Anthropic's amended S-1 warns it may soon hit its own threshold for AI-accelerated progress, the same day Q3 shows growth slowing to 47%

Anthropic's S-1/A says it may soon cross its own AI-accelerated-progress line, and that it has missed one before

The new language is specific. Anthropic writes that its "internal estimates and our expectations regarding future models" indicate it may imminently meet the Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI. It also says no current model has been designated past that threshold and that it is still assessing the data. Then it goes further than a lab has before in a legal filing: AI-assisted R&D may push capabilities faster than safety work can keep up, and "we have in the past not identified" a significant acceleration until after it occurred. The company says it is considering RSP changes to keep humans in control of frontier development.

The operational numbers in the filing back up the warning. The list-price value of Anthropic's internal Claude use grew more than 250 times from January to September 2026 and passed monthly personnel expense, excluding stock compensation, in April. Business Insider adds that more than 90% of the code added in September was written by Claude. Hacker News readers went after the accounting. The top comment in that thread says only a company chasing headlines would value internal use at the price it charges customers, not at what it costs. A second commenter asked why headcount nearly doubled if Claude writes 90% of the code. Both points are fair, but neither touches the core disclosure. The point isn't the cash. A company with every incentive to sound measured to the SEC chose to describe its own feedback loop as a material risk.

Read it two ways at once. An S-1 is both disclosure and marketing, and "we're about to build self-improving AI" sells shares. But securities lawyers don't write "we have in the past not identified" lightly, because a statement like that creates liability if it's understated. I'd treat it as the clearest public signal yet of where Anthropic thinks its internal curve sits. The thing to watch is the RSP revision it now says is coming.

We do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible.

— source

Only a company like Anthropic, desperate to make news and help their IPO, would measure by the cost they charge customers

— hungryhobbit, Hacker News

GPT-6 Astra hand-edits Qwen's weights, lifting MATH-500 from 7.9% to 48.1%: LLM-driven optimization reaches weights

The demonstration behind the lead comes from Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab at MIT CSAIL. Khattab is the DSPy author, which explains the framing: after years of using LLMs to optimize prompts and programs, they argue the same reasoning-driven optimization now works on the weights themselves. They call it weight-space programming. Shown only the raw weights, Astra named which of four games a PPO policy was trained on and which shape a small diffusion model draws. It told apart Python, English and romaji models from their middle transformer blocks alone, and it found every induction head without running the network. With zero training, it wrote Breakout, Seaquest and Asterix policies in one reply, and all three beat PPO networks trained for 100 million steps.

The headline result is the Qwen fix. Qwen3.5-0.8B's own model card warns it is prone to thinking loops, and 85 of its first 96 practice answers ran to the token cap. Astra had a Python kernel and a budget of 50 forward passes. It traced the looping to one direction written by the MLPs in layers 15 and 19, then subtracted it from 64 output columns in each, 131,072 weights in all. MATH-500 went from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval roughly tripled. The authors state the limits themselves. With 10 or 25 forward passes the gains mostly vanished. On Gemma 3 270M, which rarely loops, Astra couldn't improve math at all. Their conclusion is that editing can remove an obstacle but can't add an ability. It's a blog post with logs and the edited model, not a paper.

The Hacker News thread, at 610 points and still climbing, split cleanly. A commenter who works on weight-space learning urged people to read past the excitement. Weight editing has existed since at least ROME in 2023, they said, and the post oversells. Another argued the approach gets impractical quickly as models grow. A third camp read it as a training-data question rather than emergence: OpenAI may have fed Astra weight data on purpose as a step toward self-improvement. That last point matters most, in my view. If it's deliberate, other labs are likely doing it too, and the Anthropic filing reads less like an outlier.

Two smaller items fit the same pattern. CUA-Master, a 27B computer-use agent, mines workflows from instructional videos into an outer "world-data-model" loop that retrains on its own weaknesses. It reports 80.2 on OSWorld Verified and gains that keep coming across iterations. Trajectory raised $45 million from Sequoia at a $300 million valuation to sell continual post-training to Harvey, Rogo and others. Its founders pitch keeping humans in the loop as the safe version of RSI. Self-improvement is now a product category.

It is a cool project but not as nearly as insane as people say.

— yurimo, Hacker News

I don't think this is an emergent capability. I think Astra was trained on neural weights data as an attempt by OpenAI to get to recursive self improvement.

— vanyle, Hacker News

Anthropic's Q3: $17B revenue, 47% growth after 144%, and $895B in commitments heading into the roadshow

The revenue line: about $17 billion in Q3, up roughly 1,127% year over year but only 47% sequentially, after 144% in Q2 and 132% in Q1. Nine-month revenue was $33.24 billion. The nine-month net loss was $104 billion, mostly from fair-value changes on convertible notes. The Q3 net loss was $23.29 billion. Compute and infrastructure costs rose 329% over nine months, and committed spend now totals about $895.4 billion, "substantially all" of it compute. Matt Scherer, writing in Tech Policy Press, counts $582 billion of that as non-cancelable commitments to five companies (Amazon, AMD, Broadcom, Google and Microsoft) against $96 billion in cash. He argues the SEC should block the IPO.

Hacker News took apart both the slowdown story and the block-the-IPO argument. The top reply to Scherer said the SEC's job is disclosure, not deciding whether a company is a good bet, and that letting regulators veto IPOs invites corruption. On growth, one commenter relayed a financial analyst who put annualized revenue at $67.8 billion, up from $46.2 billion. Another noted that Opus 5.5 shipped September 22, after most of the quarter, and that September price cuts raised token volume. A third did the valuation math: at $2 trillion, Anthropic needs roughly 1.7x annual revenue growth to 2030, against about 4.7x now.

The squeeze shows up one layer down too. Oracle filed to sell $10 billion in stock and $10 billion in notes. Its shares fell 8.8% even though the raise had been announced, and its five-year credit default swaps now price like a high junk bond. Musk said Grok Bot's serving cost was "unsustainably high relative to revenue" after users found their weekly limits suddenly maxed out, even though SpaceX's AI division reported a profit. Revenue is real, but agentic inference is expensive and capacity is debt-financed. That's why flat months matter so much to everyone downstream.

The SEC's job isn't to adjudicate on whether Anthropic is a good or bad bet for investors. Its job is to ensure that investors have adequate visibility, and then get out of the way.

— marsten, Hacker News

Anthropic releases FlashProve, Claude-improved zkSNARKs for proving which model answered

The research release the IPO coverage drowned out may age best. FlashProve lets a provider commit to a model's weights and then attach a short proof that each response came from that model, without revealing the weights. On Llama 4 Maverick, a 400B-parameter MoE at 4-bit, one H100 generates proofs at about 110 tokens per second, with roughly 19 MB of proof per 1,024-token request. That costs 70 to 95 times the GPU time of serving the model on vLLM. That's expensive, but Anthropic says prior open-source systems were orders of magnitude slower. Code, a C++ verifier, a verifier proven correct in Lean and the top-level Lean proofs are all Apache 2.0. Collaborators include EPFL's Alessandro Chiesa.

Two threads from the lead meet here. Anthropic says Claude produced the speedups, and it added Lean proofs because Claude's choices in a large system are "hard to evaluate." That's a working template for trusting AI-built code. Anthropic also names AI pacing as a use: proving that compute is serving a declared model, not secretly training a new one. A lab that says it may soon cross an acceleration threshold is also building the tools a regulator would need to check. For builders, the nearer use is checking that a router actually served the model you paid for.

A smaller replication gets less attention than it deserves. Second Look Research reproduced Anthropic's "Teaching Claude Why" alignment training on Qwen 2.5 32B and 72B base models. Supervised fine-tuning on difficult-advice transcripts cut agentic misalignment roughly tenfold. Fictional stories about good AIs helped far less and plateaued. The choice of teacher model mattered more than any other data choice. Stories about a generic AI worked better than stories about a named assistant. These techniques shape production Claude models from Opus 4.5 on, and now they can be tested outside the lab.

Times and Tribune papers sue Brave as an AI data "fence": copyright fights move to the suppliers

The New York Times, and separately the Daily News, Chicago Tribune and Orlando Sentinel, sued Brave in federal court in Manhattan. They allege that its index of more than 30 billion pages, built by a crawler publishers couldn't block, was sold to Anthropic, Perplexity, Mistral, xAI and Cohere for training and live answers. The suits also target Brave's paywall blocker. The Times calls Brave a burglar and an AI "fence," and the other papers' lawyer said that buying stolen goods secondhand doesn't make them legitimate. Perplexity answered that the Times is overpaying its lawyers.

The new part is the defendant. Earlier suits went after model makers. This one goes after the search APIs labs use when they can't or won't crawl themselves. Last week's shared robots.txt opt-out bound only the labs that signed it. If this case survives, labs that buy grounding data will need to know where it came from.

Midterm day: "Super Intelligence" becomes a loyalty test, while a 400-app deepfake network slips past app review

Americans vote today, and MIT Technology Review reports that a month after Trump's executive order renaming AI "Super Intelligence," the term mostly hasn't caught on. Among the six signers of the White House accord, only Zuckerberg and Musk used it consistently on earnings calls. Trump has attacked the Wall Street Journal for not using it. The danger the article names is that what you call the technology becomes a partisan marker just as AI becomes a voting issue. On the same day, OpenAI announced a Grid Protectors program with utilities to keep its "compute campuses" from raising consumer power bills, and hired Dan Driscoll as a senior counselor on national security. Both moves fit the post-election PR push reported yesterday.

WIRED and researchers at AI Forensics and Infoblox traced more than 400 apps and domains, apparently run by Chinese developers, that pass app store review as pet-portrait or tutoring tools. Then a server-side switch loads an explicit deepfake generator, often for users arriving from Meta ads. The apps were downloaded more than 350,000 times on Google Play. Apple says the apps were compliant when it reviewed them and the content was swapped afterward. Review of the binary no longer tells you what an AI app does.

BYOD converts four 8B models to diffusion LMs in under four Colab hours each

For tinkerers: BYOD turns Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B and Ministral 8B into masked-diffusion models by training only LoRA adapters, in 3.1 to 4.0 hours each on one Colab GPU. The Gemma conversion nearly matches LLaDA 8B on a nine-task suite. Generation quality at low step counts still trails by a wide margin, so it's a cheap way to experiment, not to deploy. On video, Wave Forcing from SJTU pipelines staggered denoising across GPUs and reaches 134.8 end-to-end frames per second on eight H200s for streaming generation.