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

Episode 1 · Researchers Push Back on Dario's Claim That RSI Has Begun (2026-09-12, 7 posts)

On September 12–13, the developer community engaged in a concentrated debate over whether "Recursive Self-Improvement (RSI)" has already arrived, with multiple developers concluding it has not: current models cannot autonomously generate research ideas and can only marginally assist with experiments, while the real priority should be risks from AI misuse.

Confirmed

• tszzl stated flatly that RSI "hasn't arrived at all": today's models cannot autonomously produce research ideas and can only marginally help run experiments—something entirely different from autonomous research. Even if private models like OpenAI/Anthropic's Claude might be stronger, he does not believe RSI has been achieved. sebkrier reshared in support, saying he has held a similar view for months.
• johncrickett questioned the claim that "the threat comes from RSI": RSI has not been realized and may never be; meanwhile, AI already has many real risks and misuse problems that deserve priority attention.
• gerardsans responded that debating AI hype terminology is futile—RSI, like AGI and ASI, is a meaningless Silicon Valley dinner-table acronym whose only function is to shield investors from real technical questions. He argues there is no "self," "recursion," or "improvement," only iterative optimization or evolution.
• gerardsans also argued technically that RSI is a pseudo-concept: software is not smooth or infinite; code is done once it meets specs. AI is probabilistic but still software—once the network layers, inputs/outputs, and parameters are set, they are fixed; transformers have a capacity ceiling and cannot be trained indefinitely—feeding in more data only leads to catastrophic forgetting.

Why it matters

• The debate reflects a split in the developer community over AI narratives: one camp (tszzl, sebkrier) evaluates autonomous research capability based on research practice, while the other (johncrickett, gerardsans) argues for shifting attention from distant or illusory concepts like RSI/AGI toward present-day misuse and safety risks.
• gerardsans's technical arguments (model capacity ceilings, catastrophic forgetting) provide a concrete engineering perspective for the position that "RSI is not feasible," though these claims themselves remain contested within 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

Episode 2 · Mollick: OpenAI and Anthropic Hint Recursive Self-Improvement Has Begun (2026-09-13, 3 posts)

Ethan Mollick notes both Anthropic and OpenAI made their clearest statements yet that some form of recursive self-improvement (RSI) has been achieved, though early-stage; he cautions RSI fears may be overblown given possible hard limits in compute, architecture, and AI-driven research itself.

• 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