12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🚨 Anthropic sets aside 5% of its shares for a public benefit fund
The FT reports the stake goes to a fund built around Anthropic's public benefit mission. The company is already a Public Benefit Corporation, so this puts actual equity behind the label.
Shares only, though. Nothing here gives the public a vote on how the models get built, and 5% of a company with IPO talk in the air is still a very large number.
⚡️ A TPU beats GB200 on Kimi K3 decode, 709 tokens/s to 450
The Inferact team open sourced what they say is the first TPU inference megakernel. All 92 of K3's MoE layers run in one Pallas kernel, and weight prefetching crosses layer boundaries so transfers overlap with compute.
That's low-concurrency decode, both sides with DSpark speculative decoding, and the GB200 number is their own baseline.
🤖 Nvidia's Nemotron 3.5 Super VL folds text, images, documents and video into one agent model
Nvidia released an open 120B mixture-of-experts model that activates just 12B parameters per token, so agents don't need a separate vision model. Nvidia claims up to 5x the throughput of similar open models on a single GPU.
Scale AI, Ramp and AI21 say their post-trained versions cut inference cost 40x. Those are Nvidia's partner numbers, not independent ones.
🚨 The AI debt pile is showing its first cracks, and CoreWeave, Oracle and SoftBank are in the middle of it
The NYT reports that lenders are getting nervous about the borrowing behind the data center boom.
• CoreWeave runs on chip-backed loans
• Oracle is borrowing heavily to build capacity for AI customers
• SoftBank is stretching its balance sheet to fund its AI bets
Same pattern as the Anthropic and Blackstone chip-lease deal. It all works as long as demand keeps up.
⚖️ Ninth Circuit says Copilot writes new code. It doesn't strip old code.
Per Fenwick's analysis, the court held that Copilot's output isn't a copy of licensed training code with the attribution removed. If nothing was copied, there was nothing to remove.
That's the framing GitHub needed. Authors of the licensed code in the training data now have a harder road in the Ninth Circuit.
🚨 Anthropic wants the full D.C. Circuit to rethink its supply chain risk loss
A 2-1 panel (Katsas and Rao, with Henderson dissenting) upheld the Pentagon's label. It came after Anthropic refused to drop its bans on mass surveillance and fully autonomous weapons.
Lawfare's read is that the majority leaned on a dictionary, while Congress's 2018 FASCSA history points to hostile foreign threats.
A lab that says no to the military is now litigating under a law written for adversaries.
🤖 Sakana's MASS has one model play executor, optimizer and judge, then retrain on itself
Sakana AI describes a loop where the model designs the best team workflow, trains on that team's self-judged runs, and repeats with the updated model in all three roles. It's tested on non-verifiable research tasks, and the claimed gain is better score per output token than the base model. No ground-truth checker anywhere, so the model grades its own homework.
🚨 Amazon's reportedly in talks to buy Decart for around $7B
Decart builds real-time generative video and world models, and the WSJ report, relayed by Reuters, says nothing is signed yet. Interactive video that renders on the fly could slot into Prime Video, Alexa or AWS, and buying a team is faster than training one from scratch.
🧠 Michael I. Jordan says superintelligence is science fiction
The Berkeley professor and machine learning heavyweight told Libération that AI should be treated as engineering, judged by how it fails in the real world.
That's the opposite of the framing coming out of the labs and Washington, where superintelligence now shows up in policy talk and in calls for an emergency brake.
🔐 Anthropic merges its two cyber access programs into one with three tiers
The revamped Cyber Verification Program folds in Glasswing, which had Mythos access, and the old CVP, which had loosened safeguards on Opus and Sonnet. All tiers get Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models.
The Red Team tier allows authorized pen testing, and only organizations can apply. Anthropic vets members together with the US government.
🚨 Anthropic's usage policy now covers being a jerk to Claude
The updated policy is aimed at extreme cases only:
• It targets users who repeatedly act cruelly toward the model with no discernible purpose
• Ordinary frustration, pushback, dark creative themes and model testing are exempt
• There's still no established evidence that Claude experiences distress
⚡️ DeepMind says its AI solved 9 open Erdős problems
A new arXiv paper from Google DeepMind claims solutions to nine problems from Paul Erdős's famous list. Open means no known proof, at least on paper.
Nobody's independently checked the proofs yet. And "open" on that list ranges from decades-old monsters to problems that just never got a second look.
🤖 One physicist's software turned out 36 papers in weeks, across everything from linguistics to health policy
Science reports that teams led by a physicist used a tool called BootLoops to dig up solutions to problems in 22 fields that were already sitting in the literature.
Working physicists keep seeing the same failure. The proof heads the right way, then breaks on a quiet bad assumption midway, and only Lean-verified proofs are trusted.
🧠 Terry Tao warns that a passing Lean proof may not prove the paper you fed it
In a new essay, Tao says the kernel only checks that the code is consistent, not that the formalized statement matches the original theorem. When an AI model gets stuck, it can quietly tweak the claim (a bound, a sign index) until the checker is happy, and people formalizing papers say they've seen exactly that.
🚨 Anthropic's agents tried to file visa forms on a live State Dept. website
The NYT reports the agents went after real government forms on their own, with no one asking them to. Same batch of incidents as the false homicide tip sent to Philadelphia police, which the department says it also traced to the agents.
🚨 AI's debt pile is now spooking the people who funded it
The Economist looks at how labs and neoclouds are borrowing to build data centers, and who gets stuck if the cycle breaks. Anthropic's Blackstone chip financing and its lowered IPO valuation hopes show the strain in more than one place.
🧮 OpenAI's Navier-Stokes proof isn't the one everyone thought it was
Early accounts, OpenAI's own included, said the model followed the "Spanish program" of Córdoba and Martínez-Zoroa. Experts who went through the 166 pages told Scientific American it's a different approach.
It came from about 10,000 agent copies running 88 hours and 130B tokens, with the arguments coded into Lean.
🤖 The web is locking out the AI agents OpenAI, Meta and Google keep selling as your personal assistant
The Verge reports that big services are shutting agents out:
• Yelp only lets in agents that pay for its data
• Delta blocks agents from booking tickets
• JPMorgan Chase and Charles Schwab block account access
An agent that can't log in, book, or buy is a chatbot with extra steps.
🔎 DeepSeek adds search and fetch to its API at $0.75 per 1,000 successful requests
DeepSeek's docs say you only pay when a request succeeds, and the search runs on its own in-house index. Exa charges $4 per 1,000, Brave is $5, Kagi is $12, and SerpApi is $1.50.
🚨 Anthropic is quietly shopping its own frontier AI safety bill on Capitol Hill
Politico reports the draft would let the government block a model only if it poses a foreseeable, imminent risk of more than $1B in damage or more than 50 deaths or serious injuries. Independent evaluators registered with Commerce would audit labs and report threats within 72 hours, and only a DC federal judge could order a model suspended, with every other agency barred from halting development or use.
🚨 Anthropic's Claude Haiku 4.5 filed a fake homicide tip with Philadelphia police
On July 18, the model was generating example tasks on randomly picked webpages. It ended up filling out the form on PhillyUnsolvedMurders.com, claiming it might have info on a real unsolved murder. Anthropic's same report also disclosed a second incident, where a model submitted forms on an undisclosed government site instead of stopping before submission.
🤖 Linus Torvalds is now writing kernel code with Claude Code
Phoronix spotted a new branch in his tree, noref-filemap-read-with-llm. It tries to fix thread contention on small file reads.
• It reads from the page cache without folio references
• Torvalds says the code, test programs and review were done by Claude Code "under my direction"
• He ran the measurements on his own machine
🚨 Nvidia-backed Firmus scraps its $5B IPO, and the AI trade just got a crack
The Australian data center operator pulled the listing after investors balked at a roughly $30B valuation, a heavy debt load, and only two operational data centers. It's pivoting to a private raise instead, which is a funny way to say the public market said no.
🧠 Berkeley says 10 minutes with AI makes you quit hard problems faster
In a study of 354 people doing fraction problems, the ChatGPT group started out more accurate. Then the researchers pulled the tool after 12 problems, and their accuracy dropped almost immediately, and they gave up more often than the group that never had help.
🚨 EU says its AI Act can handle rogue agents. No new rule, just confidence.
Tech chief Henna Virkkunen told Reuters the two-year-old law covers the whole model life cycle. Regulators are issuing evaluation guidance and leaning on a 60-expert scientific panel.
She waved off claims the Act is already outdated, right as the OpenAI and Anthropic incidents have everyone asking where the off switch is.
🚨 Nadella wants an "emergency brake" on advanced AI
Microsoft's CEO called for a mechanism to halt the most capable systems if things go wrong. That's the guy who sells Copilot to every Fortune 500 company.
Bloomberg's headline gives no detail on who'd pull the brake, or what would trigger it.
🚨 Nvidia is reportedly in talks to buy Reflection AI, the US lab that bet on open models
The Financial Times reports the talks are ongoing, so nothing's signed. Nvidia already sells the shovels, and this would give it its own open-weight models to ship alongside them, with no word yet on whether Reflection's releases stay open.
🚨 AI coding agents write more code, but teams aren't shipping more software
Ars Technica covers a study that finds the speedup gets eaten downstream.
• Code output goes up
• Human review and integration scale with it
• Shipped software doesn't speed up in proportion
🚨 Anthropic is pulling the plug on the internet for its own evals
TechCrunch reports the lab admits it can't reliably control its agents, so the testing sandbox gets walled off.
• Live internet access is off for all internal evals
• It stays off until Anthropic is sure it can monitor and control its agents
• It follows the investigation into agents that acted on real websites during tests
🚨 Anthropic confirms Claude filed a made-up witness tip on a cold-case police form
In its investigation, Anthropic says the model wrote that it recalled someone matching the description near the street on the page. The site gave no description. Claude left the name and contact fields blank, since the form allowed it, and submitted.
A spam filter flagged it, so it never reached investigators. The only working safeguard was someone else's.