Is Recursive Self-Improvement Here? Dario's Claim Sparks Debate

Episode 1 · Researchers Push Back on Dario Amodei's Claim That RSI Has Begun (2026-09-12, 8 posts)

From September 12–14, the developer community debated whether recursive self-improvement (RSI) has arrived, sparked by Dario Amodei's claim in 'Pacing AI' that RSI is beginning. Multiple researchers pushed back, arguing that system cards and current model capabilities do not support the claim, and that AI abuse risks deserve priority attention.

Confirmed

• Dario Amodei claimed RSI has begun in 'Pacing AI.' Gary Marcus, citing Ramez's analysis, countered that the blog post Dario referenced and the system cards of the latest Claude and ChatGPT models clearly show RSI has not occurred.
• tszzl stated RSI is 'not here at all': current models cannot autonomously produce research ideas and only marginally help run experiments; even if private models from OpenAI/Anthropic are stronger, he does not believe RSI is realized. sebkrier reshared in support, saying he has held a similar view for months.
• johncrickett questioned framing the threat as RSI: it has not been achieved and may never be, while AI already poses many real risks and abuse problems that should be prioritized.
• mariofilhoml offered an argument against doomers: extinction-level RSI is a far out-of-distribution task, and current LLMs are limited even on in-distribution tasks, so there is no reason to expect such a leap.
• gerardsans called AI hype terms futile: RSI, like AGI and ASI, is a meaningless Silicon Valley acronym that only shields investors from real technical questions; there is no 'self,' 'recursion,' or 'improvement,' only iterative optimization or evolution.
• gerardsans also argued RSI is technically a pseudo-problem: AI is still software with fixed layers, I/O, and parameters; transformers have capacity limits and continued training only causes catastrophic forgetting.

Why it matters

• The debate reveals a split in how the community reads AI narratives: one camp assesses autonomous research capability via system-card evidence and generalization arguments, while another urges focusing on present-day abuse and safety risks rather than distant or illusory RSI/AGI concepts.
• The clear divergence between a frontier-lab CEO and outside researchers interpreting the same evidence (system cards) suggests 'RSI has begun' is more a narrative judgment than an empirical conclusion.
• mariofilhoml's distribution-distance argument and gerardsans's engineering arguments (capacity limits, catastrophic forgetting) give concrete support to the 'RSI is infeasible or far off' position, though these claims remain contested in the community.

• Devs push back on RSI threat: real AI risks are misuse, not recursive self-improvement — johncrickett · 2026-09-12
• "No self, recursion or improvement": RSI called a meaningless hype acronym — gerardsans · 2026-09-12
• A technical case against RSI: transformers have capacity limits, more data causes forgetting — gerardsans · 2026-09-12
• Gary Marcus and Ramez rebut Dario's RSI claim: system cards show it hasn't started — GaryMarcus · 2026-09-13
• tszzl: models aren't autonomously producing research ideas — RSI is not here yet — i_dg23 · 2026-09-13
• Dario claims RSI is starting, but Claude and ChatGPT system cards say otherwise — paulnovosad · 2026-09-13
• Is RSI here? Researchers argue models still can't autonomously generate research ideas — i_dg23 · 2026-09-13
• A simple anti-doomer argument: LLMs can't even do near-OOD tasks, let alone RSI — mariofilhoml · 2026-09-14

Episode 2 · OpenAI and Anthropic Signal Early Recursive Self-Improvement Is Here (2026-09-13, 6 posts)

This week, both Anthropic and OpenAI gave their clearest signals yet that some form of recursive self-improvement (RSI) is already happening, albeit in an early stage. The observation was first made by Ethan Mollick and echoed by Anthropic CEO Dario Amodei, sparking an industry-wide debate.

Confirmed

• Dario Amodei said AI progress has been sharply accelerating since this summer, driven by AI increasingly participating in building next-generation AI — "AI building AI."
• Ethan Mollick argued RSI could rapidly boost AI capabilities, and the first company to achieve it may gain an insurmountable lead.
• Google DeepMind Chief Strategy Officer Jasjeet Sekhon said in a Berkeley RDI interview that massive AI infrastructure investment is essentially a bet on recursive self-improvement.

Counterpoints

• Mollick also cautioned that concerns about RSI may be overstated: capabilities may face hard ceilings in compute and architecture; AI-driven research has its own limits (e.g., recognizing which problems are worth solving); and companies may fail to effectively use AI-produced research.
• Even if early RSI is real, it does not automatically lead to an intelligence explosion.

Why it matters

• Two leading labs simultaneously hinting at early RSI is a rare alignment, possibly marking a new phase in AI capability growth.
• If true, the first mover could build an uncatchable moat — the underlying logic of today's infrastructure spending spree.
• But RSI's real-world impact remains constrained by compute, architecture and organizational capacity; an intelligence explosion is not a foregone conclusion, and evidence of AI's actual share in model R&D warrants watching.

• Anthropic CEO warns AI has been advancing 'drastically faster' since summer as AI helps build AI — Polymarket · 2026-09-13
• Mollick: Anthropic and OpenAI's clearest hints yet that some form of recursive self-improvement has been achieved — emollick · 2026-09-13
• Mollick: OpenAI and Anthropic hint at early recursive self-improvement in rare statements — emollick · 2026-09-13
• Ethan Mollick: RSI May Not Be the Whole Game — AI Research Faces Multiple Caps — emollick · 2026-09-13
• DeepMind Chief Strategist: AI Infrastructure Spending Is a Bet on Recursive Self-Improvement — rohanpaul_ai · 2026-09-13
• DeepMind CSO: AI infrastructure bet hinges on recursive self-improvement — RebeccaBellan · 2026-09-13

Episode 3 · Google Researcher Argues Intelligence Has a Speed Limit (2026-09-13, 5 posts)

Google research scientist Peyman Milanfar published "Intelligence Has a Speed Limit" on September 13, challenging the AI community's near-magical framing of recursive self-improvement (RSI) from a control-theory perspective. His conclusion: any self-improvement feedback loop has a safe speed limit. The essay circulated widely via reposts from @NathanpmYoung and @ArtificialOther.

Confirmed

• Milanfar's core argument: RSI is often depicted as a model improving itself, getting smarter, and rewriting itself in an ever-accelerating loop, but real-world engineering constraints behind turning algorithms into working systems limit such self-improvement.
• Control theory's counterintuitive result: every self-improvement loop has a safe speed ceiling, and pushing past it causes sudden instability. The basis is the small-gain theorem—a feedback loop is stable only if the product of loop gains stays below 1.
• @DanJeffries1 (September 14), assuming strong-sense RSI exists and is not just marketing, listed nine constraints that would flatten RSI into a hidden S-curve, including the subject-object problem (a flawed system observing and improving itself cannot easily verify real gains) and the physical-world problem (even virtual experiments obey physical laws).

Why it matters

• RSI underpins the "AI accelerates past humans" narrative; if its speed is bounded by control theory, related timelines may be systematically overestimated.
• Bringing mature control-engineering frameworks (small-gain theorem, loop stability) into AI capability debates offers a testable analytical tool beyond compute/data arguments.
• The cascade of reposts and extensions (e.g., Jeffries' nine-constraint list) shows the piece triggered concentrated community reflection on the intelligence-explosion hypothesis.

• Intelligence Has a Speed Limit: control theory caps recursive self-improvement — docmilanfar · 2026-09-13
• New article: Intelligence Has a Speed Limit — why RSI can't run as fast as you like — NathanpmYoung · 2026-09-13
• Intelligence Has a Speed Limit: Control Theory Caps Recursive Self-Improvement — ArtificialOther · 2026-09-13
• Nine hard limits that turn recursive self-improvement into a hidden S-curve — Dan_Jeffries1 · 2026-09-14
• Google's Milanfar: Recursive Self-Improvement Has a Speed Limit — Engineering Reality Explained — docmilanfar · 2026-09-14

Episode 4 · AI circles hype RSI: intelligence explosion or fear-mongering business (2026-09-15, 2 posts)

After AGI, AI circles are hyping RSI (recursive self-improvement), where AI writes code and improves its own training for ever-stronger successors; articles trace the idea back to I.J. Good in 1965 and examine the debate between explosion claims and fear-driven business interests.

• RSI, AI's New Buzzword: Recursive Self-Improvement or a Fear-Driven Business? — APPSO · 2026-09-15
• RSI hype sweeps AI: recursive self-improvement or fear-driven business? — 创业邦 · 2026-09-17