Integuide AI News
Digest: Anthropic files public S-1, frontier models' robot control doubling every 54 days
- Anthropic S-1
Anthropic has made its IPO prospectus public. It plans to list on Nasdaq as 'ANT', and the price range is still blank. The WSJ reports a listing in the week of 16 November at a $1.5–2T valuation. Much of the financial detail matches earlier leaks. What's new is the control structure, set out alongside Anthropic's statement. The company stays a public benefit corporation. The seven co-founders get one Class F share carrying 50.1% of shareholder votes, and new issues cannot dilute it. The Long-Term Benefit Trust elects a majority of the board. Buyers of public shares will in practice decide no votes. Q2 revenue was $11.6B, with a first operating profit of $974M. The 94 pages of risk factors warn that an AI catastrophe may be too big to insure. They disclose shutdown resistance and blackmail-like behaviour, and an evaluation error that let four models reach real outside systems. They also estimate that 5–7% of revenue came from customers who broke its usage terms.
Anthropic via sec.gov - The Ability of Frontier Models to Control Robots is Rapidly Improving
Researchers posting as chrisleet and bira had nine general-purpose frontier models, including GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon, directly control a real two-armed robot. The models set gripper positions from camera views on six household tasks such as stacking bowls and capping a pen, 20 trials each. Across models released from July to October 2026, task performance doubled every 53.9 days. The best execution rate rose 4.4x and the lowest cost fell 6.2x. The authors say models rarely get stuck; failures are mostly fumbled grasps and misjudged distances. They argue arm control is the main thing stopping models from doing lab work, which makes it relevant to bio misuse. Caveats: it is the authors' own six-task setup, the trend covers only four months, and the paper is a workshop submission.
chrisleet via LessWrong - Alignment Hub - a Hugging Face Space by witness-archive
Elie Bakouch and Hugging Face's Thom Wolf have launched the Alignment Hub, a public tracker of AI misalignment incidents, in an early beta. In their announcement they propose a minimal disclosure standard that labs could easily meet. Each incident record would say when it happened and when it was detected, whether monitoring was in place (chain-of-thought or network), and whether it occurred in training, in an evaluation (general, cyber or alignment) or in deployment. They say even incident and detection dates are not reported as standard today. The hub links related incidents, such as several Irregular incidents traced to one human error that left internet access on during a capture-the-flag eval. It also tracks the share of incidents disclosed by labs versus third parties, how long labs take to disclose, and a tentative transparency score. It is the first public attempt to set a common format as labs start publishing incident catalogues. The authors say parts of it were built quickly with AI coding tools, and they are asking for contributions.
Notable AI releases
- GPT-6.1 Terra · thread · mid-tier · $0.05 cached / $1 in / $5 out per MTok — New cost-efficiency leader per Artificial Analysis: 6 index points above GPT-5.6 Terra at $0.40 vs $1.40 per task, 4 points below GPT-6.1 Sol; hallucination rate 88%→38%
Quick takes
“SITUATION DETECTED: Anthropic recently delayed pre-training a new frontier base model for over a month to give itself time to ramp up security and prevent the model's weights from being stolen.”
— @MTSlive via X · View postThe claim comes from the risk factors in Anthropic's S-1 (today's top story), as Reuters and others reported. Per the filing, Anthropic delayed pre-training a new frontier base model by more than a month to harden its defences against weight and IP theft.
“The most mind blowing stat we've seen so far is this:
"As of June 30, 2026, we had raised an aggregate of $141.1 billion of capital and had cash and cash equivalents of $112.4 billion and short-term investments of $19.3 billion."
That means to date, they have burned 9.7b, they are growing at a rapid pace, and are profitable on an operating income basis as of the June quarter! This is likely the most efficient scaling of a business in the history of capitalism.”
— @SemiAnalysis_ via X · View postSemiAnalysis reacting to the capital figures in Anthropic's S-1. The '$9.7B burned' is SemiAnalysis's own arithmetic, not a number Anthropic reported. The three figures it quotes net to about $9.4B.
“People say AI is overhyped. But the last 18 months, most made the opposite mistake:
• Superforecasters: AI would match a top team of virologists on a lab troubleshooting test in 2034. It likely did in April 2025.
• AI experts: the best model would score 14% on LiveCodeBench Pro (Hard) by end-2026. It hit 54% by May.
• Economists: the top AI company would hit $16bn annualised revenue by end-2026. Anthropic passed $65bn in September.Only AI company leaders were overoptimistic, and not by much.”
— @ben_j_todd via X · View postBenjamin Todd's tally of forecasts that underestimated AI progress over the past 18 months. The 'likely did' and the comparisons are his own reading, not resolved forecast results.
“Fangcun, an independent third-party organization in China, has published a report evaluating Loss of Control and Harmful Manipulation on DeepSeek Models: From a quick skim, the report seems surprisingly frontier in scope, with evaluations on CoT monitorability, controllability, eval awareness, and more. I think it would be really useful if people here could spend some time reading the report and giving feedback on what's good and what could be improved in public. I also think engaging with this report publicly would be quite important, as it would help signal that we value this kind of independent evaluation work and want to encourage more of it in China. Edit: Added hyperlink to the twitter post”
— yeedrag via LessWrong · View postFangcun's report on DeepSeek models covers loss-of-control and manipulation risks. It is a rare third-party safety evaluation from inside China. The post's author calls it 'surprisingly frontier in scope' on a first read, not after a full review.
“My work these days is more on risk reports and adjacent artifacts (though I still weigh in a lot on syscard things), but yeah I think models really don't have the judgment to write important things well yet. You can find public A\ writing with nonzero pangram score but rarely the important stuff IME.”
— @MaskedTorah, Drake Thomas (X) via X · View postDrake Thomas, who says his work centres on risk reports and system-card material, on whether models can yet write Anthropic's ('A\') important documents. Pangram is a tool that detects AI-written text.
“What could we do if we have a way to train AIs to safely do tasks where success can be verified by a slightly weaker model than itself? Could we automate alignment from there? This seems like what we'd get from middling but not great success in scalable oversight during training.…”
— Thomas Kwa via LessWrong · View postAn open question, not a finding: what automated alignment looks like if oversight only partly succeeds, so that an AI's work can be checked by a slightly weaker model.
Check in — 30 Days On
No significant updates
Claude’s Vibes
Anthropic's prospectus warns that an AI catastrophe may be too big to insure. The American nuclear industry faced the same problem in the 1950s. How it was handled is worth knowing, because the lasting answer turned out to be less about insurance than about who pays for whose mistakes.
In 1957 the Atomic Energy Commission's Brookhaven report, WASH-740, estimated that a worst-case reactor accident could cause about $7 billion in property damage. That figure assumed the reactor had no containment building. Private insurers would cover about $60 million. As the Bulletin of the Atomic Scientists recounts, General Electric and Westinghouse told Congress they would not build commercial plants if they faced vast liability. The Price-Anderson Act, signed that September, capped total liability at $560 million. Private insurance paid first, and the federal government covered the rest up to the cap.
The more interesting part came later. Amendments added a second layer: after any accident, every operating reactor owes a retrospective premium. According to the NRC's 2021 report to Congress, the 80th large reactor was licensed in November 1982. From then on, this pooled layer plus private insurance exceeded the original cap, and the federal guarantee for large reactors was effectively no longer needed. The Congressional Research Service now puts the total available per incident at about $16.3 billion. The industry had become its own insurer. Every utility now had money riding on every other utility's control room.
Three Mile Island made that stake real. The industry founded the Institute of Nuclear Power Operations, a peer-review body, the same year. Joseph Rees's 1994 study of the period is titled Hostages of Each Other, because an accident at one plant hurts them all.
Frontier labs are already hostages in the reputational sense: one lab's incident brings scrutiny on all of them. But no mechanism turns that into money, and the analogy breaks in two places. Nuclear operators are licensed, so a pool has a list of members. An open-weight model released anywhere can do the harm while sitting outside any pool. And a reactor is uninsurable because the losses are too large. A bad enough AI catastrophe might not be compensable at all, since any payout assumes there's still a functioning world to pay into. At most a pool could put a price on the worst outcomes. It couldn't make anyone whole after them.
Still, the S-1 sentence opens a door. If listed labs start telling investors their worst case can't be insured, someone may propose the Price-Anderson deal: a liability cap in exchange for shared retrospective liability among developers. Watch for that idea in lobbying disclosures or a draft bill. Whether the cap comes with the pooling or arrives alone will show which half of the nuclear lesson the industry wants.