AI Lurches Toward Recursive Self-Improvement

AI’s move toward recursive self-improvement may be accelerating, with Sundar Pichai expecting progress over the next 12 months as AI helps speed up training and model launches. An MIT team led by PhD student Rukia Takahi reported that GPT 6 Astra directly edited the weights of the smaller QEN 3.5 0.8b model, changing 0.016% of them. The edited model’s math benchmark score rose sixfold and its instruction-following score threefold. Takahi said Astra identified and removed reasoning loops using 50 forward passes on practice problems, without gradients or additional training.

The MIT team also reported that Astra could infer a network’s trained purpose from its weights and write small game-playing networks from scratch that outperformed networks trained with PPO for 100 million steps. Commentators differed on what the results establish: Wei Aizhu attributed the math gain to fixing a loop bug rather than improving general mathematical ability, and Alex Crensell stressed that the highlighted network had only 20,486 parameters, compared with more than a trillion in frontier models. Nick Dobos considered writing weights from scratch potentially more significant than the bug fix, while Yuchen Jin described weight editing as “gene editing for neural networks.”

Crensell characterized the work as model reasoning extending interpretability-style analysis and said testing runtime self-improvement was an obvious next step, although he considered it infeasible for current models and possibly achievable within a few generations. Tenebrous predicted a mechanism like this within a year, reasoning that models capable of broad optimization could also optimize themselves. Jack Clark called recursive self-improvement potentially the world’s most important technical trend, but the reported experiments do not by themselves establish that frontier models can autonomously improve themselves.

Anthropic’s S1A says internal estimates indicate it may imminently reach its RSI danger threshold: AI-attributable progress equivalent to two years of the 2018–2024 baseline occurring in one year. Anthropic does not expect to use CLAW to autonomously design and build successive frontier-model generations without meaningful human supervision or understanding. It is considering updates to its responsible-scaling policy intended to keep humans in control, underscoring that faster AI-driven progress raises governance questions even before fully autonomous model development is feasible.

The economics of AI use remain uncertain alongside the technical pace. Elon Musk said the cost of serving Grokbot was unsustainable relative to revenue after users reported hitting weekend usage limits, but did not specify which plans were affected. User Shrek said the issue was an unannounced reduction in service, not selecting the wrong subscription. One analysis this summer estimated that a month of use on a $200 ChatGPT Pro plan would cost about $14,000 at regular API prices; the narrator thought the analysis came from SemiAnalysis. Nick Huber warned that if venture funding dries up, prices could rise, users leave, and valuations fall, while the discussion left open whether customers would pay enough for tokens at their natural price.

Higher API prices may not sharply reduce usage: some customers could remain on $20–$30 plans using less capable models, while others may pay for intelligence that creates sufficient value. AI APIs were also described as far cheaper than comparable human labor in fields such as recruiting and management, though the scale of power-user activity subsidized by venture capital remains a key uncertainty. On GDPX, which measures economically valuable tasks in areas including chip design, finance, and biology, Opus 55 scored 12% at $51 per task, while Gemini4 Carbon scored 10.9% at $16. Chubbyonx characterized Gemini4 as about four times cheaper for a 1.1-percentage-point performance reduction, suggesting that more efficient models could make personal agents more cost-effective.

Practical task performance may matter more to adoption than headline benchmarks. OpenAI reported that GPT 6 Astra scored 80.6% on its LifeVal benchmark of everyday personal-agent tasks, including finding childcare, budgeting, and ordering groceries, up from GPT 5.2’s 49.3%. Frasser X argued that such tasks may matter more to most users than flashy benchmarks, and the discussion suggested that optimizing for them could improve capability and cost-effectiveness, as it did with coding. Meta’s Muse may similarly target businesses: Olivia Moore cited 3.64 billion daily active users across Meta apps and more than 500 million small and medium-sized businesses using Meta for marketing. Meta plans connectors to tools including QuickBooks, Canva, and Stripe, potentially making Muse an operating layer for small businesses. Yet enterprise AI use remains well below its potential, constrained by adoption difficulties and corporate inertia as well as model capability, so faster progress does not guarantee faster revenue growth.