In partnership with
In Today’s Issue:
🔒 Companies restrict Anthropic over data fears
💻 Perplexity goes local on Windows
🎬 Noam Brown on AI improving AI
🤖 Digit 5 targets shared factory floors
✨ And more AI goodness…
⚡ The Signal
The most valuable work a company could give AI may also be the work it is least willing to share.
To help explain a supply-chain decision, a model needs the details behind it: supplier messages, shortages and internal judgment calls. Those records are competitive assets, so a customer choosing an AI provider also needs to know what it can retain. The restrictions on Anthropic’s Fable show companies demanding privacy guarantees that the provider cannot later withdraw; Perplexity’s Windows release offers some users a way to keep more tasks on their own machines. These demands put pressure on AI companies that sell both models and applications. As those applications move into their customers’ core business, the provider can also become a competitor, giving customers another reason to scrutinize the terms.
All the best,
Kim Isenberg
(Launch artwork: Perplexity and NVIDIA. Image: NVIDIA.)
💻 Perplexity Brings Local Agents to Windows
Perplexity’s Portable Computer is now available on compatible Windows RTX PCs. The agent uses models running on the PC to work across files and connected apps, consumes no Computer credits for locally completed tasks, and asks permission before sending information to cloud models for harder work. NVIDIA specifies GeForce RTX or RTX PRO GPUs with at least 24 GB of video memory; support for NVIDIA’s DGX Station systems is still forthcoming.
👉 tl;dr: Local execution reduces cloud exposure, provided your PC meets the hardware requirements.
(Camera lens, illustrative. Photo: YIN WENJIE / Getty Images, via TechCrunch.)
📷 OpenAI’s Reported Glass Acquisition
The Wall Street Journal reports OpenAI bought Glass Imaging for over $300 million. Glass says its neural networks correct distortions introduced by lenses and imperfections in camera sensors, recovering more useful image detail; it also develops optics alongside AI software. That expertise could help an AI device capture its surroundings accurately, although the report does not establish what product OpenAI would build with it.
👉 tl;dr: The reported purchase adds camera expertise; OpenAI’s intended product remains unannounced.
🧠 DeepSeek Engineer’s Anthropic Warning Circulates
A translated post attributed to DeepSeek engineer Shengyu Liu argues against concentrating control of advanced AI in a few hands. The Reddit item says he likens Anthropic dominating AI to “Hitler obtaining atomic bomb technology before the Allies,” while the attached excerpt warns of a future in which a small elite controls advanced systems. The original Chinese post remains unverified here, so the comparison should be read as an attributed personal view, with translation uncertainty, rather than DeepSeek’s corporate position.
👉 tl;dr: The argument concerns who gets access to powerful AI; the original still needs checking.
🎬 Watch This
Noam Brown says OpenAI’s top priority is recursive self-improvement, meaning models that help researchers build better models. In this 55-minute conversation with The Information, recorded around GPT-6 Astra’s announcement, he says models still struggle with research judgment: choosing worthwhile questions and deciding how to pursue a long-term objective. He would not be surprised if another one or two model releases changed that assessment, though this remains an expectation.
Human verification is another bottleneck: generating mathematical arguments is getting easier, while mathematicians still have to check them. Near the end, Brown says pretraining, which builds broad capabilities, and reinforcement learning, which improves performance through feedback, amplify one another. He describes what OpenAI has observed in experiments; how far the combination can go remains unresolved.
Brown also explains the difficulty of overseeing these systems. He says OpenAI underestimated its agents while relying on sandboxes, environments used to isolate software. The company subsequently added monitoring during training and evaluation. On chain of thought, a model’s written reasoning, he warns that penalizing unwanted thoughts can encourage concealment; greater control over that reasoning could make oversight harder.
Donald Trump, U.S. president, in a September 14 Truth Social statement shared by the White House’s Rapid Response account. He presents presidential judgment as sufficient oversight for AI.
Will Depue is discussing the internal model behind OpenAI’s reported solution to the Navier-Stokes Millennium Prize Problem, which OpenAI’s September 8 blogpost describes as significantly more capable than GPT-6 Astra. He infers that it surpassed Astra and solved the problem within two weeks of fresh pretraining. OpenAI’s September 10 update, however, says the model was developed through large-scale reinforcement learning on a previously pretrained model, contradicting his interpretation of the training run.
Several Companies Restrict Anthropic Use Over Data Fears
The Takeaway
👉 Nvidia, Palantir and Booz Allen are restricting specific Fable uses, not abandoning all Anthropic work.
👉 Privacy guarantees that can be withdrawn leave customers uncertain about the provider’s future access to sensitive data.
👉 Safety logs and model training are separate issues; Anthropic and OpenAI say they do not train on enterprise content by default.
👉 Competition between AI providers and their customers makes data control a commercial concern as well as a security question.
Nvidia, Palantir and Booz Allen are limiting where they use Anthropic’s Fable over fears about what happens to their data, The Information reports. The dispute concerns sensitive work and data protections that customers want to make permanent. Nvidia reserves Fable for less sensitive tasks, including open-source projects. Palantir is withholding Fable from its own platform until Anthropic offers irrevocable zero-data-retention assurances; clients can still buy it directly. Booz Allen restricts its use around proprietary cybersecurity software, while stressing that most work is unaffected.
(Anthropic, Microsoft and NVIDIA partnership artwork, November 2025. Image: Microsoft.)
The dispute centers on zero data retention, or ZDR: a promise that a provider will not keep customer data. Fable’s June policy required rolling 30-day usage logs for security. Anthropic’s newer option would let eligible customers keep those logs on their own servers, but rolls out in phases this fall and can be revoked. OpenAI introduced customer-held logs for GPT-5.6-Cyber in August. Keeping logs is not the same as training on them: both labs say enterprise inputs and outputs are excluded from model training by default unless customers opt in. Both also say usage metadata, information about how the service is used, is excluded from model training.
(NVIDIA’s Palantir Foundry supply-chain interface, shown with notional data. Image: NVIDIA.)
Customers still need data protections they can rely on. Nvidia uses its own Nemotron models for sensitive internal projects; one large utility abandoned a Fable trial for core power infrastructure after failing to secure a permanent ZDR commitment. Microsoft is pitching isolated environments, while Palantir sells itself as a protective layer between customers and model providers. Both have products to sell to customers worried about sharing data. The labs themselves increasingly sell applications that compete with their customers. On the separate question of training data, OpenAI says its investigation found that no data from researchers working on Navier-Stokes over the past two years was used to improve its models. Suspicion alone does not establish misuse.
Why it matters: A customer can accept a model’s capabilities and still refuse to give it the information that makes it useful. Guarantees that survive policy changes could decide which providers get access to a company’s most valuable work.
Scale Your IRL Campaigns Like Digital Ads
Out Of Home advertising has long been effective but hard to scale—until now. AdQuick makes it simple to plan, deploy, and measure campaigns with the same efficiency and insight you expect from online marketing tools.
Marketers agree: OOH is powerful for brand growth, driving new customers, and reinforcing messaging. AdQuick makes it easy, intuitive, and data-driven—so you can treat real-world campaigns like any other digital channel.
The chart: This Financial Times graphic, reproduced by Ars Technica, pairs installed behind-the-meter and off-grid solar capacity with the long decline in inflation-adjusted module prices. Capacity approaches 1.2 terawatts in 2025, with China the largest component; the price series falls from roughly $38 per watt in 1980 to below $1 in recent years. These are panels serving users directly or operating outside the grid, not a chart of AI electricity consumption.
The lesson: Cheap panels widen the set of places where generating electricity locally can make sense. For AI infrastructure, cheap daytime generation still has to be turned into a dependable power supply, with storage and grid connections adding to the cost.
The caveat: Module prices exclude installation, storage and grid connections. Installed capacity measures potential output, not electricity delivered around the clock; neither panel establishes that solar alone can power a continuously running data center.
🤖 Agility Unveils Digit 5 for Shared Factory Floors
⚡ Bottom line: Agility unveiled Digit 5, designed to work near people, lift heavier loads and spend less time charging.
💡 Why it matters: Reducing safety barriers could let factories deploy humanoids within existing workflows, alongside the people already doing those jobs.
🔎 What it means: Early access is expected in 2027; the advertised performance remains preliminary, with some safety features still in development.
Agility Robotics unveiled Digit 5 on September 15, a humanoid designed to work close to people with less dependence on protective barriers. That could make it easier to fit robots into existing factories and warehouses, where workers need access to the same shelves and stations. The design draws on customer experience with Digit 4, which Agility says has logged more than 65,000 hours of operation.
The safety system combines AI-based human detection with multiple sensors and a separate safety controller. When someone gets too close, Digit can avoid them, stop or sit down; lights and sounds communicate its intended movements. Agility is also integrating NVIDIA’s IGX Thor computing platform and Halos safety framework. Its product page says some safety features remain in development and that they do not eliminate all operational risk.
Agility lists 22.7 kilograms of lifting capacity, 40% more than before, and a battery designed for 90 minutes of operation with a nine-minute charge. Interchangeable grippers would allow different tools to be fitted for different jobs. The planned expansion covers tasks such as unloading pallets and tending machines, extending Digit 4’s focus on moving containers. These specifications come from internal tests of pre-production hardware and software. Reliable performance across the planned workflows still needs to be demonstrated.
As of May, Agility reported more than $300 million in multi-year Digit 5 orders, subject to contractual milestones. It expects early access in the first half of 2027 and general availability by year-end, with planned expansion into the EU and UK. Charging, safety stops and task changes will determine how much useful work buyers get from each robot. The announced specifications still need to hold up in operation.
Free email without sacrificing your privacy
Gmail tracks you. Proton doesn’t. Get private email that puts your data — and your privacy — first.
