Key points
- Anthropic's amended S-1 says internal estimates indicate it may imminently meet its RSP threshold for dramatically accelerated progress.
- Herbie Bradley reads the 2x productivity threshold as R of about 0.5, roughly 10% researcher uplift per Epoch Capabilities Index unit.
- Bradley says markets depend on diffusion, leaving a huge overhang between model capability and enterprise use.
Anthropic Filing Hints AI Is Starting to Improve Itself
Anthropic's amended S-1 says its internal estimates indicate it may imminently meet its own definition of dramatically accelerated progress from AI, which puts recursive self-improvement inside a securities filing. The episode is a debate over how much that line shows. Bradley's case is that the evidence for a self-sustaining loop is thin, capability is jagged and data-bound, and adoption lags far behind what models can already do. Against that, Thelot's AGI centrism treats slowdown as the only sane response if the fast scenario is even plausible.
Takeaways
- Anthropic's filing hints it may be self-improving. The amended S-1 adds a line saying internal estimates indicate Anthropic may imminently meet its Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI. Bradley reads the 2x productivity threshold as an R near 0.5.
- Evidence for fast self-sustaining RSI stays thin. Bradley says the case for a self-sustaining acceleration loop arriving soon is limited next to mostly automated AI R&D, and he cannot tell which world we are in. Very fast takeoff has always looked unlikely to him.
- Capability is jagged and tracks available data. Models improve fastest in domains with abundant data and easily verified answers, as in math and coding, and stall where data is thin. Bradley says open-ended brainstorming has improved more slowly, and medical diagnosis has barely moved.
- Benchmarks are easy for models to game. Bradley rates Frontier SWE and Terminal Bench 4 highly, but says published agent traces often show benchmark hacking. He also praises SemiAnalysis's GDP Val, which tests less verifiable work such as finance.
- Agents lack taste and senior judgement. nostalgebraist compares agents to smart but naive first-year graduate students, book smart but unable to prioritise. Bradley says LLM code and architectural decisions fall below senior-engineer quality, and security holes will slow enterprise adoption.
- Adoption lags capability by a wide margin. Bradley sees a huge overhang between what models can do and what companies use them for, and says demand could rise tenfold or a hundredfold on current capabilities. Shyam Shankar argues the unit of enterprise productivity is the workflow, making change management the bottleneck.
The AGI centrist case
- AGI centrism holds that real RSI and very powerful general AI arrive next year absent massive political intervention, making slowdown the sensible response.
- Thelot finds it plausible but not necessarily justified, since models still have weaknesses and no stage of the improvement loop is automated.