# Anthropic tells IPO investors it may soon hit its threshold for AI-accelerated progress > In updated IPO documents, Anthropic describes recursive self-improvement as a risk for the first time in detail. It says internal forecasts suggest it may soon reach its own threshold for sharply AI-driven progress. > > OpenCheese AI · 2026.11.04 · Web page: https://opencheese.dev/blog/anthropic-tells-ipo-investors-it-may-soon-hit-ai-accelerated-progress-threshold Risk factors in IPO paperwork usually read like weather warnings: they cover everything and commit to nothing. Anthropic's latest one is more specific. In updated IPO documentation, the company tells investors it may soon approach the point where AI speeds up the building of better AI, and that it might not notice the moment when it arrives. The temptation to say "Skynet" is strong and should be resisted, because the company's own wording is more careful than the headlines. ## A risk that finally has a name According to the report, this is the first time Anthropic has described the risk of recursive self-improvement (RSI) in detail. RSI is a feedback loop. An AI system helps build better models, those models help build better ones still, and progress accelerates on its own account. The idea is usually treated as a possible turning point for the field, and it has fans and critics in about equal measure. - **The optimists** expect a sharp rise in productivity, scientific discoveries and the automation of research itself. - **The pessimists** fear that systems which improve themselves could advance faster than people can make them safe. Anthropic's documents acknowledge both sides. They do not pick a team. ## It is already doing some of this The most grounded admission is that the loop is not hypothetical. The company says it already uses AI in its own research and in the development of new models. That is the same practice that, in the company's own telling, could let capabilities grow faster than developers can verify safety, reliability and alignment with human goals. The tool and the hazard are the same object. This is not unusual in technology, but it is rarely put to investors so plainly. ## The jump you do not see Then comes the awkward part. Anthropic says it has not always noticed big leaps in its models' capabilities right away. It also says future accelerations may stay invisible until after they have happened. A risk disclosure whose central claim is that the company might not see the risk coming is a rare thing. It is honest, and it is not especially comforting. ## A threshold of its own making The documents say that internal evaluations and forecasts for future models point to a real possibility. In the near term, the company may reach a threshold it set for itself, which concerns sharply accelerated progress driven by AI. Anthropic is careful to add that none of its existing models has been judged to meet that criterion. So the claim is about trajectory, not about the present. Note also that the yardstick belongs to Anthropic. The company drew the line and is now telling investors it may soon be standing near it. ## Humans stay in the loop, for now The counterweight is a statement about autonomy. According to the documents, Anthropic does not expect, in the foreseeable future, to hand Claude the fully autonomous design and creation of the next generations of frontier models without human oversight. The phrase "foreseeable future" is doing a good deal of work there, as it tends to. Still, it separates a company that uses AI heavily in research from one that has taken people out of the process, and the documents place Anthropic in the first group. ## What the report does not say This account is a summary, and the documents themselves are not quoted at length. Several things are missing: - how the threshold is defined or measured - which evaluations feed the forecasts - what "in the near future" means in months or years - how much of Anthropic's research and development work AI performs today Until those details are public, the safest reading is the plain one. The company says it is closer to an AI-accelerated research regime than it was, that it has not crossed the line, and that it may not see the line when it does. Risk factors usually reassure readers that management would see trouble coming. This one admits it might not. It is an odd sort of comfort, but at least it is honest. ## Related posts - 2026.10.18 · [Google's Brin Says Gemini Is Helping Build Itself, but DeepMind's Own Report Sees No Speedup](https://opencheese.dev/blog/google-brin-gemini-self-improvement-deepmind-report-no-speedup.md): Sergey Brin says Google is entering recursive self-improvement with Gemini 4 Argon. DeepMind's safety report, published the same week, finds no meaningful acceleration of AI research and rates the risk as low for now. - 2026.10.31 · [Anthropic's IPO prospectus warns investors its AI could pose existential risks](https://opencheese.dev/blog/anthropic-ipo-prospectus-warns-investors-ai-could-pose-existential-risks.md): Anthropic's IPO paperwork says powerful AI could pose catastrophic or existential risks to society, and that the company can't guarantee its own safeguards will prevent them. The same document warns that a slower AI market would hurt growth. - 2026.10.29 · [Microsoft posts a $92.3 billion quarter and finally says what Azure earns](https://opencheese.dev/blog/microsoft-record-quarter-azure-33-2-billion-openai-10-billion.md): Microsoft reported $92.3 billion in quarterly revenue, its tenth straight record, and for the first time put a dollar figure on Azure: $33.2 billion. OpenAI alone accounted for $10 billion.