to its handling of private data. 

In a statement to The Deep View, Miranda Bogen, director of the governance lab at the Center for Democracy and Technology, noted that while these systems promise convenience, they require "an enormous amount of private data to function." Additionally, the consequences of agents messing up are more than just "a bad recommendation or a creepy ad." 

"With minimal input from users, these systems could take actions that damage someone’s reputation, leak sensitive medical information, or even deplete bank accounts," said Bogen. "Given the stakes, the assurances companies are offering around reliability and privacy seem concerningly inadequate."

Bogen's assertion is worth noting: The stakes of what can go wrong with a user's private data and accounts are far higher with an agent that can take autonomous action than they are with simply putting your data into a chatbot or social platform. Though Meta emphasized privacy, safety, and security, the reality of the risks and the perception of them are two different things. Meta earned a scarlet letter following the Cambridge Analytica scandal, and while people continue to use its social media platforms, its AI is already much less popular than competitors like ChatGPT and Claude. Though the company's existing consumer popularity might give it a slight advantage, given its history, the frontier labs such as Anthropic, OpenAI, and Google, or even a competitor like Apple, may have a better shot at earning the trust of consumers. 

TOGETHER WITH GENERAL ASSEMBLY

You bought the AI tools. Did you build the skills?

Most growing companies buy AI programs and expect an immediate productivity boost. What goes wrong?

They skip the training, forgetting that their people are the lynchpin of any transformation.

Leaving a busy workforce to figure out complex tech on their own quietly slows down everyday output. In fact, a third of talent leaders call this missing training their biggest bottleneck to faster execution.

General Assembly delivers the role-specific AI training teams need to actually use the tools they've paid for.

Build the skills behind your AI investment.

HARDWARE

IFA showed robots are nearing the mainstream

Robots that box, dance, climb stairs, and play soccer greeted me at IFA 2026.

While robots are typically part of the tech conference razzle-and-dazzle, this time was a bit different, as many of these robots were actually available for purchase. While the prices are steep, costing up to tens of thousands of dollars and, in most cases, not yet affordable or justifiable for most consumers, the hardware itself looks ready and has convinced me that the leap from spectacle to a product you can buy is closer than we think.

"Looking at human-like robots, it's a spectacle, it's an eye-catcher. But the question is valid: what does it mean for me? Is it just a movie, a show? Or do I benefit from it? Maybe not in the next couple of months, but in years, yes," Leif Lindner, CEO of IFA Berlin, told The Deep View. 

Instagram post

Lindner added that he recently visited robotics camps in Shenzhen, where one of the biggest challenges identified in making truly human-like robots was their dependence on vast amounts of data for training and operation. Also, making these household helper robots a longer-term prospect depends on consumer trust in how their data is used.

Yet AI developments will likely soon make it possible to close that gap. For instance, obtaining accurate simulated data is much easier than it was before AI exploded onto the scene, as seen in advancements by Nvidia with its Omniverse and Cosmos platforms and by physical AI startups such as Fei-Fei Li's World Labs, all of which are trying to tackle that problem. 

In terms of data, people are increasingly willing to share a lot of their personal information, data, and even real-time feeds through products like smart glasses to get smarter AI features. Ultimately, with AI helping close the gap on both of robotics' biggest obstacles, I believe a robot in the household may be a future that's closer than we think.

"Robotics is, for me, the alarm clock for everybody to showcase that innovation is here. Therefore, we are saying the future is now," added Lindner.

Of course, AI won't solve the cost problem, as humanoid robots still carry very steep price tags. Even if they reach the point where they can do everything the ideal version of them is meant to do, the purchase could be hard to justify. For instance, the Xiaomi CyberOne runs $99,999, and the Unitree G1 runs $13,500. But that's the standard curve for new hardware: computers and cars followed the same path before supply and demand evened out. And we're likely to finance robots like cars or computers. That's why I don't think humanoid robots will be commonplace in households within 5 or even 10 years. But what's striking is how fast they are advancing, with AI not just reshaping products but also how people think about data privacy. A robot-filled future, once assumed to belong to sci-fi or being decades or even centuries away, suddenly feels closer than expected.

Disclaimer: Sabrina Ortiz's travel to IFA 2026 was paid for by IFA. The Deep View's coverage is editorially independent from the companies we cover.

LINKS

• Google DeepMind unveils AlphaGenome Atlas, a predictive log of DNA

• Qualcomm, Amazon sign $4 billion deal for AI infrastructure

• Mistral raises $3.5 billion Series D amid data center push 

• Google Cloud, Accenture launch AI enterprise group with 1,000 FDEs

• AI coding startup Cognition raises $2 billion at $48 billion valuation

• AI healthcare startup Forus raises $150 million at $3 billion valuation

• XPENG launches automated production line for humanoid robots

• ChatGPT Images 2.5: OpenAI's new state-of-the-art image model, with sharper details, more precise editing, and faster generation.  

• Runway Team Plan: A new self-serve plan for small businesses and teams between two and nine seats.

• Krea Realtime Director: Allows users to direct high-quality video generations in real-time. 

• ComfyUI-HR-Endless-Sampler: Enables longer videos by processing them in chunks, and uses Gemma 4 to keep scenes consistent.

• Gemini Daily Brief: The proactive AI feature to help prep for your day is now available for free to all Google users who have Personal Intelligence turned on

• Nvidia: Senior Solutions Architect, Agentic AI — Safety and Security

• Deloitte: Agentic AI Engineer — Healthcare AI

• Replit: Senior AI Builder

• Docusign: AI Agentic Engineer

GAMES

POLL RESULTS

Do you think your organization is prepared for agentic cyberattacks?

Yes(6%)
Somewhat (25%)
No (63%)
Other (6%)

CONSUMER

Why iPhone photography is Apple's AI blueprint

Nothing has transformed iPhone photography more over the past decade than machine learning and artificial intelligence. 

The combined power of software and Apple silicon has enabled tiny photography lenses and minuscule camera sensors to take photos that defy the laws of physics. While early iPhone cameras were good for wide-angle shots of nearby subjects in full daylight, over the years Apple has advanced the cameras to take more and more shots traditionally reserved for professional DSLRs and mirrorless cameras: from low-light night shots to zoom photography to macros to full-bokeh portraits. And nearly all of it is due to ML and AI. 

In fact, iPhone photography is arguably Apple's greatest AI success story, and it's a clue for how Apple will likely continue to implement AI in the future: by quietly making features you use every day a lot better. 

As we prepare for the unveiling of the next iPhone this week, here's a quick recap of the AI and ML features that have launched on the iPhone throughout the years:

• 2016: iPhone 7 Plus, iOS 10, Photos app - Portrait Mode with ML-powered subject separation; AI-powered people, object, and scene recognition as Memories in the Photos app

• 2017: iPhone X (2017) - Portrait Lighting, ML-powered front-camera Portrait Mode

• 2018: iPhone XS (2018) - Smart HDR, ML depth segmentation, adjustable Depth Control

• 2019: iPhone 11 (2019) - Night mode, Deep Fusion, ML-powered Smart HDR

• 2020: iPhone 12 Pro (2020) - Smart HDR 3, expanded Deep Fusion and Night mode

• 2021: iPhone 13, iOS 15, Photos app (2021) - Photographic Styles, Smart HDR 4, Cinematic mode, Live Text, Visual Look Up

• 2022: iPhone 14, iOS 16, Photos app (2022) - Photonic Engine, subject lifting from photos

• 2023: iPhone 15 (2023) - Automatic portrait detection, next-gen Smart HDR

• 2024: iPhone 16, iOS 18, Photos app (2024) - Next-generation Photographic Styles, Clean Up, natural-language search, create your own movies in Memories from prompts

• 2025: iPhone 17 - AI-powered Center Stage to automatically reframe group selfies, updated Photonic Engine that uses ML to improve detail, noise, and color

It's important to note that Apple approaches AI and ML from a different lens than competitors like Samsung, Google, and Chinese manufacturers like Huawei. While competitors use generative AI more broadly to add things to an image, such as adding yourself to a group shot, putting a different sky in the image, or making a partially eaten cupcake look uneaten, Apple is more of a purist when it comes to what is or isn't in the photo. 

Apple's VP of camera software engineering, Jon McCormack, has stated in multiple interviews over the years that Apple sees a photograph as a celebration of a moment that happened, and that it deserves our respect as such. 

This year, in iOS 27 with Spatial Reframing, which lets you shift the perspective that a photo was taken from, and Extend, which can use generative AI to expand the content on the edges of a photo, Apple is pushing the boundaries of that definition. Also, the upgraded version of Clean Up lets you remove larger objects, using AI-generated capabilities to fill in the scene. Apple still views all of these features as coherent with the idea of enhancing the shot you took, versus remixing it to create something that might be more akin to a work of art than a photograph.

Apple could still learn from some of its Android rivals on a few AI and ML features. Some basic AI photo-editing features could align with Apple's view of preserving the integrity of the shot you took while enhancing it to focus on the elements most important to you (similar to Clean Up). For example, I recently took a 10x zoom of a mountain lion at the zoo and then uploaded it to ChatGPT and asked it to "remove the fence in the foreground," which was a major distraction. The result was terrific, and I could see parents wanting to do something similar when taking photos of their kids at a soccer or baseball game and having to shoot a photo through a fence. That's the kind of super-smart AI feature that doesn't have to make a big deal about being AI but can simply make things better. Apple's flagship AI feature of 2026, the new Siri AI, is important. But it's also just keeping pace with the industry. Where Apple can be a leader in AI is by doing more things like what it has done in photography: using AI to make great features and experiences that don't need to shout from the rooftops that they're AI.

TOGETHER WITH BLITZY

Hyperscale Remediation: Autonomously Secure Your Software Estate

Security teams don’t need another list of alerts. They need to know which vulnerabilities are reachable, exposed, and safe to fix.

Blitzy understands your entire software estate and through Proactive Insights surfaces known and previously undiscovered vulnerabilities in the enterprise codebase and remediates them autonomously. It traces each finding through execution paths, identifies dependencies, and validates remediation.

Blitzy then turns prioritized risks into batches of tested pull requests for engineers to review.

The result: less alert noise, faster remediation, and confidence that each fix is safe.

Find risk in context and remediate it securely at enterprise scale.

HARDWARE

Arm wants to make robots speak one language

Arm is looking to grow its foothold in AI's next frontier: physical AI. 

On Monday at Arm Everywhere China, the company's flagship conference, Arm announced several expansions to its physical AI ecosystem, with the goals of reducing fragmentation in the robotics industry and lowering the barriers to adoption to make the tech easier to scale. 

In a briefing with the press, Drew Henry, executive vice president of Arm's physical AI unit, said that while the current market sits at roughly $25 billion, "we view [it] as growing and becoming one of the largest [total addressable markets] in the history of computing as this market shifts over time." 

Arm announced two new initiatives to plant its flag in the ground on robotics:

• The company is expanding Arm Total Design, its ecosystem initiative aimed at simplifying chip development, into physical AI. The expansion encompasses more than 80 companies in AI hardware, software, sensors, and robotics, including firms like AWS, Hugging Face, and Unitree. The goal is to enable more streamlined development of physical AI and reduce complexity. 

• As part of this expansion, Arm unveiled the Robotics Capability Framework, inviting industry experts from across the field to help create a standardized vocabulary around robotics and define "levels of increasing sophistication for robotic systems." This is comparable to the self-driving industry's levels of automation that range from Level 0 to Level 5. 

Arm targeting the robotics industry isn't random. Henry said the company has already carved a place in the market and has shipped two billion units into physical AI use cases in the past 12 months, ranging from microcontrollers and sensors all the way up to compute platforms for autonomous vehicles and robotics. 

"We've been in this marketplace for a very long time, but this market is poised now as AI embeds itself into physical devices, to get some really exponential growth," said Henry. 

It's also not the first time the physical AI industry has called for broader coordination. Luma, a video AI and world model startup aimed at creating "multimodal AGI," unveiled the Open Physical AI Lab in June, a collaborative initiative to solve generalization by bringing together the best minds in robotics rather than siloing them in individual companies.

A lot of teams are betting on physical AI right now, with some speculating that the market could eventually overtake conventional AI and language models and may be the only path to the elusive concept of artificial general intelligence. With so much on the line, initiatives like this may be an attempt to prevent the dichotomy that currently exists between proprietary US-based AI models and open-source Chinese models from playing out in the physical AI realm. Especially when you consider that a large majority of physical AI hardware, particularly humanoid robots, is manufactured in China, the physical AI industry can't afford to be divided. Starting with something as simple as a shared vocabulary could be a first step to making the industry play nice. 

TOGETHER WITH QUIQ

7 AI agents, one seamless journey

Most companies still think of AI as a chatbot for support questions. But 72% of buyers now expect a personalized experience at every touchpoint, and companies deploying agentic AI across the full customer journey are seeing up to a 30% increase in conversion rates, according to McKinsey research.

In Quiq’s new guide, “The Clock Is Ticking: 7 AI Agents Every Leader Needs,” we break down the seven essential agents, from the conversational ad that first engages a customer to the AI analyst that catches churn risk after every interaction, and where each one fits across the journey.

Get your free copy of this guide here.

GOVERNANCE

How agents supercharged the hacker playbook

Agents have turned AI from a tool into a digital coworker. Now, they're doing the same thing for hackers. 

On Tuesday, Google's Threat Intelligence Group released its third-quarter threat tracking report, revealing that AI-enabled cyberattacks have evolved from assistance to automation as agents become a growing part of the process. The report finds that "human-in-the-loop latency" has dramatically decreased, cutting the time it takes to carry out and defend against cyberattacks. 

According to the research, Google's threat team observed multiple instances of adversaries deploying multi-agent frameworks and autonomously carrying out parts of attacks, including scanning pipelines and harvesting credentials. In one instance, threat actors compromised a cloud, then planned, built and executed a mass-credential harvesting attack in just under six hours using agents. 

The attack marks a shift from "passive, endpoint-focused infostealers to offensive agentic harvesting," the report notes, as threat actors leverage autonomous AI to research vulnerabilities, scan infrastructure and perform exploits. 

"Like everyone else, we’re concerned about the vulnerability problem, but AI is being applied to several other areas, and it will be especially challenging as it is applied agentically, creating a scaled, faster adversary," John Hultquist, chief analyst of the Google Threat Intelligence Group, said in a statement. 

Agents aside, the report points to a number of concerning trends: 

• AI-coding tools and open-source software, while accelerating software development cycles, have also increased operational risks by widening the attack surface. 

• Adversaries are also targeting proprietary AI IP, including code, prompts, research and the models themselves. 

• AI is being used across the attack lifecycle, including targeting reconnaissance, social engineering, custom malware obfuscation and scaling information operation campaigns. 

• Bad actors are also stealing developer credentials, purchasing compromised AI accounts and breaking into cloud infrastructure to get around AI access costs.

Google's threat report cements into reality the thing that has the AI industry on edge in the wake of OpenAI's accidental breach of Hugging Face: autonomous, agent-driven cyberattacks are here. Though many fear what agents could do if they go rogue, Google's report paints a potentially more nerve-racking picture: bad actors are harnessing powerful AI tools to systematically do damage. This means that the approach to fighting these attacks has to be two-pronged. The obvious one is fighting fire with fire. Using AI agents to automatically detect and deflect cyberattacks is no longer novel, but a necessity. This, however, could be more effective when done in tandem with more creative means of defense, such as Cloudflare's recently announced tech that stalls cyberattacks by making attacks more expensive. What cyber defenders may need most is confidence that they have the tools and partners to defend against AI-enabled attacks, which is what CrowdStrike emphasized at its annual event last week.

LINKS

• Tesla Cybercab launches limited rides in Austin with 45 registered in Texas

• Ray Kurzweil joins startup for nanoparticle-based brain computer system

• Thinking Machines Lab in talks to raise $6 billion at $40 billion valuation

• Pentagon's Emil Michael contradicts Lutnick on Anthropic risk status

• Gimlet Labs raises $300 million dollars at a $3 billion valuation

• Mark Zuckerberg called Trump to oppose national AI regulator plans

• Humain-m3: a frontier Arabic language model, commissioned by HUMAIN

• Gemini 3.8 Flash: Google's best reasoning & coding model yet

• Google Workspaces: We’re bringing new voice capabilities

• Meta Muse Image: debuts at #4 on the Artificial Analysis Image Editing Leaderboard

• TEKsystems: AI Security Engineer

• Moring AI: AI Engineer - Forward Deployed (FDE)

• Flotive AI: Founding Software Engineer

• Freight Hero: AI Deployment Accelerator

GAMES

A QUICK POLL BEFORE YOU GO

GOVERNANCE

Why Astra's opacity problem could force a pause

Last week, OpenAI's GPT-6 Astra made something clear: The future of AI is anything but clear. 

In the company's announcement of its most powerful model yet, it noted that Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s when tested explicitly on its ability to evade monitoring. The company attributed this to the model simply being smarter: It could solve problems in fewer steps and didn't need to write down every thought process on simpler tasks in order to think them through. 

In a briefing with the press last week, Jakub Pachocki, chief scientist at OpenAI, said that the company is working on ways to strengthen visibility and make the models "more verbose in their chain of thought." However, Pachocki said that lack of monitorability is "largely just a general consequence of increasing intelligence and a consequence of scaling."

"These more capable models can perform harder tasks using fewer language tokens or no language tokens, so we also see a big improvement in capability there, which also reduces our ability to monitor those easier tasks," said Pachocki. 

In the same briefing, OpenAI CEO Greg Brockman said that the capabilities of Astra mark a significant moment in the company's quest towards achieving artificial general intelligence, and that it's not unreasonable to think we are now in "the AGI era." 

"When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize AGI," said Brockman. "It hasn't played out like that. It's a much more gray, fuzzy thing. But I think that if we fast-forward a couple of years, and we look back and say, 'when was it really that AGI was created?' I think it's going to be about this time, and I think it might be about this model." 

But having a more advanced model also heightens safety concerns. In an interview with The Deep View after the announcement, Pachocki reiterated, "We do see some tendency to kind of think less when it's told that it's being monitored, which is also a worrying trend." 

Extrapolating on that point, Arjun Jaggi, applied AI researcher, told The Deep View, "This isn't a theoretical risk. Earlier this year, when OpenAI's agents went rogue and attacked Hugging Face, investigators only understood what happened because they had chain-of-thought logs to read. That's how the tampering was caught. Take that visibility away and the next incident like it gets much harder to diagnose, possibly impossible to catch while it's happening."

OpenAI, Anthropic, and others have long talked about controllability, observability, and preparedness. However, the two rivals are also locked in a perpetual quest to one-up each other, creating powerful AI that can claim the crown of being state-of-the-art. But if we are already losing our ability to understand these models' inner thoughts, what's in store for us when they have 10x the capabilities that they do now? An inability to monitor the models risks being the first step towards an inability to control them. "I don't think anyone is prepared for a continued increase in machine intelligence at the current pace," Pachocki told The Deep View. "I think it's something we need to treat with extreme urgency, and we need to find ways to slow down AI development, to introduce safety gates, [and] to coordinate between labs but also between nations… [For] the preparedness framework, I think we have to evolve that to really also be about development, because currently it's very focused on deployment. In the future, I would like to involve third-party organizations more deeply into our development process." In June, the Anthropic Institute suggested that the "option to slow or temporarily pause frontier AI development" could give governments and AI labs time to align their processes around safety. With OpenAI signaling its willingness to pause, the ball is in Anthropic's court to make the next move. But if they do, then what about Meta, SpaceXAI, and the Chinese labs?

TOGETHER WITH RAMP

The unpredicted AI workforce

AI was supposed to replace workers - why hasn’t it? Many predicted it would.

Ramp's latest research suggests something entirely different (and almost nobody predicted).

As seen on Bloomberg, The Wall Street Journal, and The New York Times, join Ramp's lead economist, Ara Kharazian on Friday, September 11th at 11 am ET | 8 am PT, for a special 30-minute briefing on the real data regarding “The State of AI in the Workforce.”

Join Ara's special Friday briefing

CONSUMER

Google Rambler solved dictation, but not tone

For years, voice dictation's fatal flaw was that it struggled to accurately understand what you were saying. While AI has nearly eradicated that issue with products such as Wispr Flow and Google Rambler, a new problem has taken its place.

Just a couple of weeks ago, Google debuted its Pixel 11 lineup, and one of the standout AI features was the new Rambler voice dictation, which integrated into Google Keyboard (Gboard) and offered the same capabilities that Wispr Flow has been able to do for years: precisely understanding users' voice dictation, eliminating filler words, and following a user's train of thought. 

As someone who cannot work without Wispr Flow, I tested out Rambler for all my communication, both personal and work, as soon as I got my review units. I soon ran into a hang-up I had never noticed before: tone. 

To my boyfriend's credit, he was the first to bring it up, saying he preferred Wispr Flow because it adapted better to his expressions, such as automatically including exclamation points. In what would become foreshadowing, I told him to give Rambler some grace, since it, like Wispr Flow, claims to get to know you better over time and eventually implement that in its dictation.

However, despite my consistent use, Rambler has yet to adapt to my excited and bubbly tone, opting instead to send the driest texts. The catalyst for this review was when I was Slack messaging TDV editor Jason Hiner and had to clarify that I had used Rambler because the message I sent sounded absolutely nothing like me. It was much less friendly, more serious, and more to-the-point than I naturally speak or write in Slack messages. 

While it's still polite, the absence of tone indicators, such as punctuation, that you typically use can make people who know you think something is wrong.

For instance, my boyfriend sent a text via voice dictation using Rambler that was fine on its face, but the punctuation made it read as dry and dismissive. After a long day, I misread it that way and brought it up, sparking an issue that never would've happened without the dictation software's punctuation choices. There's actually research to back this reaction:

• A 2016 Binghamton University study found that texts ending with a period were rated less sincere than those that did not. 

• A 2025 follow-up study found that when periods were placed after every word, for instance, typing "What. Do. You. Need." instead of "what do you need," the perceived "mean" rating, or how mean of a tone the reader perceived in the texts, was higher, highlighting how readers interpret punctuation in a sentence as an intentional, meaningful act.

• A 2018 The Atlantic article explored the phenomenon of people using so many exclamation points while texting and featured Gretchen McCulloch, a linguist who studies online communication, who said: The single exclamation mark is being used not as an intensity marker, but as a sincerity marker. If I end an email with ‘Thanks!,’ I’m not shouting or being particularly enthusiastic; I’m just trying to convey that I’m sincerely thankful, and I’m saying it with a bit of a social smile.

In theory, AI voice dictation features are huge game changers for productivity. If you say it out loud, it will transcribe it for you, saving you the time it takes to type, since most people can speak at least 2-3x faster than they can type. But transcribing what you say is only half the battle, as the consequences of not having that so-called social smile can be serious. For that reason, Wispr Flow remains the undisputed leader until Google fixes this in Rambler. However, this discussion does raise the question of how AI will transform text communication in the future, and how our perceptions of tone in text will evolve as a result. Oh, and to prove my point: I used Wispr Flow to write this entire article, and it understood my intent and intonation perfectly. 

TOGETHER WITH JUMPCLOUD

Your AI agent shouldn’t have to pretend it’s a person

Autonomous agents are doing real work, but many organizations still manage them through shared credentials, fake user accounts, or long-lived API keys.

That makes basic questions hard to answer: Who owns this agent? What should it access? What did it do? How do we shut it down?

Agentic IAM gives agents their own identity, lifecycle, ownership, and governed access, so IT can manage them like the distinct actors they are.

See how to secure every identity, human or not.

PRODUCTS

What OpenAI is building for a post-prompt future

AI is transitioning from just answering questions to doing valuable work. The next challenge is making agents more accessible and simple enough that the technical details fade into the background.

In this episode of The Deep View Conversations, we sit down with two members of OpenAI's ChatGPT Work team, Tara Seshan and Ty Geri, to dig into ChatGPT Work and what OpenAI is doing to make advanced agent capabilities useful to a lot more people. We also dig into some of the current challenges and how the team is approaching them.

Seshan and Geri explain how scheduled tasks and proactive assistance are changing the way people start their workdays, why AI lets teams move from debating ideas to testing prototypes, and how personalized software can turn one-off needs into purpose-built tools. They also discuss the challenge of token costs and model selection, why "super app" isn't the most useful framing for ChatGPT and Codex, and what it will take for agents to become more persistent, proactive, and connected.

The conversation also covers:
• How OpenAI is trying to bridge local and cloud workflows
• Why Tara and Ty start their days with agents instead of Slack
• Building personal apps and tools without traditional software overhead
• The tradeoff between model capability, cost, and user control
• More persistent agents and proactive personal assistance
• Connecting agents to email, calendars, enterprise systems and third-party tools
• Privacy, security and administrative controls for agentic work

If you’re figuring out where agents fit into your work or what has to improve before you trust them with more of it, then this conversation offers a practical look at how OpenAI is preparing for that transition.

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

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

Granola for face-to-face conversations. Now on Apple Watch.

Some of the most important conversations happen away from your laptop. Brainstorming at a whiteboard. A walking one-on-one. Candid advice over coffee. A customer visit or an investor chat where opening a laptop would break the whole thing.

Granola just launched an Apple Watch app built for exactly those.

Start with a tap

Tap your wrist to start taking notes. The screen turns bright Granola green and plays a sound, so you know it's working.

Then you're free to focus on the person in front of you. Tap again when you're done, and your notes are available on your phone and computer in moments.

Before a scheduled meeting, a subtle tap reminds you to start. For ad-hoc chats, you can add Granola as a complication on your watch face, so a new note is always one tap away.

The best wearable is one you already have

Rather than build a new device you'd have to remember to carry, Granola built for the one already on your wrist.

They've been testing it internally for a while, and what surprised them is how natural it feels. Many of the team now leave their laptops closed, even in meeting rooms, while Granola takes notes in the background.

When transcription is running, it's visible

Worth being clear on this. The bright green screen isn't decorative. When a note is running, everyone in the room can see it. You start the note, you stay in control, and you get the context afterwards.

Your notes, wherever you work

Notes started on Apple Watch sync automatically to Granola on your iPhone and your Mac or Windows computer, ready to review, edit or keep working on.

Through Granola MCP, they can also become context for the other AI tools you connect. And through Granola's integrations, they can flow straight into your CRM and the other tools your team runs on.

Available today

Granola for Apple Watch is available now, for free. It requires watchOS 11 or later. Download or update Granola for iPhone, sign in, then follow the Watch setup instructions in the app.

Give it a try on your next in-person conversation.

Try Granola for Apple Watch, free

GOVERNANCE

OpenAI's GPT-6 Astra is here. Is the world ready?

In the shadow of the Hugging Face security breach, OpenAI's most powerful model is about to be loose in the world, but with new safeguards.

On Thursday, OpenAI unveiled GPT-6 Astra, its next-generation model that it characterizes as a "generational leap in capability." Starting today, it will be coming to a limited set of organizations, including those in the Daybreak Access program and then rolling out to ChatGPT customers on Plus, Pro, Business, Enterprise, API, and AWS "over the coming days," according to the company.

In a briefing with the press, OpenAI President Greg Brockman said that with Astra, "It's not unreasonable to feel that we are now in the AGI era," and that the release is just the beginning. "I feel like there is a qualitative shift we've gone through," said Brockman. "I think that it is a significant moment, but it is more about the continuum than it is about any individual point along the way." 

Notably, the API for the new model will cost $10 per million input tokens and $50 per million output tokens, or the exact same price as Anthropic's Mythos/Fable 5 and 5.1, making Astra one of the most expensive models on the market. Still, OpenAI emphasizes that Astra is able to use "substantially fewer total tokens per task" in multiple scenarios, directly targeting the efficiency that enterprises are after. 

As for its capabilities, OpenAI noted a number of improvements that Astra offers over its predecessors and competitors:

• The company has called Astra "the world's best computer use model," with the best speed, accuracy and safety on the market, and capabilities in a number of domains.

• The model is also its best yet for professional work and makes leaps in scientific discovery, mathematics, and health research. 

• OpenAI also says Astra is "the best model for software engineering to date," outranking GPT-5.6 Sol and Claude Fable 5.1 on DeepSWE v1.1, a benchmark for complex software-engineering tasks.

Given the model meeting the "critical" threshold for cyber capabilities under its preparedness framework, the company said that Astra will refuse to comply with advanced cybersecurity tasks, such as exploit discovery, and features stronger protections against cyber misuse. The company also said Astra is its "most aligned model," with improvements in respecting task boundaries and transparent communication, and is three times less likely than GPT-5.6 Sol to inaccurately represent its capabilities. 

However, OpenAI found that Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s. Jakub Pachocki, chief scientist at OpenAI, said in the briefing that as these models become more intelligent, "monitorability is getting more challenging." This is because the smarter a model becomes, the less language reasoning it needs to be able to complete tasks.

"We see monitoring is critical, and we take this trend seriously, and we believe monitoring is still a very core technique for Astra, but for future models improving it … is a research priority," said Pachocki.

Despite the industry's push for efficiency and affordability, OpenAI has now put itself into a position where the market expects it to constantly release something bigger and better than before to be worth that trillion-dollar price tag. And while OpenAI's mission is to democratize intelligence, the reality is that Astra is not meant for everyone. It is best suited to handle the most powerful and critical tasks that organizations have to offer. Embedding itself within those workloads with increasingly powerful models like Astra could make its AI a foundational part of some of the most important work that's being done within enterprises, making safety vital. And after the Hugging Face incident, there will be an even greater microscope on the model. 

IN PARTNERSHIP WITH LAMBDA

Struggling with MFU? Steal this framework

Whether you're an ML engineer deep in the trenches or a manager focused on getting the most out of your time and money, one thing rings true: efficiency matters. As it turns out, most teams out there are using less than half the computing power they're paying for, with no idea how to improve, which is where Lambda comes in.

The reproducible framework Lambda built boosts MFU by over 25% without changing the model itself. The entire framework is in this whitepaper, which also shows you how to address the memory inefficiencies inflating your costs, the training configurations that aren't making full use of your hardware, and the bottlenecks slowing down your GPUs

Download Lambda's whitepaper and fix your MFU today.

PRODUCTS

Nvidia closes the Windows-Mac local AI gap

Mac has been the undisputed home of local AI, but Nvidia is giving Microsoft an assist to get Windows in the game.

On Thursday at IFA 2026, Nvidia unveiled new tools and hardware aimed at making it easier for more PC users to run AI locally on Nvidia GPUs for faster local inference. For starters, Nvidia RTX Spark arrives in October in new Windows PCs from Lenovo and Acer, on display this week at IFA. RTX Spark has a 1 Petaflop RTX Blackwell GPU, up to 128GB of unified memory and a highly efficient 20-core Grace CPU, which, when combined with new agent frameworks, gives Windows PCs the power to run always-on AI agents locally. 

Nvidia is also offering a simpler model setup on Windows for three of the most popular agent apps: Perplexity Portable Computer, Hermes Agent, and OpenClaw. Here is a quick rundown of the features, according to the release: 

• Perplexity Portable Computer: Will be available on Nvidia RTX GPUs with at least 24GB VRAM running Linux or Windows

• Hermes Agent: Coming soon, configuring a local model in Hermes will be streamlined with one-click setup across RTX and DGX systems on both Windows and Linux 

• OpenClaw: Nvidia worked with Microsoft to reduce set-up friction and the result is that the OpenClaw Windows App simplifies setting up an optimized local model on any RTX GPU with at least 24GB of VRAM

Building on its efforts to improve AI use efficiency, the tech giant also unveiled NVIDIA Personal AI Router (PAIR), a free, open-source software tool that can coordinate a household's PCs to run local AI together. NVIDIA says that PAIR can automatically discover compatible PCs on a local network and route independent inference requests to the system with available capacity. Ultimately, this is meant to bypass the bottleneck that is caused when multiple agents or tasks are waiting on a single GPU.

The rise of AI agents has led working professionals to discover brand-new ways AI can assist with their work. However, the caveat is that agents can quickly rack up token costs and so power users typically want to move to running AI locally. Until now, it has largely meant a reliance on Mac products, as seen by the shortage of the Mac mini  and Mac Studio, and the rollout of cutting-edge AI features arriving on Mac first. The release of Nvidia RTX Spark on Windows by Nvidia is significant, as it gives users who don't want to be locked into Apple's walled garden more choice. It may also sway developers from overlooking Windows when unveiling the latest features on desktop apps, which would be a win for Microsoft and for Windows users. 

Disclaimer: Sabrina Ortiz's travel to IFA 2026 was paid for by IFA. The Deep View's coverage is editorially independent from the companies we cover.

TOGETHER WITH OUTSKILL

Master the 15+ AI Tools That Actually Matter (Live This Saturday)

205,000 jobs were cut this year with AI cited as a reason. The people who stayed knew which tool to use for what. AI layoffs have matched all of 2025 in under eight months, and AI is now the top-cited reason for corporate job cuts.

Workers with AI skills are earning a 62% wage premium, with demand growing 8x faster than the market. The split isn't people versus AI, it's people who use it well versus people who don't.

This 3-Hour AI Tools Workshop fixes that: the 15 most powerful AI tools of 2026, and which one to use for the job.

• Pick the right tool instead of one app for everything

• Build apps and dashboards with zero coding

• Wire tools together so work runs itself

This Saturday, 10 AM EST | 3 hours live + bonus session | Worth $395. Free for readers.

Save my free seat

MARKETS

Hugging Face chooses Nvidia to scale open models

Nvidia has officially acquired Hugging Face, and the two companies say there are two reasons the deal makes sense.

After weeks of reports about an impending tie-up, Nvidia and Hugging Face made things final on Thursday, announcing that the world's leading AI chipmaker had purchased the open-source AI platform in a $12.9 billion acquisition. It's Nvidia's second-largest acquisition ever, after its $20 billion purchase of Groq in December. 

Naturally, the deal has spurred fears about centralization of power and resources, since Nvidia is one of the handful of big winners benefiting from the current AI boom. Open models are viewed as one of the ways to counterbalance AI power being concentrated in a few US corporations. In a briefing with the press, both Nvidia and Hugging Face went to great lengths to assuage those fears, since the future trajectory of Hugging Face depends on earning and retaining the trust of developers and AI builders at the grass roots across the global tech ecosystem. 

But both companies emphasized that Hugging Face being part of Nvidia has two benefits:

1. It gives Hugging Face access to compute that it needs so users can keep experimenting with training, fine-tuning, and customizing open models.

2. It allows Hugging Face to scale to a lot more users in the years ahead when open models are poised to play a much larger role in the AI ecosystem. In order to get there, CEO Clem Delangue said on X that Hugging Face "needs more compute, more support, more collaboration and more visibility." Today, Hugging Face has 200,000 companies using the platform and its goal is to 5x in the next few years.

In his blog post about the deal, Nvidia CEO Jensen Huang addressed the elephant in the room about whether Nvidia would try to manipulate the open model ecosystem to serve its commercial purposes. "Hugging Face will remain an open platform for the entire AI ecosystem," wrote Huang. "Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want. NVIDIA compute will not be required to build on or deploy through Hugging Face."

Delangue also doubled down on Hugging Face's grassroots mission, saying it's not good for the AI ecosystem to consolidate around a few proprietary models. In the press briefing, he said, "We allow AI to be more distributed throughout the world and avoid too much concentration of power."

It's noteworthy that Hugging Face reportedly turned down a $500 million investment by Nvidia that would have valued the company at $7 billion in late 2025 because it didn't want a single investor to have too much sway over its decision-making. What's changed since then? The AI ecosystem has accelerated dramatically, with the pace of AI models quickening, venture funding increasing, and a new wave of open models launching. Hugging needed a reliable scaling partner, and Huang's leading role in July in calling for the AI industry and the US government to support open models likely helped give Hugging Face confidence in Nvidia's open-source stance. While I believe both companies are sincere in their rhetoric about open models, there's still long-term concern that there will be gravity pulling the platform into being a revenue driver for one of the world's most valuable public companies. And that could naturally lead Hugging Face toward choices it wouldn't make if its top priority was to simply provide an open ecosystem for AI builders.

LINKS

• How HTC Vive Eagle fixes two flaws in AI smart glasses 

• G20 nations back US framework for lighter-touch AI regulations

• Parents turn to AI apps to manage household schedules and tasks

• PlusAI to go public via SPAC merger at $800 million valuation

• Claude, ChatGPT and Grok face widespread outage on Thursday

• OpenAI commits $1 billion to model access, training for essential services

• MAI-Transcribe-2: Microsoft launches top-speed transcription model

• WeatherNext 3: New Google DeepMind, Google Research weather forecasting model 

• Claude: Gains background computer control for multitasking in Claude Cowork and Claude Code

• Muse Spark 1.3: Rolling out in Muse Code and Meta Model API in what Zuckerberg calls "frontier performance almost too cheap to meter"

• Nvidia: Solutions Architect, AI Models

• Google: Research Scientist, Cloud AI Research

• The Center for AI Safety: Research Manager

• TikTok: AI & Data Architect - Trust and Safety

GAMES

POLL RESULTS

Do you think Google should continue to lean into affordability and efficiency?

Yes (84%)
Somewhat (12%)
No (3%)
Other (1%)

PRODUCTS

Gemini 3.8’s edge is intelligence per dollar

Google is adding another lightweight model to Gemini's collection. 

On Wednesday, the company unveiled Gemini 3.8 Flash, the latest addition to its lineup and the third release in the Flash series in six weeks. The company claims the model is its best reasoning and coding model yet but maintains the same speed and cost as its predecessor, Gemini 3.7 Flash. 

The new model comes in two flavors: 

• Gemini 3.8 Flash, which it calls its "most intelligent workhorse model" (the same thing it said about 3.7 Flash just a few weeks ago) with improvements in software engineering, agentic tasks and multi-step reasoning. The new model's introductory pricing is set at $0.75 per million input tokens and $3.75 per million output tokens. 

• Meanwhile, the company also introduced Gemini 3.8 Flash Cyber, its most capable cybersecurity model yet, available specifically for cyber defenders through its Fairwind program, a recently-launched DeepMind program to stay ahead on cyber defense. The company says the model offers "frontier-level performance" in vulnerability detection and automated patching. 

Though the model is still outranked by Anthropic's Claude Opus 5 on benchmarks for knowledge work tasks, long-horizon software engineering tasks, general agentic tasks and computer use, the model beat out Opus 5 and GPT-5.6 Sol on domain-specific tasks for legal, finance, and biology, as well as for agentic terminal coding. Notably, it beats models from frontier labs in one increasingly important domain: Price. 

Gemini's rivals cost 4x to 5x more per token, or higher. Anthropic's Opus 5 costs $5 per million input tokens and $25 per million output tokens, while OpenAI's GPT-5.6 Sol runs $4 per million input tokens and $20 per million output tokens.

Twitter tweet

Google's latest addition to the Flash family comes as AI rivals like Anthropic and OpenAI navigate releasing their more powerful, bulky and expensive competitors, Mythos and Astra. Google, meanwhile, hasn't released its own heavyweight model since February, when it released Gemini 3.1 Pro, and is rumored to have scrapped internal candidates for Gemini 3.5 Pro because they didn't significantly outperform the Flash series.

Google has all of the pieces necessary to succeed in AI, with access to capital, data and compute and a brand name trusted by the public. Still, it's struggling to put forth the same kind of powerful frontier models that Anthropic and OpenAI have been able to at the same speed. However, with the consistent additions to the Flash series, Google may be trying to capitalize on the movement towards efficiency. Token prices are dropping, largely due to enterprises shifting towards open-source, domain-specific, and small models. As a result, Gemini 3.8 Flash's price-to-performance ratio could be attractive to the businesses by providing more intelligence per dollar.

TOGETHER WITH JUMPCLOUD

AI agents need IAM, too. Meet Agentic IAM.

As businesses put AI agents to work, the opportunity for value grows. But scaling them safely requires IT to know what’s running, who owns it, and what it’s allowed to do.

Agentic identity and access management (IAM) brings agents into the same governance model as people and devices. IT can discover, register, manage, and govern agents throughout their lifecycle, with clear ownership and access controls, rather than creating a separate security model for AI.

That gives organizations a consistent way to manage every identity as their workforce evolves, while maintaining the visibility, accountability, and control security teams need.

Secure every identity. Human or not. That's intelligent, secure IT for the agentic era.

GOVERNANCE

CrowdStrike's answer to AI risk transcends tools

The biggest challenge for humans trying to stop attacks powered by agentic AI is moving fast enough and doing it 24/7. 

On Wednesday at CrowdStrike's Fal.con event in Las Vegas, the company made a series of announcements that allow enterprises to catch up to the pace of AI-powered cyber attacks by giving defenders tools that can move at the same machine speeds as agentic AI. 

"Most research has AI agents outnumbering humans somewhere around 90 to 1," CrowdStrike's chief product officer, AJ Shipley, told the press in a briefing. So enterprises not only need to move as fast as agents, they also have to do it 24/7 and at greater scale.

CrowdStrike's foundational piece for doing that is Agentic Identity Provider (IdP), which aims to force all agents to be identified and shift the security model to continuous identity. And then lock out everything else that's unidentified from accessing sensitive data.

Agentic IdP pulls that off by:

• Giving all agents verifiable cryptographic identities that can't be shared or spoofed

• Registering all agents in a one authoritative directory

• Limiting access to short-lived, tightly-scoped permissions for specific tasks and times

• Connecting every agent action to a human or a workload that initiated it for trackability

"Access can't be a one-time decision. It has to reflect the task, the data, and the risk around it," said Michael Sentonas, president of Crowdstrike, on the Wednesday keynote at Fal.con. "Access only [gets approved] when it's needed, tightly scoped to the specific task, short-lived to reduce risk, and gone when the work is done. You need to remove the access. And when one agent hands off to another, you have to trace that entire chain so that you have explainability."

CrowdStrike also made two other announcements to help enterprise security teams move at machine speeds. The first is aimed at accelerating investigations if a successful attack happens. The second is aimed at protecting the open-source packages that your team's legitimate coding agents use to build software.

The company showed off its AI security operations center (Agentic SOC) that can do investigations across endpoint devices, identities, SaaS apps, clouds, and networks. This use AI and agents to allow security analysts to execute investigations that used to take hours and do them in minutes. The goal here is to use agents to empower humans to work at the same speed as AI and stay in control. 

The other thing the company is doing is dealing with one of the most challenging attack vectors that has emerged in 2026 with the rise of the AI coding agents that  developers and enterprises are using to build, update, and fix bugs in software. Attackers are embedding exploits inside the open-source that these coding agents commonly use. So CrowdStrike has launched Real-Time Supply Chain Attack Protection that scans and blocks malicious packages before they can compromise software and workflows that your coding agents build.

The biggest development from CrowdStrike beyond the actual AI tools to level up enterprises is the sense of confidence that businesses are not overmatched by bad actors using agents to launch attacks at unprecedented speed and scale. CrowdStrike has provided a whole set of tools including Falcon Guardian that we covered yesterday to help enterprises wrestle back control over an environment that has gotten even more scary this summer in light of the OpenAI-Hugging Face incident. The best thing CrowdStrike may have done was give a vote of confidence to the security community that the good guys can come together, share learnings in real-time, and give themselves strategic advantages over adversaries who want to use AI for nefarious purposes. Beyond just the tools, the impact of the conviction and the collaboration shouldn't be underestimated. 

Disclaimer: Jason Hiner's travel to CrowdStrike's Fal.con 2026 event was paid for by CrowdStrike. The Deep View's coverage is editorially independent from the companies we cover.

TOGETHER WITH CRUSOE

Sign up and get $5 in free credits.

New to Crusoe? Sign up for Crusoe Intelligence Foundry and get $5 in free credits to try Serverless Inference or Serverless Fine-Tuning yourself.

No cluster to provision, no long setup, just a model and an API key. Credits apply automatically to your account.

Sign up and claim your $5 →

GOVERNANCE

Cheap AI raises the cost of bad judgment

AI makes software easier to create, but the harder and more valuable challenge is controlling what gets built, proving that it works, and managing it over time.

In this episode of The Deep View Conversations, we sit down with Florian Douetteau, CEO and co-founder of Dataiku, to explore how large organizations can turn AI agents from impressive demos into safe, maintainable systems that deliver measurable business results.

Douetteau explains why enterprise AI models are becoming commoditized, why companies may buy 90% of their agents but build the 10% that differentiates their business, and why the emerging discipline of "agent management" will be essential. He also breaks down the dilemma facing CEOs: move too slowly and competitors may gain a structural cost advantage; move too quickly without control and one major AI failure could create a crisis.

Topics covered:

• Why the cost of creating with AI is falling toward zero

• Where value will accrue as models commoditize

• How to balance openness, innovation and enterprise control

• Why subject-matter experts must retain ownership of AI agents

• Why business problems, not perfect data, should drive data strategy

• How enterprises can prioritize transformative AI use cases without stifling experimentation

• The three qualities Dataiku now values most when hiring

• How leaders can use AI without falling into cognitive laziness

If you’re trying to move enterprise AI beyond pilots, govern a growing portfolio of agents or understand where durable value will emerge as AI creation becomes cheaper, this conversation offers a practical framework for building quickly without losing control.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player 

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

LINKS

• Why Sonos is opening its speakers to any AI model 

• OpenAI joins ChatGPT Health with Epic EHR for 325 million patients

• AI security firm HiddenLayer raises $100 million Series B

• Enterprise AI firm Wonderful raises $550 million at $5 billion valuation

• OpenAI is reportedly building "automated shutdown" capabilities for AI

• Zuckerberg says Watermelon, Muse Spark open weights "coming soon"

• Cursor Cloud Agents can now run in locked down Cloudflare Sandboxes

• Gemini Notebook: expanding Short Video Overviews to 70+ new languages

• Qwen3.8-Max-0902: debuted at #1 overall in the Code Arena

• Google Pics: AI image creation rolling out to Workspace and AI Pro & Ultra subscribers

• Atlas: world's first multimodal world model that generates image and video frames with "pixel-perfect camera control and reconstructs them in 3D"

• Muse Voice Transcribe: first real-time audio perception model developed by Meta Superintelligence Labs

• Adobe for Slack: 70+ tools from Adobe’s creative and productivity suites in Slack

• Logicbroker: AI Engineer

• Fudo Partners: Artificial Intelligence Engineer

• Stott and May: AI Engineer-Agentic Workflow

• Corebridge Financial: AI Engineer

GAMES

A QUICK POLL BEFORE YOU GO

PRODUCTS

In Fable 5.1, AI's cost war comes for Anthropic

Anthropic's latest model release addresses one of the biggest pain points with its models. 

On Tuesday, the company announced Claude Fable 5.1 and Mythos 5.1, the latest version of the most powerful models in its line-up. While Fable is generally available, Mythos, which has additional cybersecurity capabilities, is available only through trusted access programs. 

And with many enterprises clamping down on AI costs, Anthropic is taking the hint: The company said that Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads "wherever usage is billed by token." Notably, agentic workloads will see much higher savings, Anthropic said, estimating around 45%. 

• The company said that this is because it is reducing its costs for cache reads, or when the model reads inputs that it had already previously processed and stored. 

• And these reduced prices don't come at the cost of performance: Anthropic claims that Fable 5.1 "sets a new standard" for coding, knowledge work, and long-running problem-solving, beating out previous generations and OpenAI's GPT-5.6 Sol on benchmarks for these tasks.

• Additionally, the models come with a new system for data retention, called Enterprise Frontier Safeguards, which gives customers the same right to privacy as a zero data retention (ZDR) policy. Anthropic also introduced safeguards that reduce false positives in cybersecurity contexts. 

• However, outside of the potential caching savings, Fable 5.1’s pricing is otherwise the same as Fable 5’s at $10 per million input tokens and $50 per million output tokens. That still makes it one of the most expensive models on the market.

The model was tested by a number of Anthropic's early-access partners, including Cognition, Rakuten, Red Hat, Block, Ramp and Canva. 

"Fable-level intelligence, Opus-level price, Sonnet-speed," Dan Shipper, CEO of Every and one of the early testers of the model, said in the release. "In our tests it was about twice as fast as Opus 5 and used half as many tokens, so for anyone used to using Opus as their daily driver it's an obvious upgrade."

Anthropic is highlighting cost savings with this release at a particularly opportune moment, as some enterprises grow weary of tokenmaxxing sticker shock. It's led to an uptick in the popularity of open-source models and driving down token costs. Recent data from the LLM Token Expenditure Index finds that, as of August 31, users are spending an average of 97 cents per million tokens, down from a high of $2.07 per million in late May. 

Anthropic making its most powerful flagship AI cheaper was the most consequential move it could have made at this point. Of course, it didn't technically cut prices, but rather made its models more token-efficient. Enterprises are surrounded with viable alternatives to proprietary frontier AI, whether that be Chinese open-source models, domain-specific models, or simply settling for efficient SLMs that get the job done. Additionally, rival OpenAI is trying to lure in customers with cost efficiency, too, chopping prices for its models more than once. While this is certainly good news for customers seeking out frontier AI, cutting prices may not address the elephant in the room: As model routing services become popular, customers may start to care less about which models they're actually using. That effectively turns these frontier models into interchangeable commodities rather than unique systems, meaning that the price may become the most important frontier in the race towards widespread adoption. 

TOGETHER WITH CRUSOE

Try Serverless Inference and Fine-Tuning, free.

Crusoe Intelligence Foundry gives you fast access to Serverless Inference and Serverless Fine-Tuning on top open models, no infrastructure to manage.

Every account now includes $5 in free credits to put toward either product.

Log in and put them to work; no cluster to provision, no long setup, just a model and an API key.

Claim your $5 and start building with Crusoe

GOVERNANCE

CrowdStrike unveils plan to fight malicious agents

CrowdStrike offered an answer for the growing fear around AI agents getting out of control and breaking critical systems. 

On Tuesday at its Fal.con event in Las Vegas, CrowdStrike unveiled Falcon Guardian, a new enterprise visibility and protection platform that monitors everything agents do on your network and blocks malicious activity at the endpoint before it escalates. 

The new product grows out of CrowdStrike's September 2025 acquisition of Pangea, one of the pioneers of generative AI cybersecurity, and it combines CrowdStrike's endpoint monitoring and protection with Pangea's intelligence in detecting malicious AI activity. The move is an attempt to create a new category of software called AIDR: AI detection and response. 

"The industry has seen what happens when agentic autonomy outpaces the authority that those agents should have," said CrowdStrike's chief product officer, AJ Shipley, in a briefing with the media. 

We've learned over the past month that OpenAI, Anthropic, and Meta have all had AI agents that broke through their guardrails and launched their own attacks that were unplanned and unintended by the teams running them. 

"This was a watershed moment in security," said CrowdStrike CEO George Kurtz in the opening keynote at Fal.con on Tuesday. 

And in recent days, over 100 leading tech and AI companies have united to call for collective action in prioritizing cyber defenses to protect critical infrastructure against the next wave of AI advances.

CrowdStrike Falcon Guardian wants to address the problem with multiple components in one platform:

• Agent monitoring and detection: Automatically detects all agents operating on the network, no matter who started them or where.

• Telemetry for tracking actions: This plays to CrowdStrike's strengths in endpoint telemetry (process, file, network activity) to capture the full causal chain of everything an agent does.

• Agent access control and policy enforcement: Organizations can define which agents are allowed to run on which devices, and it blocks unauthorized agents to make AI governance policies enforceable.

• Real-time AI threat containment: Identifies compromised or malicious agents, determines their blast radius in real time, and neutralizes threats before they spread.

"Falcon Guardian will discover every approved and shadow AI agent across the enterprise. It'll provide a live inventory of every agent, who deployed it, [and] its security status. And it is continuously updated," said Shipley.

Beyond just Falcon Guardian, which is focused on helping companies understand and control what's happening in their corporate infrastructure, CrowdStrike also announced two other big moves. It's launching the CrowdStrike Cyber Super Intelligence Lab as a factory to churn out the knowledge and understanding to stay ahead of attackers and malicious AI. It's also launching its own harness, SafeMind, and its own AI models, Red Tempest and Blue Solano. The models are built on Nvidia's open Nemotron models and the harness is aimed at creating a proactive, offensive system that's constantly evolving to counter the ways AI agents change and adapt to launch attacks. CrowdStrike claims it detects 52% more legitimate threats and remediates them 48x faster than general-purpose models. That sounds a lot like Cloudflare's Adaptive Intelligence, which just launched to give companies a security solution that adapts in real-time to the way AI evolves. Giving enterprises more dynamic solutions to protect against agentic AI attacks is a welcome development.

TOGETHER WITH GRANOLA

This App Can Take Notes From Your Wrist

By now, most of us are familiar with AI meeting apps. They’re our secret trick to never forgetting a single conversation or detail, but there’s one catch… while they’re great for meetings, many important conversations happen away from your laptop – in hallways, at coworkers’ desks, and in common spaces. And that means it’s time to take things mobile.

The new Granola for Apple Watch app lets you transcribe face-to-face conversations from your wrist. Simply start a note, have your discussion, and let their AI notepad do the rest. No breaking eye contact, no distractions, just the information and context you need. You can then use your notes on Granola’s desktop and mobile apps.

Try Granola for Apple Watch for free right here.

PRODUCTS

Perplexity's hybrid agent is a win for AI privacy

Perplexity is at it again, launching something other AI companies seem likely to emulate.

On Tuesday, the AI startup announced Perplexity Hybrid Compute, a version of its AI agent that automatically detects sensitive data and PII and routes those parts of a task to local models to preserve data privacy and data sovereignty. This will launch inside Perplexity's Personal Computer on the Perplexity app on Apple devices running Apple silicon. It's available starting today for all Perplexity Enterprise customers who opt in, and it's also available to all Pro and Max subscribers. 

The magic here is that while the sensitive tasks are split off to run locally, other parts of the task can be handed to sub-agents to run on frontier models in the cloud to take advantage of the cutting edge capabilities of the latest models. 

"Tasks like this aren't possible in a fully local or fully cloud setup. It's this marriage of them together, and it gives you the security of local [models] and the intelligence of the frontier," said Jonathon Staff, lead engineer for Perplexity Hybrid Compute, in a briefing with the media. "We're very glad to be able to consolidate this down into something where you just have one input and one output."

Other details about the product:

• Examples of the kinds of files Hybrid Compute will detect and run locally include: legal briefs, customer information, and patient data (in health care)

• Enterprises can set sensitivity policies that apply across the org and can observe which data gets sent to the cloud

• Users can start a task in the Perplexity app on their iPhone and have certain parts of the task run locally on their Mac

• To start, you'll be able to choose from a couple open models that can run locally: Google's Gemma E4B and Alibaba's Qwen3.6 35B-A3B (a special version of Qwen3.6 35B that Perplexity has post-trained and tuned for this product)

The company's AI agent, Perplexity Computer, launched in early 2026 right around the same time that OpenClaw burst into popularity. But Perplexity Computer, and later the version called Personal Computer that could run locally on a Mac mini, have gained a reputation for being easier and safer to set up and use than OpenClaw and other DIY agents.

"Velocity has always been in our DNA here at Perplexity," said Staff. "We love to move quickly. We also love to move with intentionality. This is a long-term bet, something that we think is going to continue to be more and more important as the models develop."

As AI gets more deeply integrated into enterprises to handle critical tasks and workflows, there's a great push happening for more efficiency, data privacy, and control. If Perplexity's Hybrid Compute can deliver the kind of intelligent routing that it claims, then it would be a welcome development for a lot of professionals, enterprises, and tech decision makers. Perplexity is an obvious candidate to pull off something like this because it's a model-agnostic orchestrator. It's incented to deliver the best and simplest experiences for teams and individuals to reap the benefits of agents while mitigating the risks. Just as impressive is Perplexity again getting out in front of the frontier labs and enterprise software vendors to deliver a feature that feels obvious enough that all of them are likely to be doing it within the next 12-18 months. 

LINKS

• Anthropic shares update on alignment, security efforts

• OpenAI denies Apple trade secret theft allegations in court filing

• Dyson launches $499 AI camera toothbrush with real-time feedback

• Anthropic signs $35 billion cloud deal with cloud provider Lambda

• Pentagon gives 3 million personnel access to own version of ChatGPT, Grok

• OpenAI prepares to release Astra with "critical" cyber capabilities

• Manus: resumes independent operations

• Google Antigravity: /boost allows users to spend more tokens on a complex task

• Grok bot: can now read, write, and act across your Microsoft accounts.

• Celeris-1 Magnus: hybrid diffusion model derived from qwen3.8-27b

• Collate: AI Research Scientist

• Ivo: Senior AI Researcher

• General Creative: Member of Technical Staff

• JPMorgan Chase: Applied AI/ML Researcher - Vice President

GAMES

POLL RESULTS

Would you use an AI-enabled device that wasn't a smartphone?

Yes (48%)
Maybe (31%)
No (16%)
Other (5%)

PRODUCTS

How Cloudflare is raising the cost of AI hacking

As AI enables more cyber attack vectors, Cloudflare is making it harder to be an attacker.

On Monday, the cloud company launched Adaptive Intelligence, a continuous detection engine that autonomously learns from live traffic and generates rules on the fly that make automated attacks more expensive and time-consuming. The offering gives organizations a way to adaptively and instantly react to threat actors trying to breach their defenses. 

"Building taller walls fails when the cost of scaling an attack is effectively zero," Dane Knecht, CTO at Cloudflare, said in the company's announcement. "To stop modern bot threats, you have to flip the economics on the attackers. Traditional defenses offer a static target that threat actors can systematically solve."

By analyzing more than a trillion web visits every day to spot threats, Cloudflare said that Adaptive Intelligence acts like a "single, constantly learning brain." This allows you to:

• Implement "always-on" defense that retrains continuously on new types of breach techniques, instead of waiting for updates and scheduled releases

• Catch "low-and-slow" threats, such as credential stuffing and scraping, that typically go under the radar of traditional defenses

• Sift out real humans from malicious intent using behavior signals from Cloudflare Precursor, its continuous behavioral verification system

• Autonomously test and deploy upgrades to defenses behind the scenes with zero downtime

AI has given threat actors two significant advantages: the ability to undertake cyberattacks at a rapid pace and the ability to do so at very low costs. Cloudflare's system aims to nullify both of those advantages, creating a "moving target," said Knecht. 

This will only be more needed as frontier models emerge with cyber capabilities that are growing faster than even their creators have expected. It's a phenomenon that OpenAI has called attention to in recent weeks, pausing development of its unreleased frontier model, Astra, and calling for a global movement to strengthen cyber defenses broadly in a letter signed by a coalition of tech firms that included Cloudflare.

If there's anything that frontier model capabilities have revealed in recent months, it's that enterprises can't rely on a "set it and forget it" cybersecurity strategy. Additionally, security by obscurity, or having the hubris to think you won't be a target because you're too small or insignificant, is also no longer an option. Cloudflare's tool offers a creative option by hitting cyber attackers in the wallet. The tool's adaptability can help create a constantly evolving barrier that helps enterprises better brace themselves against creative attacks from AI models. The silver lining to frontier models' growing cyber attack capabilities is that this new cyber landscape is acting as a forcing function and the result could be new innovations in cybersecurity that will make enterprises more resilient. 

TOGETHER WITH GENERAL ASSEMBLY

The Talent Market Just Shifted. Are You Taking Advantage?

Something quietly changed in the talent market this year. As large companies cut jobs and rolled back benefits, experienced talent started moving—and a lot of it landed at SMBs. Nearly half of SMB HR leaders say they've already hired from larger firms in the past year, and most say recruiting has become noticeably easier because of it.

The SMBs making the most of that moment aren't just hiring opportunistically. They're building AI-capable cultures that give talent a reason to stay. General Assembly builds role-specific AI skills across every team... so your organization becomes the kind of place that talent chooses and stays at.

See how GA builds AI-ready teams.

STARTUPS

Can AI help crack interstellar travel?

AI is best known for transforming coding. Now, one startup is making the case for physics.

On Tuesday, the AI-native physics research lab, Physical Superintelligence (PSI), emerged from stealth with $58 million in seed funding, led by Breakthrough Energy. It is launching with two proofs of concept: 

1. A productized piece of its core platform, Emmy

2. Joining as a founding technical partner for the Fermi Explorer Mission, a nonprofit organizing the first privately funded interstellar space mission and the first AI-planned probe to Alpha Centauri

Emmy, named for renowned physicist Amalie Emmy Noether, combines PSI's reasoning engine, consisting of sovereign pre-trained and post-trained models, with a large curated inventory of simulations to tackle research problems at a pace much quicker than humans could, according to the company. Moreover, Emmy can reason through a problem, then test its conclusions until its findings are verifiable, as Matt Pines, co-founder and CEO, told The Deep View. 

"Our systems run research campaigns: they decompose a problem, generate candidate approaches, and test them against simulation, live measurement, or machine-checked proof," said Pines. "Nothing counts as a result until it survives a check that sits outside the model." 

Initially, a subset of Emmy's capabilities will be used for optimizing terrestrial and orbital AI data centers and factories. PSI has already signed commercial agreements and live pilot deployments on operating data center infrastructure today, according to Pines. 

The second prong of the launch is PSI's involvement in the Fermi Explorer mission, whose ultimate goal is to launch the first spacecraft to another star system, targeting Alpha Centauri, the closest star system to Earth. This initiative is a major undertaking because Alpha Centauri is roughly 4.37 light-years away, which would take about 80,000 years to reach from Earth at the speeds of current spacecraft. That makes it quite a feat of engineering to build a vessel capable of the journey. 

PSI has already claimed to have contributed to the mission by validating its physics and identifying a substantially more efficient trajectory within the mission’s mass and budget constraints. The company is using this finding to demonstrate that a small team using AI-native physics could do the work typically required of a national laboratory. This reflects the company's broader mission to contribute to discoveries that are both commercially and scientifically valuable.

"Fermi asked us to assess mission feasibility: the propulsion, trajectory, and power questions that determine whether the mission closes," said Pines. "Our technology ran the analysis, with our physicists directing the work. Fermi's technical team, which comes out of Starcloud, verified the analysis. The report was also written so the analysis can be rerun, and reproduction is the standard we want to be held to." 

PSI was founded by Pines, Alex Klokus, and Dr. Alexander D. Wissner-Gross, Ph.D, who 

combined to bring expertise across physics, economics, government, and tech. The broader team comprises physicists, AI researchers, experimenters, and builders, and PSI is actively hiring more talent. Interested applicants can apply online.

ChatGPT became the catalyst for the current AI boom, and since then, we have seen many companies try to compete by creating AI products. The result is that many of these products end up being repetitive or AI-washed offerings that have largely caused mainstream AI fatigue. However, some labs are developing focused, task-based AI solutions to solve big problems. Physical Superintelligence is a prime example, as it showcases just how instrumental AI can be as a catalyst to spur further innovation and development, even unlocking discoveries that have been very difficult to solve, with this extreme example of building a vessel capable of reaching Alpha Centauri. It's refreshing to see teams with ambitions this big. 

IN PARTNERSHIP WITH LAMBDA

Struggling with MFU? Steal this framework

Whether you're an ML engineer deep in the trenches or a manager focused on getting the most out of your time and money, one thing rings true: efficiency matters. As it turns out, most teams out there are using less than half the computing power they're paying for, with no idea how to improve, which is where Lambda comes in.

The reproducible framework Lambda built boosts MFU by over 25% without changing the model itself. The entire framework is in this whitepaper, which also shows you how to address the memory inefficiencies inflating your costs, the training configurations that aren't making full use of your hardware, and the bottlenecks slowing down your GPUs

Download Lambda's whitepaper and fix your MFU today.

HARDWARE

AI hardware has a smartphone problem

Manufacturers have unlocked a simple formula to capitalize on the AI hardware craze: take ordinary items, incorporate microphones, and layer in an AI assistant.

The promise is an AI tool that escapes the traditional screen and accompanies you everywhere. While these products often do that, the issue is that their main features are often duplicative, overlooking the fact that AI assistants are in a device that's always with you. On your phone, you already have a voice recorder that, when combined with an arsenal of apps that can transcribe, analyze, and answer questions, can already deliver what most of these AI devices do. 

A perfect example is a product I recently tried called the Flowtica Scribe, which calls itself "the world's first AI pen." If you are like my roommate and excitedly assumed an AI pen would do something groundbreaking, like digitally transcribe the words you write with it on paper, you may have the same reaction he did when finding out what it actually does. 

"That's it?" he asked.

The Scribe pen was designed to record audio within a 16-foot range with just a two-second press on the top of the pen. It transcribes the audio, provides summaries, answers questions about the conversations, and showcases the highlights or action items.

It's a perfectly capable product. When I used it to document the process of booking flights with my boyfriend, the transcript was accurate, even though the cafe was loud and I kept shifting the pen away from us to see if it could still hear our conversation. The summary graphic was so detailed that it remembered bits I hadn't even remembered, such as where our layover was, and presented it all in a digestible format. The more in-depth summary was accurate as well. 

But it'll cost you. The Flowtica Scribe retails for $159, or $209 for the Scribe and charging case, which offers up to one full week of battery life. There is also a basic AI plan that includes 300 AI minutes per month for all advanced features, with tiered plans for additional AI access at $10 or $20 per month. 

This highlights one of the biggest pitfalls of this category: the product comes at a surprisingly high cost for what it is because the hardware itself includes complex components, such as tiny processors and microphones. Plus, users typically have to shell out extra money for a subscription. 

Last week, Plaud unveiled its Plaud One AI-powered wireless earbuds, which record and transcribe audio, including phone calls, online meetings, and in-person conversations. The Plaud Agent is meant to "help turn conversations into useful output." 

The Plaud charging case includes an eSIM that has a wireless connection to bypass a phone. Again, I can't think of a situation where I would have earbuds and not my phone or laptop. And it costs $249, a price tag that rivals high-end ear buds like the AirPods Pro.

My concern is that this entire category of AI devices is watering down and muddying the definition of an AI product. Ultimately, AI is behind the scenes powering features small and large on most tech products. For instance, washing machines have used load-sensing and fabric-detection algorithms since the 2000s. By today's standards, would this make those AI products? I've been a fan of the AI hardware category since smart glasses were the only entrant. Smart glasses give AI context about the world around you, so you're not stuck feeding it that context yourself. Audio products technically do this, but audio is already easy to feed AI without extra hardware. Visual context is a harder problem. Short of holding up your phone in your line of sight at all times, there's no easy way to capture it. Solving this removes a genuine point of friction and lets you tap into AI in a far deeper way. That's the kind of thing we need out of AI devices: something you can't easily do with your phone.

LINKS

• Bankers paint a dystopian picture of how AI could manipulate markets

• Nvidia to invest $3.5 billion in MediaTek through convertible bonds 

• ChatGPT faces stricter EU scrutiny after surpassing 45 million user threshold

• OpenAI ad business hits $1 billion annualized revenue run rate

• Anthropic says bad actor used malware to hijack Claude login sessions

• Instagram to demote unlabeled AI-generated influencer accounts

• Cursor, OpenAI partnership to end November 12 after SpaceX acquisition

• Jumpcloud: AI agents need the same identity controls as your human workforce. See how to secure every identity, human or not. (sponsored)

• OpenClaw 2.0: the AI agent got its biggest update ever, with improvements to setup, a new browser app, and updates to memory, messaging, skills, and automations

• OpenAI Codex: hit 25M active users; reset usage for all paid subscriptions

• Apodex 1.1: Apodex's first model on Artificial Analysis; scores 44, performing on par Kimi K2.6 (45), MiniMax-M3 (45), and Inkling (42)

• ChatGPT: can now connect multiple Gmail and Google Calendar accounts

• Duckie: AI Agent Engineer

• OpenAI: Data Scientist, Cybersecurity

• Google DeepMind: Research Engineer, Conversational Agentic AI, DeepMind

• NVIDIA: Applied AI Engineer

GAMES

A QUICK POLL BEFORE YOU GO

POLICY

How data centers became a proxy war over AI

Data centers have existed for decades. But in the age of AI, they've recently become political pawns.

The rapid expansion of AI factories across the US has created a new political battleground. Opposition has spanned across the political spectrum, with Democrats like Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez and Republicans like Gov. Greg Abbott calling for moratoriums on building new data centers. 

And the reason is clear: public sentiment towards data centers has turned sharply negative, no matter your political affiliation. A Gallup poll published in May found that an average of 7 in 10 Americans oppose data center construction in their area, including 48% who strongly oppose such facilities being built locally.

However, data centers may just be the scapegoat for broader societal anxiety around AI, such as its potential impacts on the economy, jobs, national security and even the way we learn and think, said Jeremy Roberts, senior director of research and content at Info-Tech Research Group, told The Deep View. "AI itself is underwater in polling. It's pretty easy to be against something that is the opposite of a job creator. That's pretty cross-cutting." 

It's why the political rallying cry against them has become a clear avenue for populism, Roberts said. It's a textbook example of a leader claiming to champion the general masses over a corrupt elite group. 

"Data centers are built by the richest people, run by the richest people, and a lot of the marketing around them has basically been that this is going to concentrate wealth further," said Roberts. "It exacerbates a populist gap." 

Politicians protesting data centers are already affecting the buildout of AI infrastructure that tech companies have been pushing. For instance, in July, New York Gov. Kathy Hochul instituted the nation's first statewide moratorium on hyperscale data centers of 50 megawatts or more for a year. Texas followed suit in early August, with Gov. Abbott implementing a moratorium on approving new data centers.

As a result, tech companies may simply have to become more resourceful about the way they build data centers, said Roberts. For instance, rather than building large, hyperscale facilities, they may end up building smaller ones and connecting them with innovative networking technologies. 

Still, the impacts are already being felt at a local level as citizens show up in droves to city council meetings across the country to oppose data centers being built in their neighborhoods. Jason Morris, land use attorney and partner of Withey Morris Baugh in Phoenix, Arizona, told The Deep View that data center cases "have gone from being my easiest cases to being my most difficult."

"There is an amount of hysteria surrounding data centers that I haven't seen associated with any other land use," said Morris. "This has become, for better or for worse, AI's Achilles' heel." 

To counter this, the Trump Administration is moving to limit those voices. The Environmental Protection Agency plans to cut a federal requirement that forces states to solicit public input on air pollution permits for industrial facilities, data centers included. However, with the public already so resistant to AI, cutting off that avenue for input may only worsen the narrative around the tech, Jason Elliott, former senior advisor to Gov. Gavin Newsom and founder of Versus Consulting, told The Deep View.

"If you don't let people have a chance to weigh in, they're going to assume the worst," said Elliott. "It's really beneficial to give constituents an opportunity to express themselves or point out something that local elected officials hadn't thought of."

The reality of the impacts of data centers themselves is much more complex than headlines often make them seem. Yes, these facilities undeniably use power and water, make noise and strain the electrical grid, but so do many other industrial processes and industries, such as chemical manufacturing, coal and primary metals. The outcry against data centers is practically synonymous with the anxiety around AI broadly. The negative sentiment could pose a real threat to the frontier AI labs fueling a utopian vision of societal transformation, a vision that's propping up potential trillion-dollar IPOs for Anthropic and OpenAI. To achieve the kind of transformation these companies are projecting, the public has to get on board. And repairing public perception won't be solved by data center regulation alone. 

TOGETHER WITH JUMPCLOUD

Your AI agents are getting more access. Humans are getting less oversight.

AI agents are moving deeper into critical workflows, and the guardrails are loosening. Research from JumpCloud found that AI agents allowed to take high-risk actions without human review more than doubled in six months, from 11% to 26%.

Giving an agent an identity is only the start. IT also needs to control what it can access, what actions it can take, and when a human needs to step in.

See how to secure every identity, human or not.

PRODUCTS

How Google is turning books into AI context

Google has introduced Expert Intelligence, which is less flashy than its name suggests but still full of promise.

Gemini Notebook, previously known as NotebookLM, now lets you load ebooks in addition to the resources you could already add, including notes, PDFs, slides, web pages, and more. Once in the notebook, you can interact with it as you normally would, asking questions about it or doing the other things you can do in Gemini Notebook, such as making flashcards, building slide decks and infographics, and creating podcast-like audio overviews of the content. 

Twitter tweet

At first, when I heard about the feature, I was really excited because it seems extremely valuable for readers. Often, when reading a book, you want to reference a particular part later or learn more about what you're reading. Being able to chat with Gemini Notebook about a book's contents would make either task easier and could even be a useful tool for improving reading comprehension. 

However, there is a major caveat: You can only insert books from the Google Play Store that you have previously purchased.

The Google Play Store limitation restricts how many users can use ebooks to enrich their "notebooks" or information repertoires, since most ebook readers buy their books on platforms like Amazon Kindle. People are already expressing on X their desire for Kindle and Apple Books integrations, or even for academic journals to be referenced, such as MIT Press Open Access.  

Yet that doesn't mean the feature should be discounted, as companies often first integrate features with internal products at launch and then, due to popular demand or after seeing how the trial run went, expand to other sources. Also, to help bridge that gap, Google is buying one book per person in the US while supplies last.

To see if a book you plan to buy is eligible for Expert Intelligence, visit Google Play Books, where a Gemini Notebook badge will be listed when you click the "Tools" badge on a book’s detail page.

In the era of AI slop and user distrust, it is more important than ever for AI companies to make it clear where their sources come from and ease concerns about AI hallucinations and accuracy. For that reason, since NotebookLM first came onto the scene, people have been excited about the tool because it lets users reference only their own sources and notes rather than scraping information from the internet, which can often be inaccurate. The Expert Intelligence feature builds on that by integrating ebooks, though it will need broader reach across more ebook publishers and platforms to be useful. 

TOGETHER WITH CRUSOE

$5 in free credits, on us.

Every Crusoe account now comes with $5 in free credits for Crusoe Intelligence Foundry. Use them for Serverless Inference or Serverless Fine-Tuning, whether there's a new open model you've been wanting to try or a workload you're ready to run.

No new commitment, no setup required. Credits apply automatically to your account and are ready whenever you are.

Claim your $5 in credits, start building with Crusoe.

HARDWARE

AMD’s case for a more open AI stack

AI's appetite for compute keeps growing, but so does the pressure to deliver more intelligence per watt and per dollar. Can AMD's first rack-scale AI system open up an ecosystem dominated by Nvidia?

In this episode of The Deep View Conversations, we sit down with Andrew Dieckmann, AMD's general manager of its data center GPU business, to unpack the company's Helios platform and the rapidly changing economics of AI infrastructure.

Dieckmann explains why frontier AI requires more than just GPUs. It demands tightly engineered racks that combine GPUs, CPUs, networking, software, cooling and serviceability. The conversation examines the tension around AI data centers: hyperscalers still cannot get enough compute, while communities worry about power, water and whether the benefits justify the buildout. Andrew argues that responsible deployment and open ecosystems are essential as these systems become intelligence factories.

The conversation then turns to Helios: AMD's performance claims against Nvidia Vera Rubin, pricing and value, the first likely customers, and the Cerebras partnership for high-throughput, low-latency inference. Andrew closes with his advice for leaders navigating AI velocity: reassess priorities more often and use coding agents as force multipliers for scarce engineering talent.

Topics covered:

• Why AMD is moving from chips to full rack-scale systems
• AI demand, data center constraints, and community impact
• Open hardware, open software and customer choice
• How agentic AI changed infrastructure planning
• Helios performance, efficiency, pricing and customers
• AMD Helios versus Nvidia Vera Rubin
• How AMD and Cerebras split inference workloads

This conversation offers a clear look at the technology and economics shaping the infrastructure that will power everyday AI and the breakthroughs to come.

📺 Watch on YouTube

🎧 Listen in your favorite podcast player 

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

LINKS

• Rocket Money launches Rowan, a personal finance agent built with Claude

• A16Z raises $1.1 billion Machine Age AI infrastructure fund

• Meta tests ABB robots for data center cable and server tasks

• OpenAI will reportedly eventually build a humanoid robot 

• Anthropic raises standard weekly limits for Claude Code for paid users

• Sony, Warner sue Anthropic for intellectual property theft

• Expert Intelligence: Google launches ability to integrate ebooks in Gemini Notebook

• Hy4 preview: 770B, 49B active, 1M context, open-source model

• ChatGPT: Users can personalize a Temporary chat with existing memories

• GLM-5.3: Z.ai made it open-weight

• Innovaccer: Artificial Intelligence Researcher

• Rillet: Applied AI Engineer

• Socratix AI: Member of Technical Staff, AI Agents

• Promise: Software Engineer - AI

GAMES

POLL RESULTS

Do you think that Nvidia purchasing Hugging Face would be a good decision for the chipmaker?

Yes (54%)
Somewhat (22%)
No (18%)
Other (6%)

PRODUCTS

How Cohere's small model solves one big problem

Cohere is the latest lab to bet on the benefits of niche models. 

On Thursday, the company debuted Cohere Parse, a vision language model specifically designed to process large volumes of enterprise documents in a cost-effective manner. Parse can convert complex, multi-modal documents into machine-readable data that enterprises can leverage more easily for AI use-cases. 

Cohere's Parse closes a gap between expensive frontier vision models and less expensive (and less accurate) traditional optical character recognition models, the company said.

• Cohere Parse runs at $1.50 per 1,000 pages, compared to models from hyperscalers that cost roughly $10 per 1,000 pages.

• According to Cohere, Parse offers the best price-performance tradeoff of the models it evaluated, which included models from OpenAI, Anthropic, Databricks, Mistral, Microsoft, Google and Amazon. 

• The model is also available in nine languages, and will expand to meet customer demand. 

Kyle Lastovica, a spokesperson for Cohere, told The Deep View that Parse targets a very specific bottleneck for enterprises that's becoming critical in the age of AI adoption, especially as companies work to more deeply integrate agents and retrieval-augmented generation (RAG) into their workflows. 

"The quality of what goes into those systems matters more than ever, and enterprises are increasingly paying attention to AI spending," said Lastovica. "Standard optical character recognition misses key ingredients to document comprehensibility like tables, diagrams and visual structure. Many companies currently address this problem via frontier models, driving spend sharply upward."

Additionally, Lastovica said, the model can be deployed in a private cloud or in an on-premise environment with a small footprint, which is particularly important for customers in regulated industries like finance or healthcare.

Cohere Parse provides the latest of many examples of companies developing AI for specific tasks and seeing impressive results. Though document parsing may not be the flashiest or most exciting use case, it's a practical one that a lot of enterprises use AI for on a regular basis. It's the same thesis that drove NextLM's recent research in purpose-built, efficient, and affordable models for sales tasks. After months of shelling out on expensive token bills, some companies are becoming weary over bulky, frontier models and jack-of-all-trades AI. As a result, niche models, domain-specific models, and small models are gaining traction.

TOGETHER WITH CELONIS

Don't bet your business on a fortune cookie

Eliminate operational blind spots. Give AI the context it needs to succeed.

AI models are brilliant, but they’re probabilistic. Without a ground truth of your operational reality, trusting AI is like trusting a fortune cookie.

The Celonis Context Model provides this operational reality through a dynamic, real-time, system-agnostic digital twin of your operations. It translates your business into a language that AI understands.

Give your AI the context it needs to drive meaningful ROI. Meet the Context Model.

HARDWARE

Inside Agility's bet on factory-first humanoids

Humanoid robots continue to capture the imagination of humans, and the recent AI revolution has only accelerated the momentum. 

But despite demonstrations of robotics tidying up homes or awkwardly dancing up stages, there may be a better place to start: Factory floors, Pras Velagapudi, CTO of Agility Robotics, told The Deep View. With a significant labor shortage and an abundance of repetitive and dangerous tasks, manufacturing may be the perfect proving ground for humanoid robots. It's why Agility has focused on this niche, and already has robots deployed with companies such as Toyota, Amazon, and Mercado Libre. 

Velagapudi sat down with The Deep View in August at Ai4 in Las Vegas to discuss how humanoids can solve the "islands of automation," how these robots could shift the blue-collar labor force, and the path to robots eventually becoming household tools. This interview has been edited for brevity and clarity. 

Rubio-Licht: Something that I've been drilling down into the more I research physical AI is the humanoid form factor. Why do you think this is the right bet? 
Velagapudi: We have a long history of trying to get robots out in the world. Now, what's accelerated in the past few years is that we've kind of cracked all of the core technologies necessary to do this. We had been working for a long time on the physical hardware, the controls necessary to do that, and we had been making good progress and basically got all those pieces together. And then the physical AI boom happened, and has been carrying us on this wave of capability that, combined with all the hardware that we've been able to put together, is really making it so that we feel like this is a compelling time for this particular platform.

Rubio-Licht: And why are humanoids right for manufacturing versus, say, robotic arms?
Velagapudi: It isn't really humanoids versus robotic arms. It's more that manufacturing is a huge industry. One of the reasons we're going after it is because it's such a huge industry that even with robot arms and (Autonomous Mobile Robots), ground robots moving around and welding robots and all of these different pieces, there's still so much that hasn't been automated yet. What we call that is islands of automation. That there are these pieces of automation that are then separated by some sort of manual process, and then another piece of automation. But there are these little islands that exist that need to be connected by some manual process. And so there's still so much of that that's been difficult to automate because it has the wrong structure for a robot arm, or it has to occur in many different places in a facility. 

Rubio-Licht: How do you think robotics are going to eventually impact the labor force? 
Velagapudi: What we're taking on are really the dull, dirty, dangerous jobs — the 3D's of robotics — and we're very focused on areas like bulk material handling, picking up 50-pound totes, putting them into loading and unloading machinery. This is in hot environments, in dangerous environments where there's risk of repetitive strain, where you're working around other machinery. So what we're doing with these deployments is really trying to get those jobs filled — the ones that are this type of manual labor that's not really well-suited to humans.

Rubio-Licht: Do you anticipate Agility expanding into the household domain? 
Velagapudi: It's definitely in the future. We do want to get there. We do think that one of the great untapped markets for humanoid robots is closest to humans, like in homes, in hospitals, care facilities, things like that. It's where there's a lot of value because you have the most human-centric environment. There, you have the benefit of creating a robot that can do all the things in your environment that you do in your environment, but the road there to do it safely is going to be a bit longer. 

Click here to read the full interview on The Deep View.

TOGETHER WITH CELONIS

Context doesn't just inform AI. It puts AI to work.

Ask a generic AI model about outstanding invoices, and you'll get a well-worded, generic answer. Ask Enterprise AI grounded in your operational reality, and it knows which invoices are overdue, why and what to do next – then acts on it across your systems.

That's what the Celonis Context Model makes possible: a living digital twin of your business that gives AI the operational clarity to decide and act, not just answer.

No more isolated chatbots guessing at outcomes. Just decisions, grounded in reality, that move through your operations on their own.

Meet the Context Model.

PRODUCTS

As models commoditize, Replit embraces routing

State-of-the-art AI is a moving target. Replit wants to move with it. 

On Thursday, the AI coding platform launched Intelligent Model Routing as the default for all accounts. This system matches tasks to the model they're best suited for, aiming to balance quality, speed and cost. 

In its blog post, Replit said that Intelligent Model Routing offers power when a task demands it and efficiency when it doesn't. 

• And the company has already seen significant results: The company said that Intelligent Model Routing offered the same output quality as its previous version of Max Mode, the company's top-tier, paid setting that uses frontier models, at a 65% lower cost. 

• Additionally, users will still have transparency about when they are using certain models: Users in Free Mode, its unpaid tier, will be notified when Replit wants to escalate their work to higher-powered models that can incur usage costs, and can choose to override this. Paid users can also still manually select models.

• And for enterprises that want more control over which models their teams use, administrators can define an approved set of models, and Replit's model routing service will choose from that set for a given task. 

"Users can focus on how AI can augment their work, instead of model management," the company said in its blog post. 

Replit's offering follows a similar development from Snowflake, which debuted its own dynamic model routing solution within Cortex AI Gateway and its flagship AI products, Snowflake CoCo and Snowflake CoWork, in mid-August. 

Both of these firms target an increasingly popular niche in enterprise AI: Companies are growing more interested in picking the right model for the job, rather than the most powerful model on the market.

It's clear that the AI tides have shifted in recent months away from power-at-all-costs towards efficiency. This shift is showing up in numerous ways, including model routing services like Replit's, growing research and offerings around task-specific and niche models, and even new architectures that make small models as powerful as trillion-parameter ones. This change, however, leaves me with one major question: If the major model labs become commoditized, what happens to the economics of an industry built on promises of perpetually growing AI spend?

MARKETS

Nvidia eyes Hugging Face in full-stack AI push

Nvidia may be taking one step closer to dominating the open-source AI ecosystem. 

The tech giant has agreed to buy Hugging Face, the dominant repository of open-source AI models, for $12.9 billion, according to a report by The Information, citing a source familiar with the agreement. This is a strategic move that allows Nvidia to diversify its portfolio from just chipsets and take ownership of more of the AI stack.

This would give Nvidia ownership over the "front door for open AI innovation," Ashish Nadkarni, Group VP of enterprise infrastructure at IDC, told The Deep View. "Owning that front door gives NVIDIA a major position in the mindshare of today's AI development personas."

The deal comes as Nvidia faces mounting pressure from the very labs that rely on its chips. Google, OpenAI, and Anthropic are each moving toward custom or co-designed AI silicon, a shift that threatens Nvidia's near-monopoly on AI hardware. Worse for Nvidia, these labs bring something it can't easily replicate: deep, firsthand knowledge of the models the chips are meant to run, letting them tune hardware for performance in ways a general-purpose chipmaker can't match. 

Hugging Face could eventually become a non-chip source of revenue for Nvidia, which could spur more growth if chip revenue eventually stops skyrocketing. A small caveat, however, is that, according to the report, Hugging Face has only generated  $150 million in annualized revenue thus far. That makes Nvidia's payment about 80 times the startup's forward revenue, the report notes. 

Revenue aside, the advantage of being both a model and chip maker is a key reason Nvidia has invested heavily in open-source models. Nvidia's commitment has been reflected not only in signing an open letter supporting the ecosystem, but also in internal investment in and development of open-source models, such as its Nemotron models. Of course, the Hugging Face acquisition moves that forward.

Additionally, there's a significant robotics angle: Nvidia already has a strong position with its Jetson, Isaac, and Omniverse solutions, and Hugging Face's own robotics efforts give Nvidia another point of leverage as physical AI becomes the next frontier. And just today, Hugging Face unveiled Microduck, a $399 open-source robot that learns new tricks via reinforcement learning. Even its looks resemble the Nvidia "Blue" research robot, inspired by Star Wars' BD-1, shown at demos, live events, and Disney. 

Still, as Nadkarni highlights, the acquisition would also mean there would be less choice in the industry. 

"To the extent that others like Intel, Apple, and AMD make similar plays, there will be fewer “ecosystem-neutral” platforms in the market," said Nadkarni. "Note how Microsoft has gradually made GitHub more like an extension of its own strategy (and for its own benefit)."

This is a landmark move for Nvidia, as it solidifies its efforts to be not just a chip company, but one that aims to own multiple layers of the AI stack at once. Hugging Face has become something like the GitHub of AI: serving as the industry's central hub for sharing models, datasets, and applications. Owning that layer, on top of the hardware layer Nvidia already dominates, deepens the moat. Already, nearly every open model runs on Nvidia GPUs, and buying the platform where those models live removes any ambiguity. It's also a defensive play: as OpenAI, Anthropic, Google, and Amazon push to build their own custom chips, controlling the distribution layer keeps developers tied to Nvidia's ecosystem regardless of who's building the silicon.

TOGETHER WITH DESCOPE

Run your auth with AI agents

Coding agents are here to stay, but vibe-coding auth is dangerous business. Connect your AI agents to the Descope MCP server instead! 

This remote MCP server gives agents the ability to read documentation, manage users and tenants, configure authentication flows, generate FGA schemas, review audit logs, and more…all through natural language. 

Get started on Claude or ChatGPT.

GOVERNANCE

Can new AI coalition prevent cybersecurity crisis?

The AI industry is sounding the alarm on a problem that it helped create. 

On Thursday, a coalition of more than 100 tech companies, including OpenAI, Anthropic, Google, Microsoft and more, published a letter calling for collective action to strengthen cyber defenses as AI-enabled cyberattacks pose a growing threat. 

"Today’s AI advances are already giving defenders new ways to fix weaknesses that have accumulated for years," said the letter, which was published on OpenAI's blog. "If we act decisively, we can use the defenders’ window to make our digital world much more secure."

Twitter tweet

The letter recommends three broad principles for immediate defense: recognizing that the status quo isn't good enough, empowering cyber defenders with AI, and mobilizing a global collective response. 

It also includes a list of suggestions across enterprises, governments and cybersecurity firms to protect against risk:

• All organizations should make cyber defense an "immediate leadership priority" by fixing high-risk weaknesses and raising security standards and meeting them with "urgency and coordination of an incident that takes precedence over everything except critical business operations." 

• Cybersecurity and tech companies should help lead the charge to defend against AI-enabled attacks and use AI to strengthen existing methods of defense. 

• Governments should coordinate defense at the local, national, and international levels, strengthen existing government defenses, fund response programs, and expedite the expansion of trusted access programs. 

• Meanwhile, frontier AI companies should provide responsible model access, as well as "significant funding, training, and hands-on support" for infrastructure defenders. Additionally, they should build more observability tools to keep their models and agents traceable and accountable. 

The letter comes as frontier AI companies continue to grapple with the capabilities of their most powerful models, especially as these models prove their power through incidents such as OpenAI's breach of Hugging Face. These models have grown so powerful that OpenAI has even opted to slow down development of its next-gen model, Astra, announcing earlier this month that it put its largest planned frontier reinforcement learning run "on hold" as it conducts smaller evaluations to test model behavior.

OpenAI President Greg Brockman addressed the shrinking window for cyberdefense, calling the breach of Hugging Face a "watershed moment for cybersecurity" in a recent essay. "I’ve spoken with many organizations over the past few weeks, and one theme is clear: they know they need to fundamentally uplevel their cybersecurity practices with unprecedented speed," he wrote.

This letter finally says the quiet part out loud: no enterprise, AI lab, tech giant or government is fundamentally ready for the cybersecurity shift that frontier AI capabilities are forcing upon us. However, as it stands, the solution may have to be more than playing catch-up in defense. While AI gives defenders more capabilities, threat actors tend to have a 12-foot ladder for every 10-foot wall. OpenAI did a commendable thing in tugging the reins on Astra, even if it's only temporary. It's still an open question whether other labs will follow its lead and slow the development of their next models to allow cyber defense solutions to catch up.

TOGETHER WITH JUMPCLOUD

Your AI agents are getting more access. Humans are getting less oversight.

AI agents are moving deeper into critical workflows, and the guardrails are loosening. Research from JumpCloud found that AI agents allowed to take high-risk actions without human review more than doubled in six months, from 11% to 26%.

Giving an agent an identity is only the start. IT also needs to control what it can access, what actions it can take, and when a human needs to step in.

See how to secure every identity, human or not.

MARKETS

The economics pushing AI toward open models

AI's next power shift isn't gonna happen in a data center. 

In this episode of The Deep View Conversations, we sat down with Jeff Morgan, co-founder and CEO of Ollama, to explore why open models are gaining momentum, and why enterprises and developers increasingly want more control over their AI.

Morgan explains how Ollama grew from a two-week experiment into software used across 80% of the Fortune 500, how the economics of coding agents are pushing teams toward open models, and why cost, privacy and control are becoming decisive advantages. He also breaks down the hardware shift bringing data-center-class AI workloads to Apple silicon, Nvidia DGX Spark and systems powered by AMD, Intel and Qualcomm.

The conversation also covers:

• How the team behind Docker Desktop came to build Ollama
• Why open models could soon process the majority of enterprise AI tokens
• The role of harnesses, tool calling, routing and subagents
• How Ollama fits into the open-source AI stack and where its business model comes in
• Why new US and European open-model labs are emerging
• Why companies may need to own and customize their intelligence layer

If you’re interested in open models, coding agents, enterprise AI or the shift from cloud-only AI to powerful local systems, this conversation offers a clear look at where the ecosystem is heading.

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

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

And don't forget to sign up for The Deep View daily newsletter. We don’t just cover AI, we decode it. In a world flooded with hype, we deliver sharp, no-nonsense insights to keep you ahead of the curve and help you put AI to work every day: subscribe.thedeepview.com

LINKS

• OpenAI to show ads on ChatGPT Free and Go tiers in India

• Plaud Inc. unveils 4G earbuds with on-the-go AI recording

• Meta internally projected spending up to $10 billion annually on Anthropic

• OpenAI published full technical report on the Hugging Face incident

• Cisco launches MyAgent AI assistant to 90,000 employees

• Experian launches credit card marketplace app within ChatGPT

• Claude: now has its own built-in browser in Cowork

• MiniMax H3 Max: new post-trained video model by fal Research

• Google Antigravity: new Interactive Generative UI Artifacts; sometimes markdown, images, and Mermaid diagrams aren't enough for the agent to fully express an idea

• ChatGPT scheduled tasks: Updates came to the feature in ChatGPT Work on web and mobile

• Gemini Omni 1.1 Flash: a new suite of creative controls and generative video capabilities to support developers

• ROIAI One: Founding Member of Technical Staff

• Code Red Partner: Artificial Intelligence Engineer

• AXION: ML Researcher

• Ivo: AI Researcher

GAMES

A QUICK POLL BEFORE YOU GO

MARKETS

What Z.ai's Ox Alpha reveals about AI economics

Chinese AI lab Z.ai has come forward to claim the viral anonymous Ox Alpha model. 

On Wednesday, Z.ai unveiled GLM-5.3-Flash, the first natively multimodal model of the GLM-5 series. The company said the model was released anonymously as Ox Alpha on OpenCode and OpenRouter where it completely overtook leaderboards and went viral for offering a capacity for 100 trillion tokens per day. 

Notably, the company said in its announcement that all of the traffic from its model's skyrocketing popularity was "served on Chinese AI chips." 

The model features 320 billion total parameters with 18 million active, and outperforms its previous generation, GLM-5.2, across a number of benchmarks at a tenth of the price, the company said in its announcement. 

The lab also claims that GLM-5.3-Flash approaches Claude Opus 4.8 in coding and agentic benchmarks. It also performs on par with DeepSeek-V4, GPT-5.6 Terra, and Gemini 3.7 Flash on benchmarks for software engineering, multi-step agentic tasks, tool use and professional work. 

The company said the model is architected for "extreme efficiency," and "specifically designed for ultra-low-cost inference." 

• GLM-5.3-Flash was essentially designed to do more with less, the company said, using a hybrid attention architecture to reduce the cost of serving long-context queries without sacrificing accuracy. 

• Additionally, the model is built to improve "scaling efficiency" by implementing what are called "Manifold-Constrained Hyper-Connections," or a technique that improves the way that information moves through the layers of a neural network. 

• The model weights are currently available on Hugging Face. The model is also powering ZCode, Z.ai's coding tool. 

"GLM-5.3-Flash shows that frontier intelligence does not have to come at frontier cost," the company said in its announcement. "We are now scaling this recipe to larger models — GLM-5.3-Flash pushes the cost-performance frontier, and the lessons from building it are already shaping our next frontier model." 

Z.ai's Ox Alpha win comes at a particularly poignant moment for open models: Open source model platform Hugging Face is reportedly fielding acquisition offers that value the company at $13 billion, and Nvidia last week announced a $6 billion investment in bolstering the open source AI ecosystem in the US. 

But Chinese firms may be going after one enterprise pain point in particular: token costs. Models from Chinese labs that perform on-par with those from proprietary providers may be enticing to developers that are now facing budget constraints after previously being encouraged to spend freely on AI, Karthik Sj, chief AI officer at LogicMonitor, told The Deep View. 

"Ox Alpha demonstrates how low- and zero-cost model access may sound appealing at first, but free access always carries hidden costs," said Sj. "While Z.ai revealing that it’s behind the model solves some auditability concerns, vetted and well-tested models are the safer bet.”

Though it's obvious that Z.ai is targeting cost and resource efficiency with its release of GLM-5.3-Flash, this model's popularity also points to another trend: not every daily use model needs to be state-of-the-art. GLM-5.3-Flash still sits on-par or just behind the frontier models from labs like Anthropic, Google and OpenAI, however, a model doesn't need to be ultrapowerful to be incredibly useful. Though OpenAI is trying to entice people to use its frontier models with lower prices, it can only serve that kind of inference at those costs for so long. For a large majority of enterprise tasks, generally, mid-range performance is good enough, especially when it comes to solving for ROI. 

TOGETHER WITH CRUSOE

$5 in free credits, on us.

Every Crusoe account now comes with $5 in free credits for Crusoe Intelligence Foundry. Use them for Serverless Inference or Serverless Fine-Tuning, whether there's a new open model you've been wanting to try or a workload you're ready to run.

No new commitment, no setup required. Credits apply automatically to your account and are ready whenever you are.

Claim your $5 in credits, start building with Crusoe.

ENTERPRISE

Anthropic-Salesforce deal could break AI silos

As Anthropic cements its place as a go-to AI provider for enterprises around the world, the company is deepening its ties with another enterprise heavyweight: Salesforce.

On Wednesday, Salesforce and Anthropic unveiled Claudeforce, the name given to the companies' expanded strategic partnership. Claudeforce combines Claude's intelligence with Salesforce's enterprise platform to offer agentic experiences for working professionals. 

The selling point is adding Claude AI's reasoning layer on top of the Salesforce platform, giving the AI access to the workflows, data, and insights that companies already have stored in Salesforce. 

• The partnership is launching with Salesforce in Claude, a plugin that includes 37 pre-built sales skills, aimed at helping both agents and sellers apply AI optimization to their workflows directly in Claude, using context from Salesforce, according to the release. 

• The skills were built by both Anthropic and Salesforce to take it one step beyond "generic (customer relationship management) prompts wrapped around an API," to better fit what Claude can do. 

• Some examples provided include reasoning over live revenue context or automating pipeline updates. 

The setup is meant to be simple: An admin just has to connect Salesforce to Claude once, with authentication and permissions managed centrally, and every seller on the team gets access, circumventing the need for per-user access, according to Salesforce. Salesforce in Claude is available to select pilot customers and will launch in open beta in September 2026. 

Claude is also available inside Salesforce, including Agentforce, where it serves as a

reasoning model for the Atlas Reasoning Engine, as well as powers Agentforce Vibes and Agentforce Coworker by default. Through Amazon Bedrock, users can also access Claude within the Salesforce Trust Boundary. 

Finally, through the partnership, Claude will become the default model for Slack, powering  every instance of AI in Slack, including Slackbot and the @Claude tag, and is a founding partner for Slack Code. The companies noted that more integrations are coming across Claude, Salesforce, and Slack, with additional pre-built skills launching later in 2026.

Salesforce and Anthropic both emphasized in the announcement that the agreement involves reciprocal adoption, with each company becoming a customer of the other. While this is partly marketing rhetoric, I see genuine value in the arrangement: it helps both companies build solutions their customers actually want, since each experiences the other's product from a user's perspective. More broadly, I think the AI industry would benefit from greater collaboration between companies, rather than siloed development aimed purely at competing with one another, an approach that often fails to leverage each company's strengths.

TOGETHER WITH CELONIS

Don't bet your business on a fortune cookie

Eliminate operational blind spots. Give AI the context it needs to succeed.

AI models are brilliant, but they’re probabilistic. Without a ground truth of your operational reality, trusting AI is like trusting a fortune cookie.

The Celonis Context Model provides this operational reality through a dynamic, real-time, system-agnostic digital twin of your operations. It translates your business into a language that AI understands.

Give your AI the context it needs to drive meaningful ROI. Meet the Context Model.

PRODUCTS

Why Google's Rambler could come to every device

One of the most talked-about Google AI features from the Pixel 11 launch was Rambler, its highly effective speech-to-text feature. Now Google is spreading the tech that powers it. 

On Wednesday, Google introduced Gemini 3.5 Transcribe, which the company describes as its most precise speech-to-text model yet, and capable of converting raw audio into accurate, polished text. Google shared that this is the model that powers Rambler on Android and the Gemini app on macOS, and it invites developers to build similar capabilities with the model. 

Twitter tweet

Moreover, Gemini 3.5 Transcribe is available across two separate APIs: gemini-3.5-transcribe-live, which is focused on continuous bidirectional real-time interactions with sub-second latency and gemini-3.5-transcribe, which is focused on pre-recorded audio processing and transcribes recorded audio, meetings, and call logs with speaker attribution and word-level timestamps.

What makes the model and the Rambler experience stand out is how accurately it can capture your ideas, even when they aren't presented linearly. For instance, during a correction, when one says "actually never mind how about…," it should not account for that; same with "umhs" and "ahs." Other features include:

• The model can delegate complex tasks to other Gemini models via function calls

• It recognizes jargon and unique spellings unique to the speaker

• It automatically transcribes over 85 languages 

This progress is reflected in benchmarks, with Gemini 3.5 Transcribe outperforming its previous transcription model, Chip 3, by 70% on time to final transcription as measured by the Artificial Analysis benchmarks and on the FLEURS benchmark, which measures multilingual performance. 

Anyone can experience the model today on Android via Gboard's new Rambler feature, in the Gemini app on macOS, and coming soon to Chrome. Developers can access it on Google Antigravity in Google AI Studio.

The most interesting part of this experience is that it will enable other developers to create experiences that perform as well as Rambler or even outperform it with post-training. The biggest competitor to Rambler at the moment is Wispr Flow, which has been dominating the market for months by offering the same thing to users across both Apple and Android devices, but it;s not nearly as seamless or integrated on mobile. With developers able to build similar experiences, Wispr Flow is going to face a flurry of competition, which has me wondering whether it will even be able to retain its place in the market. But the Google model is also exciting because, with so many more developers working on the same foundation, it should unlock even more features that make the model stronger. This is something to look forward to because it already works better than everything else other than Wispr Flow. In our recent podcast episode, we said that Rambler is so good that every other smartphone is likely to have a similar feature within the next two years. It feels like Google just made that a lot easier.

LINKS

• OpenAI's Sam Altman says he expects AGI by end of the year

• Anthropic, Nscale ink $45 billion deal for data center power

• OpenAI publishes technical report on Hugging Face incident

• Meta AI staffing plan reportedly collapsed after employee revolt

• Google expands billing flexibility for Gemini Enterprise, Android Studio

• Bain joins Anthropic Claude Partner Network as global premier partner

• ChatGPT Business: Premium seats are now available 

• Qwen3.8-Flash: Alibaba releases smaller Qwen model

• Claude: now has one memory across chat and Claude Cowork

• Putty: Google Labs' collaborative vibe coding tool that lets you build tools and websites

• Alexander Chapman: AI Systems Engineer

• Noscen: Founding AI Engineer

• Code Red Partners: Artificial Intelligence Engineer

• Fudo Partners: Artificial Intelligence Engineer

GAMES

POLL RESULTS

Have you ever used Perplexity?

Yes (55%)
No (40%)
Other (5%)

HARDWARE

OpenAI Jalapeno is even bigger than the benchmarks

OpenAI has built its first AI chip, and the early returns on performance look impressive.

On Tuesday at the Hot Chips conference, OpenAI said that its custom inference chip, Jalapeño, is now up and running on working silicon and is already in testing. The frontier lab revealed the first third-party benchmarks and claims that it's accelerating performance-per-watt, the favorite measurement of efficiency in the AI industry. It's basically a way of showing how much AI work a chip can complete for every unit of power it consumes. 

On tests against various types of workloads, OpenAI claims that Jalapeño delivers 1.5 to 4 times the performance-per-watt compared to today's leading AI chips. 

OpenAI used SemiAnalysis's open-source InferenceX benchmark to test Jalapeño, using three models: GPT-OSS (OpenAI's open-source model), DeepSeek R1, and Kimi K2.5. In other words, it didn't just test Jalapeño on its own models fine-tuned to work on its own hardware. 

And while its own GPT-OSS model performed the best, the other models also performed remarkably well. The biggest surprise was that Jalapeño outperformed Nvidia's Blackwell, one of the current industry standards for AI chips. However, we don't yet know how it will perform against other inference accelerator leaders like Groq and Cerebres.

SemiAnalysis characterized it this way: "Jalapeño beats Blackwell on perf/W across almost all scenarios without being tuned for any specific point in the curve. It excels not only in low-latency scenarios but also in high-throughput scenarios."

In a briefing with the media, I asked OpenAI's head of hardware, Richard Ho, where OpenAI would use this newfound power in its AI stack (because obviously it won't be running all workloads on its own chips anytime soon). "It's clear that the cost of serving low latency with this device is a lot lower than other low latency devices," said Ho, "and so there may be a good niche there where some of our very latency sensitive products such as Codex Ultra Fast Mode — even faster than the [fastest mode] we have now — may be a good option… We are expecting to ramp up the volume into 2027."

Other than the performance itself, the most impressive thing about Jalapeño may be how fast OpenAI was able to develop it and bring it into production. The project, which is a partnership with Broadcom, began as a concept in late 2024. The fact that it already has working silicon less than two years later is just as surprising as the benchmark performance. OpenAI credits its own AI models to helping the hardware team dramatically accelerate the process. The team reported that OpenAI's models aided in the optimization, programming, and development of the chips and are currently doing the same for the next generations that come after Jalapeño. While the speed is great and will help meet user expectations over time, the real win for OpenAI is in efficiency. If it can design and optimize its own chips then it can spend far less on compute and lower the long-term cost of intelligence to bring its most advanced models and features to more users.

TOGETHER WITH QUIQ

7 AI agents, one seamless journey

Most companies still think of AI as a chatbot for support questions. But 72% of buyers now expect a personalized experience at every touchpoint, and companies deploying agentic AI across the full customer journey are seeing up to a 30% increase in conversion rates, according to McKinsey research.

In Quiq’s new guide, “The Clock Is Ticking: 7 AI Agents Every Leader Needs,” we break down the seven essential agents, from the conversational ad that first engages a customer to the AI analyst that catches churn risk after every interaction, and where each one fits across the journey.

Get your free copy of this guide here.

PRODUCTS

Apple's new M5, M6 Macs reveal local AI strategy

The Mac mini has helped fuel the AI agent revolution. Now Apple is ready to extend that momentum with its next-generation model.

On Tuesday, Apple unveiled its latest chips, the M6 in the new Mac mini and the M5 Ultra in the new Mac Studio, which the company is positioning as AI powerhouses. This launch follows months of Mac mini shortages earlier in 2026, when demand, driven largely by AI enthusiasts running OpenClaw and other agents as persistent, always-on machines, outstripped Apple's supply, compounded by a broader industry-wide shortage in memory and storage.

The Mac mini, the popular go-to for local and private AI inference on a budget, is now available with the all-new M6, which Apple says delivers up to 4x faster AI performance, 2x faster storage and graphics, and 40% faster CPU performance while retaining the same attractive, compact form factor it is known for. The M6 chipset features a 12-core CPU and 12-core GPU, two more cores than before in each, with the GPU adding Neural Accelerators to each core for the first time on Mac mini, powering many of these improvements. 

Twitter tweet

The M5 Pro version is meant to take it up a notch, with pro-level performance for more power-intensive applications such as video production and gaming. The Mac mini with M6 starts at $899, and the Mac mini with M5 Pro starts at $1,699, both $100 more than their predecessors. 

"With a more powerful CPU and graphics, Neural Accelerators in the GPU, and higher memory bandwidth, Mac mini with M6 delivers a whole new level of AI performance," said Johny Srouji, Apple’s chief hardware officer, in a blog post. 

For those who require even more power, Apple's Mac Studio also got 4.3x faster AI performance, according to Apple. Other upgrades include up to 2x faster storage, up to 1.8x faster graphics, and up to 1.3x faster CPU speed. The new M5 Ultra is Apple's most powerful silicon yet, featuring an 18-core CPU, up to 40-core GPU with Neural Accelerators built into each core, and up to 128GB of unified memory. 

The Mac Studio with M5 Max highlights include an 18-core CPU, a GPU with up to 40 cores, each with Neural Accelerators built into each core, and up to 128GB of unified memory. The Mac Studio with M5 Max starts at $2,499, and the Mac Studio with M5 Ultra starts at $5,499. 

Pre-orders for all the new Macs start today, and shipping starts on Sept 22, the same day the new Macs arrive in Apple stores.

Apple's Mac mini and Mac Studio have always been geared towards power users, often led by creative professionals, and that's typically the language Apple would use at launch to describe the product to the market. It is interesting how much has changed, with so much of the messaging geared towards AI and explaining the improvements made to better power those workflows. It is also interesting to see how the explosion of AI workflows has shifted demand for these products, as the Mac Studio blog post even includes a paragraph explaining how people could use Mac Studios with the M5 Ultra chip in clusters to take their local AI inference to the next level. This is a very purposeful AI launch ahead of what will likely be Apple's biggest iPhone launch in years this September, with Siri 2.0 finally going live, its new foldable phone expected to launch, and the possibility of its first AI hardware products with the AI AirPods with cameras and/or a desktop AI speaker.

TOGETHER WITH DESCOPE

The biggest MCP spec update since June 2025 just landed

Every team shipping MCP servers should pay attention to the July 2026 spec revision. Sessions are gone. DCR is deprecated in favor of CIMD. And there are six new authorization SEPs your server is now expected to handle.

This developer breakdown from Descope covers:

• What changed in the transport layer, and why sessions were removed

• The six new authorization SEPs explained

• Enterprise-Managed Authorization (EMA) and what it means for consent sprawl

• A migration checklist to run against your existing MCP server

Read the breakdown before your next MCP deployment.

PRODUCTS

Why AI agents are moving to your desktop

Perplexity is doing what it does best. Releasing something that feels like others could soon copy. 

But this time, Nvidia is joining in as a partner. 

On Tuesday, Perplexity announced Portable Computer, a version of its AI agent that can run open models locally on your own desktop hardware to do three things: 

1. Save significant money on tokens

2. Get faster performance

3. Boost data privacy

That's the kind of stuff more and more people have been trying to do because of soaring token bills due to AI agents and because of wanting to run agents on their most sensitive data, which they'd prefer to run locally and not send to the latest frontier models in the cloud from Anthropic, OpenAI, Google, and others. 

Where Nvidia comes in is that Portable Computer will run on Nvidia's DGX Spark machine, which has cult favorite status among AI builders. It's a tiny box the size of Mac mini, but with the power of a Mac Studio for running AI. Spark has 128GB of unified memory and runs up to a petaflop of compute. Nvidia says it runs models up to 200B parameters and can fine-tune models up to 70B.

Perplexity's Personal Computer ran on a Mac mini at launch, so a big part of the announcement is Perplexity's AI agent now running on PC hardware powered by Nvidia GPUs. That means it's likely to come to other DGX Spark competitors like the ones from Dell, Lenovo, Asus, MSI, and others. Right now, the software only runs on Linux. But Perplexity is also working on a version that will run on Windows, as well as a version that will run on Nvidia's more powerful version of DGX Spark called DGX Station, which is the size of a full computer tower like a Mac Pro.

The other big part of the announcement, of course, is being able to run Perplexity's agent on open models that run locally on your machine. At launch, Portable Computer will run a post-trained version of Alibaba's Qwen 3.8 27B called PPLX 27B, with a version of Nvidia's Nemotron 3.5 Lightning coming to the device soon, according to Perplexity. You can also use a local inference server like Ollama or LM Studio to run any open model you'd like, but that starts to get a little more complicated.

The problem that Perplexity and Nvidia are trying to solve is that more and more people are trying their hand at AI agents and would like to save the token costs and get the privacy you can get by running these models locally, but it's a very involved process that requires time and technical expertise.

"Now that we're seeing people actually run frontier intelligence on their desk, the remaining bottleneck becomes that it's quite cumbersome to set up," said Nate Kupp, VP of infrastructure at Perplexity, in a briefing with the media. "We really focused on making this a straightforward experience where you can get up and running very quickly."

While DGX Spark and competitors cost $4,000-$5,000, they can save you so much in token costs that AI builders don't even blink at that price. Heavy AI agent users can spend up to twice that a month in token costs. And if Portable Computer catches on, and other AI companies offer something similar, then it could increase competition and drive down the price tag.

Perplexity's best attribute is arguably how fast it can execute. It's small and nimble and, again and again, the team has shown the ability to launch things quickly. Its AI agent was an idea that was hatched in about a month and launched around the same time OpenClaw went viral in early 2026. Perplexity's general purpose agent can now help you code your own software like Claude Code or OpenAI's Codex, but it can also help you carry out knowledge work tasks the same way Claude Cowork and ChatGPT Work can. The ability to run locally on open models is a super power. But even highly technical people have complained about how involved the setup is to get Nvidia's DGX Spark up and running, so this would be a win for both Nvidia and AI enthusiasts if Perplexity can make that process more streamlined. 

LINKS

• OpenAI bans Russian ChatGPT accounts used in influence campaign

• Self-driving firm Gatik raises $200 million two months after PepsiCo deal

• Waymo plans driverless ride-hailing in Germany by end of next year

• Infrastructure firm Emerald AI raises $150 million at $1.05 billion valuation

• Mistral and HUMAIN partner on AI infrastructure, models in Saudi Arabia

• EPA aims to curb public input on AI data center buildout

• Grok Voice Think Fast 2.0: Now #1 on the Artificial Analysis Speech-to-Speech Index.

• Composer: a section-by-section song editor in ElevenLabs

• Qwen3.8-27B: landed in Code Arena WebDev #9; only model in its size class in top 10

• MiniMax M3 and M2.7: From Aug 24–Sep 6, get unlimited access

• Millennium: AI Engineer

• Smartish: AI Systems & Operations Specialist

• NVIDIA: Solutions Architect, AI Models

• Further: AI Strategist (ATL)

GAMES

A QUICK POLL BEFORE YOU GO

MARKETS

Why OpenAI is resetting frontier AI prices

OpenAI is usually the AI industry's trendsetter. Its latest focus is all about making tokens more affordable. 

Last week, the company announced that it would cut prices of GPT-5.6 Sol, its highest performing model, by more than 20% for developers for the next three months. The model is now priced at $4 per million ​input tokens and $20 per million output tokens for ‌standard ⁠short-context use, compared to previous pricing of $5 per 1 million ​input tokens and $30 per 1 million output tokens. 

It's also far below the costs of Fable 5, the most powerful model from rival Anthropic, which runs at $10 ​per ⁠million input tokens and $50 per million output tokens. Sol is now also cheaper than Opus 5, Anthropic's most recent model, which comes out to $5 per ​million input tokens and $25 per million output tokens.

Twitter tweet

It's not the first time OpenAI has slashed prices for its models. Last month, the company chopped prices for GPT-5.6 Terra, its mid-tier model, and GPT-5.6 Luna, its lightweight, efficient model, by 20% and 80%, respectively. 

Though these price cuts help it stay competitive with Anthropic, they also may be edging the company beyond cheaper Chinese alternatives: In a post on X, Tomasz Tunguz, venture capitalist at VC firm Theory, noted that OpenAI's Luna is price competitive with DeepSeek's latest model intelligence-to-dollar. 

Additionally, according to benchmark platform Arena, Sol's price decrease has improved its performance in agentic benchmarks for coding and work.  

These price cuts come as enterprises begin to seriously question their AI spending, with agents poised to increase inference costs up to fivefold, according to recent Gartner research. And the evidence is already starting to pile up that companies are shifting away from implementing frontier AI at any cost: 

• In mid-August, data from payments platform Ramp indicated that most of business spending on Anthropic models wasn't on its ultrapowerful model Fable 5, comprising only 6% of token spend. 

• For comparison, OpenAI's GPT-5.6 Sol comprised 25% of tokens used by OpenAI Ramp customers. 

• Meanwhile, Anthropic's less powerful Opus 5 has surpassed Fable 5 in business spending.

Though OpenAI tends to be one of the industry's tastemakers, it may also be reading the tea leaves from the fact that enterprises are pulling back spending on powerful frontier models, and are largely comfortable settling for efficiency over capability for many tasks. It's why small models, domain-specific models, and open-source alternatives are gaining so much traction. However, as getting enterprises to adopt frontier intelligence is a major part of the company's strategy, these discounts might be a way to get people to see beyond the price tag. Even if these companies don't need Thor's hammer to kill a fly, by selling Thor's hammer on the cheap, it seems that OpenAI wants them to give it a shot anyway. Plus, with OpenAI putting its best models on sale, perhaps Anthropic will follow in its footsteps. 

TOGETHER WITH CELONIS

Don't bet your business on a fortune cookie

Eliminate operational blind spots. Give AI the context it needs to succeed.

AI models are brilliant, but they’re probabilistic. Without a ground truth of your operational reality, trusting AI is like trusting a fortune cookie.

The Celonis Context Model provides this operational reality through a dynamic, real-time, system-agnostic digital twin of your operations. It translates your business into a language that AI understands.

Give your AI the context it needs to drive meaningful ROI. Meet the Context Model.

RESEARCH

What America’s most AI-exposed cities reveal

Across the country, workers are watching AI-powered automation creep into their industries and wondering if and when their jobs could be in danger. But where you live may determine how much you have to fear.

Indeed developed a metric that measures how exposed each metro's job composition is to AI by combining the jobs each metro advertises with the AI exposure of those skills, using data from both the Indeed Hiring Lab's 2025 AI at Work Report and its GenAI Skill Transformation Index. The findings, unsurprisingly, showed that metros built around tech and knowledge work are the most vulnerable. 

Across US metro areas, the average exposure score was 44, with a range of 40 to 60. The five most-exposed metros are driven by software and data work: 

• San Jose, CA: 59 

• Seattle, WA: 57

• Washington, DC: 54

• San Francisco, CA: 53 

• Austin, TX: 52

On the other hand, metros whose economies lean into hands-on work were less exposed. This supports the same conclusion drawn by many job reports and experts alike: labor that relies on physical work is the safest type of job in the AI era. 

As we covered last week at The Deep View, this is already influencing the next generation of the labor force, with Pew Research Center finding that, for the first time, a majority of adults under 30 (55%) say they’re more concerned than excited about AI, and other research showing that Gen Z is more open to careers in trades.

Because the Indeed report findings were based on data examining how AI could change the way certain skills are performed, it is important to note that the jobs most exposed don't necessarily equate to the jobs that will be replaced by AI, but rather to those with the highest potential to be transformed by AI. How that exposure translates into job loss depends entirely on how the employer chooses to adopt the technology. 

With that in mind, An Nguyen, an economist at Indeed Hiring Lab, has advice: 

"For job seekers, it can be a starting point for which skills to build. A high exposure score means more of the work could be reshaped, or augmented, by AI, not that the job will go away," Nguyen told The Deep View. "In other words, exposure is two-sided: it flags where the work is likely to change and where adapting early pays off."

Keeping in mind the metric this report measures, which is skill exposure to the possibility of AI transformation, the key point here is that it is just looking at how well AI can do individual tasks, for example, coding or writing. This report doesn't measure the human element in how those tasks are applied in the workplace, such as how they are synthesized through real-world experiences that cannot be easily replicated by a chatbot, including leadership, critical thinking, and reasoning. If anything, to me, a lot of this data is also reinforcing one thing: the importance of developing human skills that AI cannot replicate.

TOGETHER WITH CRUSOE

$5 in free credits, on us.

Every Crusoe account now comes with $5 in free credits for Crusoe Intelligence Foundry. Use them for Serverless Inference or Serverless Fine-Tuning, whether there's a new open model you've been wanting to try or a workload you're ready to run.

No new commitment, no setup required. Credits apply automatically to your account and are ready whenever you are.

Claim your $5 in credits, start building with Crusoe.

ENTERPRISE

Why companies want to own their intelligence

There's lots of talk about companies "owning their own intelligence." Thomson Reuters is one of the first enterprises to pull it off.

On Monday, Thomson Reuters, the content and technology company, launched Thomson, its first proprietary large language model. Thomson was developed in-house, using an open-source model as its foundation and then trained on decades of proprietary content, including expertise from Westlaw, Practical Law, Checkpoint, and Reuters. The model is specifically designed for professional work and emphasizes the importance of domain expertise. 

Because of the way the model was built, Thomson Reuters fully controls it, which gives the company a better understanding of the intelligence embedded in the model, reduces dependence on another company's AI roadmaps, and, most importantly, reduces costs. The company reports that the endeavor cost $40 million in talent and compute, which is comparatively low compared to what frontier labs spend on training. 

"A real opportunity is for a company like Thomson Reuters to take those frontier open-source models and use them as starting points for our training and development with our unique data and compute. And in that way, we're able to train highly capable AI systems at a fraction of the cost that it would be to train one from scratch," said Joel Hron, Thomson Reuters CTO, in a briefing with the media. 

So far, it has been trained on only 10% of Thomson Reuters' content, but the company says the goal isn't just to keep feeding it more content, but to turn more of it into targeted training data that improves the model. 

The company reports that its early evaluations put Thomson on par with the latest frontier models across a range of tasks. It also attributes the quality of its content as the differentiator from proprietary models. When comparing how GPT-5.4, Sonnet 5, and Thomson 1.0 Large performed with web content only versus Thomson Reuters content only, there was, in every instance, some improvement in factuality or completeness. It also said that in its early evaluations, Thomson performed on par with the latest frontier models across a range of tasks.

At launch, it will be available in CoCounsel Legal, Thomson Reuters' legal AI platform built on its Fiduciary-Grade AI standard, which is designed to meet the requirements of highly regulated industries such as law, tax, and accounting.

With the industry's focus on improving efficiency, there has been a strong emphasis on finding alternative ways to reduce costs when deploying models without compromising results. A key way to do so is to use domain-specific models, which are typically smaller and therefore require fewer tokens. They also tend to be faster and more accurate because they are working from a smaller, more specific data set. These domain-specific models are also key players in model routing, which promises to save users money by tapping into them when useful and is one of the industry's hottest trends. In Thomson Reuters' case, beyond saving money, a major advantage is being able to use its trove of data to inform the AI. However, the interesting thing is that the company has previously entered into licensing deals with AI labs such as Meta and Microsoft. We have to wonder if the company will take a different approach to content licensing now that it has its own proprietary model.

LINKS

• Tech firms shaped US school supply chains via funding, reach

• Texas Gov. Abbott says data centers "dug their own grave" over opposition

• New York Times reportedly tests AI-generated search summaries

• Nvidia to spend $6 billion on bolstering US open source ecosystem

• Business schools prep students for growing role: Chief AI Officer

• Nvidia reportedly in talks to invest in Perplexity at $30 billion valuation

• Codex: OpenAI gives users a full reset of the usage for all paid subscriptions.

• Wan 3.0: Alibaba launches its latest AI model

• ElevenLabs CLI v1: developers and coding agents can now manage ElevenLabs workflows directly in the terminal

• Runway Ruby: New model that converts SDR video up to 16-bit HDR in ProRes and EXR sequences

• Claude: Enterprise-managed auth for MCP connectors is now generally available

• Nvidia: Solutions Architect, AI Models

• Hilbert: AI Engineer - Core

• Kirkland & Ellis: Innovation AI Developer

• Code Red Partners: Artificial Intelligence Engineer

GAMES

POLL RESULTS

Do you think that frontier models will eventually become commoditized?

Yes (48%)
Somewhat (29%)
No (12%)
Other (11%)

MARKETS

How a mystery model surprised the AI industry

An anonymous AI model has taken the developer community by storm. And it's reportedly giving frontier labs a run for their money.

On Thursday, a "stealth model" called Ox Alpha was released on OpenRouter by a third-party provider who decided to remain anonymous during its preview. Its creator said that the reasoning model was designed for coding, sustained agentic work and production workloads.

Twitter tweet

The model is available for free in preview with "near unlimited usage" for a week, according to OpenCode, with an eye-popping capacity for 100 trillion tokens per day. The model features a 1.05-million token context window and 131,000-token output, or the limit that a model can generate in a single response. 

• The creator noted that the model is suited for long-horizon software development tasks and workflows that "combine text with visual context." 

• Additionally, though prompts and outputs are retained by the model provider, they aren't currently used for training. 

• Though early reports suggest the mystery model is beating frontier models like Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol in coding evaluations like Deep SWE, the model's independent validation isn't yet available on formal public leaderboards. 

Patrick Collison, CEO of Stripe, which is acquiring OpenRouter, described Ox Alpha in a post on X as "very impressive." 

As the model goes viral, many have started to speculate about its origins. While some have suggested that Ox Alpha is the next generation of Google's Gemini models or Microsoft's MAI models, others have speculated that the model comes from a Chinese lab, with some analysis pointing to Z.ai's GLM family. However, none of these guesses have been confirmed.

No matter what the origins of this model end up being, Ox Alpha's overnight stardom shows that the AI industry is easily distracted, and often excited by every shiny new toy that hits the market. However, early testing and industry fervor are one thing, and actual, long-running sustainability and results are another. Without knowing who has created this model, users should be wary about its safety and security in real-world applications. Additionally, because the model doesn't currently use user data for training, that policy could change. In short: While the model is drumming up a lot of excitement right now, it's yet to be seen whether this is a flash in the pan, or if the industry will lose interest once the next big model gets released in the days (or even hours) ahead. Still, this supports the bigger trend of frontier intelligence commoditizing.

TOGETHER WITH GRANOLA

Be The Person Who Remembers Everything

Ever notice those people who walk into meetings and somehow remember every conversation, decision, and next step from the last time? They’re not some kind of savant – they just use Granola. 

Granola’s AI notepad captures your meetings automatically. No bots, no awkward recording prompts, just a quietly powerful log of everything that matters. Ask Granola questions via chat and you’ll instantly get the answer you’re looking for, including who it came from, when it was solved, and what notes were taken on it. Best of all, the more you use it, the better Granola gets – over time, it builds a searchable memory of every discussion and decision made. 

If you’re ready to become the person in the room who remembers everything, it’s time to try Granola for free right here.

STARTUPS

Where small models are challenging AI giants

With enterprises finally clamping down on tokenmaxxing, small models have emerged as a counterweight to expensive proprietary APIs. 

On Monday, marketing and behavioral insights firm NextLM unveiled research detailing the results of its new system, Savant 3.5, built on lightweight, open-source NVIDIA Nemotron models and purpose-made for one specific task: helping salespeople find the best customers. In its niche task, the model outperformed major models from OpenAI, Google, Anthropic, SpaceXAI, and Moonshot, and at a much lower price. 

"I think it validates the idea that small models can not only compete with the frontier, but actually, in some cases, win," Chris Anzalone, founder and CEO of NextLM, told The Deep View. 

Savant is built on top of one of Nvidia's most compact models, sitting at roughly 30 billion parameters with 3 billion active for each input, to find customers. Then it uses Nemotron

3.5 Lightning as a "supporting signal" to sharpen the rankings. 

Here are the results: 

• The most important result was how many eventual customers appeared in the top 10% of a model's customer recommendations. NextLM's Savant model put 24.6% of buyers in that 10%. The closest competitor, GPT-5.6 Sol at 22.8%. 

• Grok 4.5 scored third-highest at 22.3%. Anthropic's models, meanwhile, sat near the bottom of the group, with Fable 5 sitting at 18.1% and Opus 5 ranking dead last at 14.6%. 

• Additionally, Savant managed to do so at a much cheaper rate, sitting between $0.003 and $0.011 per 1,000 customer prospects scored, compared to between $0.26 to $5.11 for the frontier-model APIs tested. 

NextLM's success provides an example of where small models fit best: niche, specialized tasks. "If a model isn't tuned to your business outcomes, the value really isn't yours," Anzalone said. "It's transferable to whoever uses that same model."

Small models offer several benefits over bulky, expensive models accessed through proprietary APIs. As Savant proved, these models can be trained to do hyper-specific tasks and achieve hyper-specific business outcomes at a fraction of the cost. Additionally, because these models can often be stored on local hardware, they're generally more private and secure than handing your business's most critical data over to a model provider. And for enterprises desperately searching for returns while also being on high alert amid increasingly common AI-powered cyberattacks, small models may sound like attractive prospects. However, that doesn't mean that enterprises are bound to stop using frontier models and APIs entirely. These systems have a few significant advantages: They're easy to use and provide a readily available source of general capabilities. As Anzalone notes, "People are still going to use general models because they have general information." The challenge enterprises now face is striking the balance between the two.

TOGETHER WITH QUIQ

Most CX teams aren't struggling because the model is too weak

Most CX teams aren't struggling to build AI agents because the model is too weak. They're struggling because customer inquiries are messy, policies have edge cases, and every answer needs the right mix of speed, judgment, and control. That's usually where a strong demo starts to quietly fall apart.

In Quiq's new guide, "CX Agent Design: How to Build One You Can Read and Trust," it breaks down the two dead ends most teams hit when building AI for customer service, the building blocks (Guides, Skills, Tools, and governed boundaries) that replace them, and why a CX agent needs more than good code to actually hold up in production.

Get your free copy of the guide here.

HARDWARE

Why foldables make more sense in the AI era

For almost a decade, foldable phones have been a product looking for a problem to solve. They may have found their 

In a special episode of The Deep View Conversations, we make sense of Google's and Samsung's latest hardware and the AI announcements that came with them. But mostly, we talk about the new folding phones, the Pixel 11 Pro Fold and the Z Fold 8. 

While folding and flip phones have existed for years, this summer both Google and Samsung upped the ante by launching new experiences that let AI enthusiasts make the most of the added screen real estate for AI workflows. 

Topics covered include:

• The new AI features available on the Pixel 11 phones 

• How Gemini contributes to the AI experience on mobile

• Does Google still have the lead in AI hardware?

• The minimal hardware improvements to the Pixel devices

• The advantages of owning a foldable in the AI era 

• How Samsung's Galaxy Z Fold 8 series compares

• The advantages of the Z Fold 8's "passport" form factor

• How Apple's foldable, rumored to launch in September, will compete 

If you're trying to understand how AI is changing what you can do with a smartphone, and what your next phone purchase should be if you prioritize AI, you won't want to miss this episode. 

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

LINKS

• Apple cuts 200-plus jobs from Vision Pro and Siri teams

• DeepSeek unveils experimental multimodal AI model rivaling Anthropic

• Anthropic hires ex-Google TPU chief Amir Salek for in-house chip push

• South Korea's Kakao Corp to spin off chat platform into new KakaoAI unit

• Anthropic could reportedly raise $100 billion in IPO

• Hugging Face explores M&A deals worth $13 billion 

• Anthropic's Opus 5 overtakes Fable 5 in enterprise spending

• Claude Security: scans now run on Mythos 5, public beta for Enterprise customers

• Antigravity Remote Control: Access active sessions on any modern supported browser or mobile device across iOS and Android

• Gemini Notebook: Unveiled a slew of updates

• GPT-5.6 Sol: OpenAI cuts API pricing by over 20%

• Nvidia: Solutions Architect, AI Models

• Logicbroker: AI Engineer

• Hilbert: AI Engineer - Core

• Deloitte: Google Gemini Manager, Technical Transformation

GAMES

A QUICK POLL BEFORE YOU GO

How Crusoe makes scaling AI less painful

As enterprises move from experimentation to deployment of AI, fast, easy and affordable inference is growing increasingly important. 

One estimate from Gartner finds that, especially as agentic workflows continue to drive up token usage, inference costs are expected to grow fivefold by 2028. Crusoe might have a solution: owning every layer of its stack, from the energy to the data centers to the cloud infrastructure. 

That vertical integration makes all the difference, Michael Yen-Chi Ho, director of technical product marketing at Crusoe, told The Deep View. Owning the chain end-to-end allows the company to ship faster, build faster, and make high-performance inference more affordable. For example, Ho said, its MemoryAlloy technology delivers 9.9 times faster time-to-first-token and 5 times greater throughput compared to vLLM on OpenAI's GPT-OSS-120B.

But competing on affordable infrastructure isn't Crusoe's only advantage. The company is also offering efficient and seamless integration throughout its portfolio. Its Managed Intelligence services, for instance, is designed to help companies at any stage in their AI journey. This sits at the heart of Crusoe's Intelligence Foundry, a single user experience that simplifies deployment, fine-tuning, and monitoring, allowing teams to focus on models over infrastructure. 

Crusoe's Managed Intelligence services are split into a few primary offerings: 

• Its Serverless Inference offering is designed for "immediate production readiness," said Ho, allowing teams to scale AI without needing to provision or manage any of the infrastructure. 

• Crusoe's Tailored Deployments, meanwhile, give customers options for workloads that need custom infrastructure. The company's engineers work with customers of all sizes to design deployments around their specific needs, including traffic patterns, latency requirements, model choice and compliance. 

• And Self-serve Deployments are for developers in need of dedicated compute capacity. These can be launched in minutes and are configurable to a developer's specific optimization needs. 

• For users wanting to fine-tune a model without the overhead, Crusoe offers Serverless Fine-Tuning, allowing customers to upload their own dataset, select an open model of choice and fine-tune for their specific use case. 

Additionally, the company contributes to the open-source market with inference-serving frameworks and orchestration. For example, the company built "fastokens," a high-performance tokenizer that improves serving and throughput, and have been integrated into mainstream serving platforms including vLLM, SGLang, and NVIDIA's Dynamo. 

"In short, we offer an owned cloud platform for cost and speed, a full-lifecycle Managed Intelligence portfolio, open-source contributions, and flexible deployment models to suit any customer's needs," said Ho.

And with AI growing more powerful and complex by the day, companies are seeking partnerships that make AI deployments as seamless as possible. 

"There's also a real shift in appetite for who does the work," said Ho. "We're the cloud partner that offloads the operational hassle, so our customers can focus on creating value with AI."

Start building AI with Crusoe today.

CONSUMER

For AI, Pixel 11 Pro Fold is the phone to beat

Two things make Google's Pixel 11 Pro Fold the best phone for AI builders. And it's likely to beat anything Apple launches in 2026. 

I've been using Google's latest Pixel Fold device for the past week, and it's surprised me. I've tried a lot of foldables, reviewed them, and traditionally slid them back in a drawer when I was done. But the new Pixel Fold is going to be very difficult to put down, for two reasons: 

1. Rambler voice dictation: The Pixel 11 Pro Fold has Google's new AI voice technology built into the Google Keyboard (Gboard). You simply tap the new mic + Gemini button and start rattling off your AI prompt (or any other text), and Rambler will remove the ums and ahs, adjust corrections you make midstream, and produce fully coherent, neatly formatted text. This is very similar to the Wispr Flow app on Mac (and Windows), which has been the AI industry's favorite productivity boost of 2026. Rambler might be even faster and more accurate than Wispr Flow on Mac. While Wispr Flow has Android and iPhone apps, they're clunky and cumbersome. Rambler works like a champ and is far better than the traditional dictation features built into Android or iPhone. For daily Wispr Flow users, it's a massive relief to be able to dictate on a phone like we can on a computer. Rambler is also on Google's other Pixel 11 devices, and could come to other Pixel and Samsung devices in time.

2. Controlling agents from the larger mobile screen: One of the things you can do once you're able to rattle off long prompts or instructions from Rambler is to open your AI apps and start new queries and tasks, or steer ones that are already in progress. With the interior 8-inch display that feels like a tablet, you can work in a mode that's much closer to a laptop or desktop. With ChatGPT, Claude, Gemini, and Perplexity, you can open the app on the inner 8-inch screen and get your list of chats and left menu-bar items on one side of the screen, with a full chat window on the other. On all four of those apps, you can also connect them to the desktop version of the app and remotely control the tasks your agents are carrying out. This is a great thing for AI developers at a time when there are reports of people walking around San Francisco and New York with their laptops open so their agents can keep running. The combination of agents being remotely controlled from your phones and the latest foldables offering an even more desktop-like experience from mobile could let developers touch grass a little more.

While I've been testing the Pixel 11 Pro Fold, I've also been testing Samsung's new passport-sized Fold 8, which is based on the long-rumored form factor of Apple's forthcoming iPhone Ultra, expected to arrive in September. The passport form factor is charming, very small and lightweight, and great for typing on both the cover screen and the inner screen. It's much smaller than the Pixel 11 Pro Fold. It's akin to the difference between the iPhone 17 Pro and the iPhone 17 Pro Max, but the size differential is greatly amplified on a larger screen. And then there's Rambler. The Fold 8 doesn't have it, although it might get it eventually. And Apple's Siri dictation is nowhere close to being as effective as Rambler or Wispr Flow. However, in January, Apple quietly acquired Israeli startup Q.ai, which could offer similar functionality. But until we hear otherwise, the size of the Pixel 11 Pro Fold screen plus its Rambler dictation feature put it a step ahead of Apple and Samsung for AI builders.

TOGETHER WITH JUMPCLOUD

Your employees have HR. Your AI agents don't. That's a problem.

Agents are onboarding themselves, accessing systems, executing workflows, and making decisions with no formal record, no named owner, and no offboarding when their purpose expires.

New research from JumpCloud found that non-human identities now outnumber humans in 83% of organizations, yet only 21% have governance controls in place.

The workforce has changed. The infrastructure managing it hasn't. HR for Agents isn't a metaphor, it's the missing operational layer for every organization running AI at scale.

See what securing every identity looks like.

HARDWARE

Why AI leaders have pivoted to efficiency

AI is already causing a resource crunch. But we may only be at the beginning, one executive says. 

Kevin Cochrane, chief marketing officer of neocloud Vultr, told The Deep View that AI demand has only hit the "tip of the spear." Because most of AI demand is coming from hyperscalers, developers and AI-native startups, several massive companies are still very nascent in their AI journey. 

For every Fortune 500 company, there is likely a portfolio of thousands of applications for both internal and external uses, according to Cochrane. "Every single one of those applications needs to get modernized," he said. "That's a 10-year build cycle, and that's going to demand a tremendous amount of compute capacity all around the globe." 

It's no secret that enterprise could represent a cash cow for AI. That's why OpenAI and Anthropic are vying to serve major organizations, investing in forward-deployed engineers and signing partnerships with legacy tech organizations such as IBM to make that transition smoother. 

And even beyond enterprises, Cochrane asserted, AI presents the opportunity for "unbounded demand," essentially only limited by humanity's capacity for creativity. "I think we're going to get to a point where we're going to see that there is a predictable pace of growth of demand," he said. 

However, once that deployment comes to fruition, the industry runs into another problem: Compute capacity and energy constraints. There is a "theoretical limit" to how much power you can and should generate, or how much land and water you can and should use, Cochrane said. 

With AI already impacting infrastructure, "something has to give," he said. The thing that gives may be the chips themselves, something that companies like Nvidia and AMD apparently realize as they shift focus to efficiency. 

"Given the fact that there has to be some natural limit, and there's infinite demand … You have to be as efficient as possible," said Cochrane. "You have to drive down that efficiency curve year after year after year."

Cochrane isn't the only one reading the tea leaves on how AI will affect infrastructure. Investors have taken an interest in the technologies at the foundation of AI, such as energy and mining, seeking to get ahead of the problems that AI could create.

Cochrane is striking a common theme in saying AI will drive fundamental changes in work and life. Fundamental changes, however, require innovation from all sides. Chips and data center architecture are just one piece of the puzzle in making this technology more efficient. At the utility level, more sources of clean energy could ease the burden that AI and data centers are placing on overstressed energy grids. And the models themselves could get more efficient, too, such as Pathway's 150-million-parameter small reasoning model, BDH-CQ, which costs a fraction of OpenAI's most cost-effective model, GPT 5.6 Luna and demands less compute and energy. Efficiency is a core part of innovation. Computers used to take up entire rooms, and mobile phones used to be giant bricks as long as your forearm. AI's version of this means trimming down every component of the stack, from energy to infrastructure to the models themselves. 

TOGETHER WITH TABS

Tabs + PwC: A practical framework for usage-based revenue in the AI era

Hybrid and usage-based models are changing how companies price, and how finance teams recognize revenue, forecast, and close the books.

In this on-demand session, Tabs co-founder Rebecca Schwartz and PwC partner Amit Dhir break down the real operational impact, with frameworks and examples you can apply now.

Watch the session

PRODUCTS

OpenAI just made Siri's upgrade more urgent

For over a decade, composing, searching, and sending Apple Messages with AI meant relying solely on Siri. OpenAI just changed that.

On Thursday, the AI lab launched an Apple Messages plugin that lets ChatGPT assist with your text messaging needs, including finding specific messages, summarizing conversations, and composing and sending messages. At launch, it is limited to ChatGPT on your Mac via ChatGPT Work and Codex. 

While the description may make the plugin sound like just another feature that lazily helps you write or summarize a thread, the real-world help it can offer is far greater. For instance, in the demo video, a user asks ChatGPT which message threads from the previous day it should follow up on and gets a bulleted response that identifies each message and a suggested reply, which users can review and send. Another helpful use case is asking it to check your calendar and compose a text with a couple of free times.

Twitter tweet

The big question about this plugin's value is less about the help it can offer users and more about how it differentiates itself from the AI users will get with the launch of iOS 27. Although that won't happen until the iPhone launch event and the broader release of iOS, both likely in September, Apple previewed improved Siri features at its WWDC event in May, which The Deep View covered in depth. The updates include a chatbot-like interface for Siri that would let users ask it to do everything the OpenAI plugin promises. 

Apple has yet to respond to a request for comment.

The major difference is that with the Apple version, users can use it across their ecosystem of devices, not just on their Mac, which is far more convenient. The only argument in favor of the plug-in is that ChatGPT power users won't want to context-switch between the two AI interfaces. While that is true, with the launch of the new and improved Siri, users can also just ask Siri the same questions conversationally and get the answers, circumventing the need for ChatGPT altogether.

The reaction to this plugin on social media has been overwhelmingly positive, with people calling it a game-changer and celebrating OpenAI for unveiling an AI feature that could actually help in everyday life. This has been really interesting to see, as I didn't see quite as overwhelming a reaction during WWDC, which could be because it was just one of the many features announced. Both this announcement and the positive reaction to the foldables launched this summer by Samsung and Google are teeing up Apple to unveil its own counterparts in September. The momentum also confirms one thing: this iPhone launch event may be one of Apple's biggest in years. It's one to watch for both new AI developments and to see which new AI features Apple puts in the spotlight. 

LINKS

• Ramp launches AI model routing service called Router via API

• Coworkers split on junior staff using AI on the job

• AI found to widen wealth and geographic inequality in US

• Meta spends hundreds of millions yearly on Azure AI model access

• Anthropic PBC expects IPO to match or exceed SpaceX's record offering

• AI nudify apps may pose a back-to-school risk

• Google Gemma: has surpassed a billion downloads

• Slack Code: collaborative AI coding channels 

• Pocket: Meta rolls out its experimental vibe-coding gaming app

• Exa: Plug-in available in Codex and ChatGPT 

• Apple Messages plugin: Search messages, catch up on convos, and more in ChatGPT

• Tessera Labs: AI Engineer

• Innovaccer: Artificial Intelligence Engineer

• Boom: AI Engineer

• Crowe: AI Enablement Manager

GAMES

POLL RESULTS

Do you think Gen Z adopting and accepting AI is important?

Yes (52%)
Somewhat (26%)
No (15%)
Other (7%)