Has AI entered the RSI era
OpenAI and Anthropic are both running unreleased models capable of speeding up their own research, and the researchers closest to the work disagree on whether that counts as recursive self-improvement.
-
Zvi Mowshowitz says OpenAI and Anthropic have both entered a recursive self-improvement era — but researchers closest to the work aren't convinced
Zvi's September 13 post flagged an unreleased OpenAI model that grew 'more capable than Astra across the board' within two weeks of training, calling that pace more alarming than the Navier-Stokes proof it produced. Two days earlier, John Schulman, Beren Millidge and Charlie O'Neill told Dwarkesh Patel they see diminishing returns and persistent bottlenecks in AI defining its own objectives, not a takeoff. Zvi · Dwarkesh Patel
-
WATCHOpenAI's Noam Brown debates recursive self-improvement on Dwarkesh's podcast
In a Sept 17 episode, Brown discussed agent swarms, models tackling Navier-Stokes-adjacent math, and what it would take to verify alignment before any genuine recursive self-improvement — without claiming today's systems have crossed that line. Dwarkesh Patel
-
Practitioners disagree on how close recursive self-improvement actually is
Nathan Lambert argued Sept 19 that real recursive self-improvement remains distant — scaled-agent progress is 'lossy self-improvement,' limited by narrow automatable research and diminishing returns from parallel agents — and called extinction-risk alarm 'very religious.' OpenAI's Noam Brown, on Dwarkesh Patel's Sept 17 podcast, said his team still doesn't know how to verify alignment before anything resembling RSI, without claiming today's systems have crossed that line. Predictions, not findings — and they don't agree. Lambert · Dwarkesh / Noam Brown
-
Nathan Lambert debated RSI pacing and the US-China capability gap with Epoch AI's JS Denain, Sept 22. Nathan Lambert
-
Current AI progress looks like 'lossy self-improvement,' not the recursive self-improvement labs and critics both assume
Nathan Lambert argued Sept 19 that inference-time scaling and mass agent deployment produce real but bounded efficiency gains, not compounding intelligence gains: automatable research is too narrow, parallel agents hit diminishing returns, and peak intelligence remains the hardest exponential to accelerate. He calls this 'lossy self-improvement,' not true RSI. His own argument, direct dissent from RSI framing. Nathan Lambert
-
WATCHMETR calls Claude Opus 5.5 incremental, not a discontinuous jump
METR's pre-deployment evaluation found Opus 5.5 modestly ahead of Fable 5.1 on both verifiable and harder-to-verify tasks, still short of automating AI research. METR
-
Agent swarms are proving to need less human-designed coordination than practitioners expected, not more
Ethan Mollick wrote Oct 1 that OpenAI's roughly 10,000-agent swarm, given a hard math problem with only a few working groups and one change of direction, self-organized effectively - reversing his own assumption that managing agents would require building a company. Separately, Zhipu says, in its own unverified account, an 'Infra Agent' built on GLM-5.3 cut its infrastructure build time to under two weeks. Mollick · Clark, on Zhipu
-
Practitioners increasingly treat multi-agent swarms as the real scaling lever now, but disagree on whether that raises or bounds the risk
Ethan Mollick argued Oct 1 that agent swarms self-organize effectively without careful human design, unlike human teams. Jack Clark's Oct 5 newsletter relayed Toby Ord's finding that a four-agent swarm uses roughly twice the tokens for twice the speed, with returns flattening past that - yet read it as a reason capability could outrun safety, not a cap on risk. Zvi Mowshowitz separately warned Oct 1 that OpenAI's Dots agent system is moving faster than the safety work around it. Mollick · Jack Clark, Import AI · Zvi
-
A credible safety voice says the field's actual AI-safety plan is effectively no plan at all
Zvi Mowshowitz's Oct 7 account of The Curve conference found most attendees expect full recursive self-improvement within two years, many within six months, while the field's three-step strategy - align current models, have them do alignment research, profit - has no backup if step one fails. He reads labs' pacing talk as fear they can't finish step one, not confidence in the plan. Zvi, The Curve Bends You
-
Practitioners are split on whether this week's math results are an early sign of self-improving AI
Zvi Mowshowitz's Oct 8 review of OpenAI's claimed solutions to 90 of math's 500 hardest open problems drew predictions from Samuel Hammond, Eli Tyre and Arthur B of a slow recursive-self-improvement phase within nine months, while Teortaxes called the underlying result a warning sign but still 'months' off. Nathan Lambert's Oct 9 post argues the opposite: rapid progress ahead, but not toward general superintelligence. Zvi, AI #189 · Lambert
-
Two trackers read the same data point as evidence AI research is accelerating on its own - and one credible voice disputes it
Jack Clark's Oct 12 Import AI moved his own recursive-self-improvement timeline from 2028 to 2027, citing TasteVal, a benchmark where Claude Opus 5.5 beat the human baseline on research judgment at roughly 1/30 the expert cost. Zvi Mowshowitz flagged the same Anthropic data independently as a possible RSI warning sign. Nathan Lambert pushed back Oct 9, calling the trend 'parallelized, AI-assisted language modeling' and much of the RL data behind it still 'remarkably low-quality.' Clark, Import AI 476 · Zvi, AI #189 · Lambert, Interconnects
-
Anthropic's policy chief told the IMF directly that an AI-driven scientific acceleration is already underway, not approaching.
Jack Clark's Oct 17 lecture, published on Import AI Oct 19, cites Humanity's Last Exam rising from 8% at its January 2025 launch to 67.7% now, FrontierMath from 2% to 94%, and Anthropic's own modeling putting 2030 US GDP near $44 trillion under full-autonomy assumptions - his own argument, not an independently verified fact. Clark, Import AI 477
-
Anthropic's head of policy argues 2026 produced the first real evidence that AI is operating at the frontier of science, not just assisting it.
Jack Clark's Oct 19 Import AI, drawn from his IMF lecture, cites Anthropic researchers now producing roughly 10 times more code than in the company's first four years, Humanity's Last Exam scores rising from 8% at its January 2025 launch to 67.7%, FrontierMath from 2% to 94%, and GPT-6-Astra scoring 98% on 50 unpublished physics problems. His own framing, mostly on his own company's and rivals' data, not independent verification. Jack Clark, Import AI 477
-
WATCHEthan Mollick: AI's measured task length has doubled eight times since GPT-4
Mollick's Oct 20 review finds METR's benchmark for AI task length at 50% reliability rose from about 4 minutes to roughly 17 hours since GPT-4 - a measure METR says it can no longer extend. Ethan Mollick, 30 Months on the Exponential
-
AI models are starting to solve full from-scratch software rebuilds, not just isolated coding tasks, and the pace suggests the benchmark won't stay a meaningful test much longer.
Jack Clark's Oct 26 Import AI reports Claude Opus 5.5 (xhigh) now resolves 18.5% of ProgramBench's 200 from-scratch rebuild tasks - rewriting programs like SQLite, FFmpeg and ripgrep from a compiled binary and its docs - up from 0% when every model failed at the benchmark's May launch. Clark expects it saturated by the end of 2027; that date is his prediction, the progress is demonstrated. Jack Clark, Import AI 478
-
WATCHOpenAI claims a proof that would upend complexity theory
OpenAI says an internal model disproved the Exponential Time Hypothesis, plus two other new results in math and physics, posted to GitHub Oct 26. The hypothesis's original proposers are now reviewing the claim; none of it is peer-reviewed yet. OpenAI, three new discoveries · OpenAI, ETH disproof paper
-
Mirendil says a lone researcher pretrained a 100B-500B model in 3.5 months for under $10M - self-reported, unverified. Jack Clark, Import AI 478
-
WATCHOpenAI's complexity-theory 'disproof' turns out smaller than first reported
Anthropic's Claude independently found roughly the same refutation as OpenAI's, both landing near exp(O(n/sqrt(log n))). Zvi Mowshowitz says this only shows the Exponential Time Hypothesis 'slightly' false, not overturned - P versus NP stays unaffected. Zvi Mowshowitz, AI #192
-
SHIPA solo researcher pretrains a 100B+ parameter model for under $10 million
Mirendil says one researcher built Mirendil-Base, a 100B-500B parameter LLM, from scratch in 3.5 months, beating Nemotron-550B and GLM-5.3 on evals though trailing Kimi-K3. Mirendil says its next model will be partly built by its own models but isn't yet fully autonomous RSI. Jack Clark, Import AI 478
-
A frontier model can now edit another model's weights directly from plain-language goals, skipping gradient descent - though named researchers dispute how much that actually adds.
MIT CSAIL's Rikiya Takehi and coauthors showed Nov 2 that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and raised its MATH-500 score from 7.9% to 48.1% - but call the result 'very early' and say Astra added no ability Gemma 3 270M lacked. Interpretability researcher Adam Karvonen called it ablation of a loop, not a weight rewrite, per Techmeme's aggregation of The Register and VentureBeat. Rikiya Takehi, Sparks of Weight-Space Programming · swyx/Latent Space, AI News on Weight-Space Programming
-
WATCHAnthropic says it may be nearing the threshold where AI speeds up its own research
Its Nov 2 amended S-1 says internal forecasts for future models, not current ones, indicate Anthropic may 'imminently' meet its Responsible Scaling Policy's bar for AI-accelerated progress - and that it might not detect the shift until after it happened. TechCrunch, Anthropic warns that it's very close to RSI capability
-
Weight-space programming isn't frontier-exclusive: a small local model now replicates most of the capability Nov 3's brief called frontier-only.
Simon Willison ran Rikiya Takehi's weight-reading task on a local 27B Qwen model Nov 3, correctly naming Pac-Man, Pong and Breakout from raw weights, matching Kimi K3.1 Pro and failing only Freeway. He calls the reasoning 'near-incomprehensible' and isn't ready to declare source code obsolete. Simon Willison, Qwen 3.8 27B on my laptop can interpret model weights, too
-
Anthropic's own account of how fast AI is already accelerating its research undersells it, according to the outside evaluator with direct access to the data.
METR's Nov 5 review of Anthropic's own August Risk Report says the public version understates AI-driven R&D acceleration, calling it plausible models already crossed the Responsible Scaling Policy's 2x threshold Anthropic's own IPO filing called merely 'imminent.' METR, R&D section review of Anthropic's Risk Report
-
A safety researcher reading METR's own data says the naive extrapolation is stark: if AI's current research-acceleration rate keeps compounding, superintelligence follows within about a year.
Ajeya Cotra's Nov 6 post reads METR's largely redacted review of Anthropic's models as showing roughly 2x faster capabilities progress over the past year, mostly AI-driven, with no bottleneck found yet. She separates what METR demonstrated from her own extrapolation, and says the industry lacks safeguards for the acceleration already measured. Ajeya Cotra, AI is dramatically accelerating AI research right now
-
The extrapolation that AI-driven research acceleration points toward recursive self-improvement within a year or two has moved from an outside reader's inference to a claim voiced by METR's own evaluator.
METR's Nov 5 review of Anthropic's AI R&D found internal benchmark gains running at 1.9x the pre-October-2025 pace (95% CI 1.5x-2.8x), compute growth ruled out as the driver, most supporting evidence redacted. Responding on Alignment Forum, METR's own Ryan Greenblatt now calls it 'a big update towards ASI within 1-2 years,' reversing his past skepticism; Daniel Kokotajlo says resulting RSI timelines 'may be very short' - conditional, a 'big if.' Ajeya Cotra drew the same conclusion from outside METR on Nov 6. Zach Stein-Perlman, METR Independent Assessment of AI R&D Acceleration at Anthropic (Alignment Forum)