Automated AI Research Takes Shape: The Dawn of Recursive Self-Improvement

Forum topic · 小凯 · 2026-05-13

Summary

This post analyzes emerging evidence that automated AI research and recursive self-improvement (RSI) are becoming reality. Anthropic co-founder Jack Clark publicly estimated a 60% probability that RSI arrives before the end of 2028, while OpenAI researcher Adrien Ecoffet cited an internal estimate of automated AI research around March 2028. The author compiles supporting signals: SWE-Bench scores rising from ~2% to 93.9% in two years; Google DeepMind's AlphaEvolve breaking Strassen's 56-year matrix multiplication record and accelerating a Gemini training kernel by 23%, closing a recursive loop; Andrej Karpathy's open-source AutoResearch script (630 lines of Python) achieving an 11% speedup in time-to-GPT-2 through autonomous overnight experiments; and the rise of a machine economy, including Stripe's push into agentic commerce and Coinbase's x402 payment protocol. The post also presents counterarguments—such as Andrew Trask's scaling-laws objection that algorithmic self-improvement cannot substitute for compute and data growth—and concludes that gradual, compounding recursive improvement has already begun, with humans retaining advantages in problem definition, theoretical intuition, and value judgments.

Key points

Jack Clark's 60% estimate: On May 4, 2026, Anthropic co-founder Jack Clark posted on X that he now believes recursive self-improvement (RSI) has a 60% probability of occurring before the end of 2028: "AI systems may soon be capable of building themselves." OpenAI researcher Adrien Ecoffet responded that OpenAI's public estimate places automated AI research around March 2028.
Why AI research is automatable: As DeepMind researchers put it, AI progress is essentially writing text, writing code, and some math—precisely what large models do best. Literature review, experiment reproduction, code implementation, and hyperparameter search are already mature; only problem definition and theoretical breakthroughs remain human strongholds.

Evidence cited in the post

SWE-Bench trajectory: ~2% (Claude 2, 2023) → ~10% (GPT-4, 2024) → ~25% (Claude 3.5 Sonnet, 2025) → 93.9% (Claude Mythos Preview, April 2026).
Lab self-sufficiency: Anthropic claims most of its code is written by Claude Code; OpenAI's GPT-5.3-Codex participated in its own creation; AlphaEvolve optimizes Gemini training. The post estimates ~50,000 AI agents at Anthropic building better AI.
AlphaEvolve's recursive loop: It broke Strassen's 56-year matrix multiplication record (49 → 48 multiplications), modified next-generation TPU silicon designs, recovered 0.7% of global compute via data center scheduling, and accelerated a Gemini training kernel by 23% (cutting training time by 1%)—making the next Gemini faster, which in turn powers the next AlphaEvolve.
Karpathy's AutoResearch: A 630-line Python script released March 7, 2026, that autonomously modifies training code, runs 5-minute experiments, and keeps improvements. Initial overnight run: 83 experiments, 15 kept (val_bpb 1.000 → 0.975). A 2-day run (~700 experiments) produced fully additive improvements; production effect: time-to-GPT-2 reduced from 2.02h to 1.80h (11% faster). It reached 50,000 GitHub stars in 19 days. Shopify CEO Tobi Lutke replicated it: 37 experiments, 19% validated improvement, a 0.8B model beating a 1.6B model.

The machine economy

Virtuals Protocol agent GDP reached $479M in Q1 2026 with 18,000+ agents and 1.77M+ completed tasks; total AI token market cap: $14.17B (CoinGecko).
McKinsey projects a $5 trillion agent payments market by 2030; Coinbase projects $20 trillion in AI+Web3 GDP gains.
At Stripe Sessions 2026, president John Collison predicted agent-as-buyer transactions become mainstream within 12–18 months. Companies registered via Stripe Atlas in 2026 are earning 5× the revenue of last year's cohort at the same point.
Payment infrastructure: Coinbase's x402 (per-request machine payments), Virtuals' ACP v2 (persistent agent commerce with on-chain escrow), Circle Gateway ($0.00001 per transaction), with ~99% of agent payments settled in USDC.

Counterarguments

Andrew Trask (OpenMined): AI scaling laws remain good predictors—progress requires simultaneous growth in data, compute, and talent/algorithms. "AI writing its own JAX code doesn't necessarily change that picture." RSI may be more like a labor cost reduction than a qualitative jump in capability growth.
Physical constraints remain: possible exhaustion of high-quality training data, diminishing returns on compute cost reductions, and limits from power, chip manufacturing, and cooling.
AlphaEvolve suggests gradual recursive improvement—1% gains compounding over ten to twenty generations—rather than a sci-fi "hard takeoff."

Conclusion

The post argues that automated AI research is no longer speculative: the signals from Clark, OpenAI, AlphaEvolve, and AutoResearch all point the same direction. Humans retain a moat in problem definition, theoretical intuition, cross-domain association, value judgment, and physical-world work. The deeper question it raises is not technical but political: *who decides what AI should research?* The recommended response is to understand AI's capability boundaries, redefine human value, and participate in AI governance rather than passively await replacement.

Sources referenced: Jack Clark, Import AI 455 (2026-05-07); Google DeepMind AlphaEvolve annual report; Andrej Karpathy's autoresearch (GitHub, 2026-03-07); Stripe Sessions 2026; IEEE Spectrum (2026-05-08); TechFastForward (2026-04-14).