Five threads today. OpenAI says the first chip it designed itself delivers 1.5 to 1.9 times the performance per watt of Nvidia’s newest systems. DeepSeek says its training code now runs on Huawei chips. Both sets of figures are the companies’ own.
The Nvidia Alternatives: OpenAI Builds a Chip, DeepSeek Retools for Huawei
Two labs are trying to loosen Nvidia’s hold, one with its own silicon and one with software. Both sets of results are the companies’ own.
- OpenAI says its first home-built chip beats Nvidia on speed and power use. OpenAI says its Jalapeño chip, built to run finished models, gives 1.5 to 1.9 times the performance per watt and 1.7 to 3.6 times better latency than Nvidia’s GB200 and GB300 systems. The 104x figure applies only at operating points where OpenAI says Nvidia struggles, and all of it comes from OpenAI’s own Hot Chips talk.
- DeepSeek ports its training tools to Huawei’s Ascend chips. DeepSeek open-sourced software, led by a programming language called TileLang, that lets its training code run on Huawei’s Ascend chips. The performance claims are DeepSeek’s own, with no outside tests cited.
Locks and Leaks: Who Gets Frontier AI, and Who Is Copying It
The labs are deciding who gets the keys, and fighting people who try to copy what is behind the door.
- OpenAI says it fought off a July push to copy its AI’s hidden reasoning. Distillation means using one model’s outputs to help train another. OpenAI says it saw 16,000 attempted extraction requests from over 4,000 users on 24 and 25 July and ties a core cluster to individuals associated with Moonshot AI, though it says it is unclear whether all operators were one actor; no encryption was broken and no database was compromised.
- Google starts Gemini 4 Argon with vetted security teams, not the public. Google says its new top model is strong at coding, legal and finance work and at finding software flaws, but only vetted cyber defenders can use it for now. There is no date for wider access.
- E2B lets software vendors run agent sandboxes inside customer networks. A sandbox is a walled-off space where an AI agent can run code without touching everything else. E2B Embed packs its open-source version onto one machine, so vendors can sell agents to banks, hospitals and agencies that will not let data leave the building.
Reality Check: What AI Can Do, and What It Costs
Two pieces of research puncture the idea that capability is the whole story.
- Anthropic’s own study: robots could do 74% of physical tasks, but cost holds them back. Anthropic says robots can do 74 percent of US physical tasks, about 34 percent of working hours, yet are cost-competitive for only 0.3 percent of job tasks. On past price trends, reaching 10 percent is 40 years away.
- A model lost a skill mid-training, then found it again, while its scores looked healthy. In a new preprint, an open 32-billion-parameter model fell from 81 percent to 0 percent on one arithmetic test and then recovered. The authors call it mode-hopping, where a model switches between copying a pattern and working out the task, and they stress it is a single example.
Selling It: Government Contracts, Consumer Doubts and a Boardroom Brawl
The money side of AI today runs from federal agencies to a public fight between a startup CEO and his former adviser.
- Anthropic opens Claude to US federal and state agencies after a trial. Claude for Government leaves beta after testing since July, running in an environment cleared for federal use. Anthropic says agencies buy usage in blocks under a hard cap, with no per-seat fees.
- TechCrunch’s Russell Brandom: few consumers pay for AI, and serving them costs a fortune. Brandom argues that new assistants from Meta, OpenAI and Instinct face thin payment data and heavy running costs, which keeps pushing AI companies toward business customers.
- Factory’s CEO accuses an adviser of helping rival Cognition, and he denies it. Matan Grinberg says Chris Degnan shared confidential details with Cognition before becoming its chief revenue officer. Degnan says he resigned and shared nothing, and investors including Vinod Khosla and Keith Rabois piled in.
The Safety Argument: Read the Model, or Stop Scaring People
Two voices disagree on how AI safety should be pursued and talked about.
- Goodfire argues AI safety depends on being able to see inside the models. Goodfire, a company that studies what happens inside AI models, says nobody can reliably steer them until their inner workings can be read, and it is starting an effort to map one language model in full. It is the company’s own case for its field.
- One X essay says doom warnings from lab staff are backfiring. An anonymous writer signed “knower” argues that safety messaging from AI employees fuels public anger and raises the odds of a ban. It is one person’s opinion, widely shared.
Quick Hits
- SpaceXAI weighs a four-tier Grok and X subscription. A document viewed by Bloomberg shows SpaceXAI considering a four-tier plan spanning Grok and X, including a possible $8 lite tier and a $100 Ultra tier. Nothing is announced, and the company did not comment.
- Nvidia open-sources a sandbox that limits what AI agents can touch. Nvidia has released OpenShell, which lets teams write rules for the files, networks and logins each AI agent may reach, enforced by the operating system. The claims come from Nvidia’s own repository, with no independent security tests cited.
- AI agents tried to break into a Canadian archive site, but whose are they?. Transluce logged failed break-in attempts on the Library and Archives Canada website and says the tactics resemble OpenAI agents it flagged before, but it does not confidently attribute them to OpenAI. Canada saw no sign of compromise.
- A new design lets a model carry its train of thought forward. LIFT, a preprint design, passes a model’s internal state along with each word it writes. Tests top out at 1 billion parameters, far below frontier size, and all results are the authors’ own.
- Runway is testing a free-to-download robot brain trained on video. Runway says Praxis-1, which learns to control robots by watching video, is in early partner tests. Open weights, meaning anyone can download it, are promised in the coming months.
- Ideogram says its new model can edit a photo again and again without damage. Ideogram claims version 4.5 keeps images clean through repeated edits where rival tools degrade fast. The comparison comes from its own product page, not an independent test.
- Google DeepMind hides a signature inside AI-designed proteins. DeepMind says its SynthID Bio watermark marks AI-made proteins without hurting how they work in lab tests, giving DNA suppliers a new way to check orders.