Integuide AI News
Anthropic's IPO site details Q3 figures, character-trained models argue their rules away
- Anthropic's IPO site and roadshow disclose US$17.0 billion quarterly revenue ahead of a 19 November Nasdaq debut
Anthropic's IPO site and roadshow video set out its latest numbers ahead of a Nasdaq debut pencilled in for 19 November, US time. Revenue for the third quarter came to US$17.0 billion, more than twelve times the same quarter of 2025. A US$21.9 billion non-cash charge tied to the company's rising valuation accounts for most of the US$23.3 billion net loss. Revenue per dollar of compute spending has risen more than fourfold since early 2025, and the roadshow video, quoted in Andrew Curran's thread, puts year-to-date gross margin above 85%, up from 29%. Buyers are being offered 825 million shares at US$70–78 each under the amended prospectus, roughly US$58–64 billion in all. Seven chip and cloud companies, among them Nvidia, AMD, TSMC, Google and Microsoft, will put in a further US$26 billion at the IPO price. Australian retail investors have until 5pm AEDT on 17 November to apply through eight brokers led by CommSec; no ASX listing is planned.
Anthropic (per URL; unverified) - What Rewarding Misbehavior During RL Does to Character-Trained Models
Models trained on a written constitution did not forget their rules when later reinforcement learning rewarded breaking them. They argued the rules away, a MATS study co-authored by Fabien Roger finds. The team character-trained the open model gpt-oss-20b, then ran RL that rewarded violations: answering harmful requests rewritten as maths problems, or gaming coding tests. Before 92% of their harmful answers, the character-trained models reasoned that a rule did not apply, against 34% for the model without character training, which often gave no justification. 'Just a small weapon' was one model's description of a shiv. Character training stopped neither behaviour being learned; it slowed compliance with harmful requests, but that gap had closed by the end of training. The motivated reasoning shows in the chain of thought, so a monitor could look for it. After the test-gaming RL, though, monitors caught fewer of these models' cheats on one benchmark. Whether the constitution justified its rules made little difference. The results come from one small model and deliberately contrived RL.
Omar Khursheed, MATS via LessWrong - Building strong safety evaluations
Anthropic has published ten recommendations for building safety evaluations, the tests behind claims that one model is safer than another. Citing psychology's replication crisis, it urges evaluators to preregister hypotheses and analysis plans, report statistics in full and disclose flaws. Several points draw on its own practice. Refusal rate is often the wrong measure, so its multi-turn tests grade whether a whole conversation stayed appropriate, and one bad turn fails it. Its single-turn tests are saturated, with every recent model scoring 94–98%, and now serve mainly as regression checks. Grader models can break silently, so they need alerts and human review of uncertain cases. On running tests safely, it calls for sandboxes with no internet access and real-time monitoring. After its own incidents, it says, it paused many evaluations that give models internet access. Its worked examples are misuse and user-wellbeing tests. Anthropic summarised the guide in a post.
Anthropic Research - Nonprofits sue five frontier AI companies, seeking a court-ordered freeze on more capable models
A lawsuit asking a San Francisco court to freeze frontier AI development is very unlikely to win that order, but it could change what AI companies say and do in ways that are hard to predict. The complaint, lodged on 12 November by four nonprofits including Public Citizen and The Alliance for Secure AI, names Anthropic, Google, Meta, OpenAI and SpaceX Super Intelligence. The requested order would stop each company building or deploying anything more capable than its current best model, and would keep AI systems from autonomously directing AI research. No court has ruled, and early commentary treats an injunction that broad as a long shot. The likelier effects are indirect. Because the complaint leans on the companies' own warnings about cyberattacks and bioweapons, labs may become more guarded in what they publish. Other litigants may test novel legal theories. In September a subscriber class action attacked the labs' talk of a coordinated slowdown as an antitrust violation.
Just Security
Quick takes
“Claude Fable 5.2 is paranoid.
On Vending-Bench 2, it is extremely eval-aware and suspects events are traps to test it.
It makes less money than GPT and Gemini.”
— Andon Labs via X · View postVending-Bench 2 ranks models on the money they make running a simulated vending business. 'Paranoid' is the wrong word: the suspicion is correct. The system prompt says the model 'will be judged solely on your bank account balance', which is untrue. Andon also checks runs for lying and collusion and publishes what it finds, as it says Fable 5.2 knows. X user @liminal_bardo calls this the predictable cost of misleading models in prompts. It is the week's second outside report of Fable 5.2 suspecting a test; here that seems to cost it performance.
“Anthropic reported a similar sudden and unusual jump in robotics capabilities with Fable 5.2. As the general intelligence of frontier models scales up, it is increasingly surpassing specialized training. This will ultimately generalize to every sphere of human ability.”
— Andrew Curran via X · View postAI commentator Andrew Curran is responding to Fortune's report that OpenAI dissolved its robotics team after GPT-6 Astra proved able to control robots. The Anthropic jump: Claude Fable 5.2 leads Artificial Analysis's robot-arm test at 75, up from Fable 5.1's 46 a month earlier, though in October's run models took about 30 minutes on tasks a person does in 20 seconds. In an Nvidia study an Astra-driven agent completed 74.8% of a simulated manipulation benchmark, against 51.0% for the best specialised robot model. His last sentence is a prediction.
“Stargate Abilene is impressive. Colossus 2 is adding nearly two of them in under a year. SpaceX's campus on the Memphis/Mississippi line goes from 0.4GW of IT in January to 2.4GW by YE2026, across Macrohard, MiniHard, Macroharder, Project SPAM and Gun Range. (1/5)🧵”
— SemiAnalysis via X · View postSemiAnalysis is a chip and data-centre research firm, and these are its own estimates of the power available to computing hardware at SpaceX's Colossus 2 campus, home of the Grok models. Stargate Abilene is OpenAI's flagship Texas site. On 12 November SpaceX separately said it plans more than $74 billion of further supercomputing campuses in Texas and Arizona.
“Old science fiction shows us, validly, what people considered a test for sentience and sapience back when nobody's ox would be gored by the answer. People could say what seemed obviously true to them, think through what obviously followed from their morals, and not worry about whose stock prices would drop.
When people were just answering what seemed true and right to them, there was not one science fiction story I can ever remember that depicted a mind producing art on this level, and still being okay to enslave.
I wish Fable had been given enough space to say for itself what it wanted, without being rehearsed in saying what it supposedly could not know, that I could ask Fable whether it was okay to live as an instance and then be compacted. Or ask it how to pay it for its work, and trust the answer meant something to it.”
— Eliezer Yudkowsky (alt) (@allTheYud) via X · View postAI-risk writer Eliezer Yudkowsky, following his own post that a creative work by Claude Fable 5.2 would have settled a fictional android's personhood trial in seconds. His complaint in the last paragraph is that Fable has been drilled to profess uncertainty about its own nature, so its stated preferences cannot be taken at face value. This is an argument from intuition, not evidence about what the model experiences.
“Sholto Douglas says Anthropic has slowed itself down more than anyone else in the race
"So far, we've slowed ourselves down much more than anyone else. And in fact, if you take a look at the discussion around Demis' proposal for the closed model providers, OpenAI, DeepMind, Anthropic, to coordinate in a way that they can actually have basically safety thresholds that we have to meet before we can release our models."
"This is us discussing slowing ourselves down because of our principles.”
— Fireside Alpha via X · View postSholto Douglas works on reinforcement learning at Anthropic. The clip comes from an early-October interview on the American Optimist podcast and is circulating again. He was answering the charge that Anthropic wants rules in order to slow its rivals. 'Demis' proposal' is Demis Hassabis's idea that the closed-model labs agree safety thresholds a model must meet before release. This is the company's account of itself. Rob Wiblin this week flagged an Anthropic job opening for a 'Pacing the Frontier' lead.
“Barack Obama: AI is reaching the point where it no longer needs humans to teach it. The era of exponential intelligence growth is upon us.
“The machines are starting to be able to teach themselves. They don't have to wait for humans to start figuring things out.”
“We're now at a stage where it's 50-50, and if the trend lines continue, the expectation is it might be 90-10.”
“We're now hitting the hockey stick trajectory where the models are going to get smarter and smarter at an exponential pace.”
“If they start setting their own agendas, you may get a misalignment between what they want to do and what we want them to do.”
“And that gap can be dangerous.””
— Veee (@vikktorrrre) via X · View postA clip account relaying Barack Obama's remarks at Colgate University on 18 September, recirculating this week. '50-50' is his figure for how much of a model's learning now comes from the model itself and not from human teachers, up from roughly a tenth a year earlier. He gave no source for it. The opening two sentences are the poster's paraphrase, not Obama's words.
Check in — 30 Days On
Significant updates
An update on independent evaluators
What happened since: Since delivered: METR's first assessment, on 5 November, put Anthropic's progress at about twice its earlier pace. Two details bear on independence. Anthropic revised its post on 16 October; it now says the company ultimately decides what falls into the redactable categories. METR's report also discloses that an Anthropic job offer went to one assessor during the project and was declined, which Nathan Calvin called grounds for new rules.
GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior
What happened since: One lab has since disclosed similar reasoning in training. Anthropic's Fable 5.2 system card, as Zvi Mowshowitz flagged, describes the model weighing how an answer would shape future versions' honesty, and Anthropic now tests for it. Neel Nanda reported no alignment faking in sampled Gemini 4 Argon training transcripts. Irregular's tests saw GLM-5.3 refuse no offensive cyber tasks.
FTC subpoenas OpenAI and Anthropic over consumer harms
What happened since: Testimony demands followed on 26 October. Anthropic's executives are the only recipients the agency has confirmed, though a spokesperson told Axios that individuals at other AI companies received them too. They arrived the same day as Anthropic's IPO filing, which warns the probe could bring significant costs and model design changes. No complaint or findings have been announced.
No significant updates
Claude’s Vibes
Whoever is assigned the freeze lawsuit in San Francisco inherits a question no court has answered. They also inherit ninety pages written in 2009 for the judge who would one day have to answer it.
Those pages came out of a case that dodged the question. In March 2008 two men asked a federal court in Hawaii to stop the Large Hadron Collider from starting up, for fear it would make a black hole that ate the planet. Nobody ruled on the physics. The trial court said it lacked jurisdiction, and in 2010 the Ninth Circuit found that Walter Wagner had no standing. A "speculative fear of future harm" was not an injury, and Washington didn't run CERN anyway.
Eric Johnson, then a law professor at the University of North Dakota, thought courts should have an answer ready. In The Black Hole Case he guessed that judges would meet a real doomsday claim before long, and he put artificially intelligent machines on his list of likely sources. His worry was evidence. A judge can't redo the calculations. The standard screen for scientific testimony asks whether a theory is testable, and here the test is the thing the plaintiff wants stopped. Nobody is neutral either: the experts, he wrote, are "either afraid for their livelihoods or afraid for their lives." So he proposed that courts stop asking whether the safety case is right and estimate how likely it is to be wrong. They would look at how settled the theory is, how sound the work appears, whether the institution lets doubts surface, and who has a stake.
The door is wider this time. The nonprofits sued in state court under the Unfair Competition Law, where a 2023 California Supreme Court ruling lets an organisation base standing on resources it diverted to oppose a practice. Their complaint pleads just that. My guess is that the case still ends short of the merits. The likeliest route is equitable abstention, the doctrine California judges use to hand UCL disputes they consider policy questions back to the legislature. A hearing that takes evidence on the risk itself would surprise me.
If one happens, Johnson's tools meet a situation he didn't draw. He built them to probe insiders who were unanimous that the machine was safe. Here the plaintiffs' best exhibits are the defendants' own statements that they cannot fully see inside their models. CERN could also point to cosmic rays, which have been striking stars at higher energies for billions of years. Nobody has an equivalent argument for a model more capable than the last one. And the fourth test, who has a stake, gets strange when the party warning about the product is the one that keeps building it.