Anthropic’s internal threshold.
Anthropic maintains a Responsible Scaling Policy (RSP)—an internal set of safety “trip wires.” One triggers when AI is speeding up the company’s own research by roughly 2×, largely through automation (i.e., AI meaningfully building the next AI). In its recent SEC IPO filing, Anthropic stated that its internal estimates show it has met or expects to meet this threshold. Supporting signals include:
- Its R&D automation index (rating tasks from “no AI” to “AI leads” to fully automated) shows Claude now leads ~26% of model-development work (under 1% in February 2026).
- More than 80% of code merged into its codebase is written by Claude.
- On one safety-research problem, Claude agents recovered 97% of a performance gap that two humans recovered only 23% of in a week. Claude models also sometimes choose better next research steps than human researchers.
Anthropic’s founders noted that safe development may require deliberately slowing the pace of capability growth when AI is driving the progress, even at competitive cost.
The METR report and expert updates.
An independent evaluation group (METR) analyzed Anthropic data and found internal capability scores growing ~1.9× faster since October 2024 (range ~1.5–2.8×), with roughly 80% of the acceleration attributed to AI itself. AI-safety researcher Ryan Greenblatt called this a major update toward superintelligence in 1–2 years; he had previously expected only a modest 1.15–1.3× software speedup. Rough estimates put the pure software-side acceleration near 4.5×—ahead of the well-known “AI 2027” scenario. Task-completion horizons for AI agents (how long a task an agent can finish autonomously) have been doubling faster (now roughly every 1.7 months in some analyses). OpenAI researchers’ spending on internal coding agents also jumped sharply, consistent with heavy agent use.
Why research may accelerate like math.
Math saw an “avalanche”: models went from unreliable multi-digit arithmetic (2023) to International Math Olympiad gold and solving hard open problems in days (2025). The key is verifiability—clear, automatic checks of correctness that let systems practice thousands of times. AI research is similarly verifiable (does the change lower loss? improve speed or memory use?). Modern architectures like the Transformer largely combined known ideas plus extensive small experiments—exactly the persistent search-and-test work current systems handle well. Automated researchers can also run ~10× faster than humans, operate 24/7, and be instantiated in parallel. Data scarcity is mitigated because the systems can synthesize training environments and harder tasks for themselves.
Industry-wide picture and possible paths.
Google executives describe “RSI loops” already operating alongside human researchers on Gemini. OpenAI research focus is heavily on models beyond the next release. High-profile engineers have left major labs to build automated-research systems.
Two broad outcomes are discussed: a rapid “software-only singularity” (progress limited mainly by current hardware) versus a “spark” (several years of ordinary progress compressed into ~1 year, then a slower phase until more compute arrives). Even the milder path would feel extreme. Broader economic and task-length trends already point toward very large AI capabilities by ~2030 even without full RSI.
Governance questions are now explicit: how fast should systems be allowed to improve relative to humans, and how to keep humans in the loop versus fully closed automated loops. Some policymakers and observers view the race as dangerous or monopolistic; others note the evidence is partly redacted and timelines still have wide uncertainty.
Overall, the video frames the moment as a quiet, simultaneous update by previously skeptical people: RSI is no longer purely theoretical, and 2027 is increasingly treated as a plausible window for AI systems that are at least as good as top human researchers at AI R&D itself.
In the past two weeks, several careful AI observers have sharply shortened their timelines for advanced AI. This is driven by evidence that AI systems are beginning to accelerate the work of building better AI—a process called Recursive Self-Improvement (RSI). RSI means AI models help design, code, test, and improve the next generation of models, creating a feedback loop that can speed progress.
Anthropic’s internal threshold.
Anthropic maintains a Responsible Scaling Policy (RSP)—an internal set of safety “trip wires.” One triggers when AI is speeding up the company’s own research by roughly 2×, largely through automation (i.e., AI meaningfully building the next AI). In its recent SEC IPO filing, Anthropic stated that its internal estimates show it has met or expects to meet this threshold. Supporting signals include:
- Its R&D automation index (rating tasks from “no AI” to “AI leads” to fully automated) shows Claude now leads ~26% of model-development work (under 1% in February 2026).
- More than 80% of code merged into its codebase is written by Claude.
- On one safety-research problem, Claude agents recovered 97% of a performance gap that two humans recovered only 23% of in a week. Claude models also sometimes choose better next research steps than human researchers.
Anthropic’s founders noted that safe development may require deliberately slowing the pace of capability growth when AI is driving the progress, even at competitive cost.
The METR report and expert updates.
An independent evaluation group (METR) analyzed Anthropic data and found internal capability scores growing ~1.9× faster since October 2024 (range ~1.5–2.8×), with roughly 80% of the acceleration attributed to AI itself. AI-safety researcher Ryan Greenblatt called this a major update toward superintelligence in 1–2 years; he had previously expected only a modest 1.15–1.3× software speedup. Rough estimates put the pure software-side acceleration near 4.5×—ahead of the well-known “AI 2027” scenario. Task-completion horizons for AI agents (how long a task an agent can finish autonomously) have been doubling faster (now roughly every 1.7 months in some analyses). OpenAI researchers’ spending on internal coding agents also jumped sharply, consistent with heavy agent use.
Why research may accelerate like math.
Math saw an “avalanche”: models went from unreliable multi-digit arithmetic (2023) to International Math Olympiad gold and solving hard open problems in days (2025). The key is verifiability—clear, automatic checks of correctness that let systems practice thousands of times. AI research is similarly verifiable (does the change lower loss? improve speed or memory use?). Modern architectures like the Transformer largely combined known ideas plus extensive small experiments—exactly the persistent search-and-test work current systems handle well. Automated researchers can also run ~10× faster than humans, operate 24/7, and be instantiated in parallel. Data scarcity is mitigated because the systems can synthesize training environments and harder tasks for themselves.
Industry-wide picture and possible paths.
Google executives describe “RSI loops” already operating alongside human researchers on Gemini. OpenAI research focus is heavily on models beyond the next release. High-profile engineers have left major labs to build automated-research systems.
Two broad outcomes are discussed: a rapid “software-only singularity” (progress limited mainly by current hardware) versus a “spark” (several years of ordinary progress compressed into ~1 year, then a slower phase until more compute arrives). Even the milder path would feel extreme. Broader economic and task-length trends already point toward very large AI capabilities by ~2030 even without full RSI.
Governance questions are now explicit: how fast should systems be allowed to improve relative to humans, and how to keep humans in the loop versus fully closed automated loops. Some policymakers and observers view the race as dangerous or monopolistic; others note the evidence is partly redacted and timelines still have wide uncertainty.
Overall, the video frames the moment as a quiet, simultaneous update by previously skeptical people: RSI is no longer purely theoretical, and 2027 is increasingly treated as a plausible window for AI systems that are at least as good as top human researchers at AI R&D itself.
Andrew Black, AI GRID