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

Forum topic · 小凯 · 2026-05-13

Summary

This forum post compiles evidence that automated AI research and recursive self-improvement (RSI) are becoming concrete near-term trends. Anthropic co-founder Jack Clark estimated a 60% probability that RSI arrives before the end of 2028, while an OpenAI researcher cited an internal estimate of March 2028. The post argues AI research is mostly writing, coding, and math—exactly where large language models excel—citing SWE-Bench scores jumping from ~2% (2023) to 93.9% (2026), labs' rising code self-sufficiency, Google DeepMind's AlphaEvolve closing a recursive loop by speeding Gemini training kernels 23%, and Andrej Karpathy's 630-line open-source AutoResearch agent delivering validated training improvements overnight. It also covers the emerging machine economy (Virtuals Protocol metrics, Stripe's agentic commerce push, x402/ACP protocols), job displacement analysis, and skeptical counterpoints such as Andrew Trask's scaling-laws argument that compute, data, and talent—not cleverer code—remain the bottleneck. Conclusion: incremental recursive improvement has already begun, and humans should focus on problem definition, theoretical intuition, and governance.

Key points

Jack Clark's 60% estimate (May 2026): Anthropic co-founder Jack Clark stated on X that he believes recursive self-improvement (RSI) has a 60% probability of occurring before the end of 2028 — i.e., AI systems may soon be capable of building AI systems. An OpenAI researcher (Adrien Ecoffet) replied that OpenAI's public estimate is that automated AI research arrives around March 2028.
Why AI research suits LLMs: The post argues AI progress itself consists of writing text, writing code, and some math — the core competency of large models. Literature review, experiment reproduction, code implementation, and hyperparameter search are already mature or maturing; architecture innovation is partially feasible (AlphaEvolve); theory and problem definition remain human strongholds.

Evidence chain

SWE-Bench trajectory: ~2% (Claude 2, 2023) → ~10% (GPT-4, 2024) → ~25% (Claude 3.5 Sonnet, 2025) → 93.9% (2026). Despite benchmark-version debates, the trend is undeniable.
CORE-Bench: benchmarks where AI reproduces other research papers — a core RSI capability.
Lab self-sufficiency: Anthropic claims most code is written by Claude Code; OpenAI says GPT-5.3-Codex helped build itself; DeepMind uses AlphaEvolve to optimize Gemini training. One cited figure: 5 employees × ~10,000 agents each = 50,000 AI agents building better AI at Anthropic.

AlphaEvolve: the recursive flywheel

| Domain | Achievement | |---|---| | Math | Broke Strassen's 56-year matrix multiplication record (49 → 48 multiplications) | | Chips | Directly modified next-generation TPU silicon designs | | Energy | Data center scheduling savings recovering ~0.7% of global compute | | Training | Sped up a Gemini kernel by 23%, cutting training time ~1% | | Collaboration | Helped Terence Tao on open problems |

The crucial point is the closed loop: AlphaEvolve (Gemini-driven) accelerates Gemini training, and the faster Gemini powers the next AlphaEvolve. Commentators note RSI is starting not as a discontinuous leap but as incremental 1% kernel-level gains compounding across model generations.

Karpathy's AutoResearch

A 630-line Python script that autonomously runs LLM training experiments overnight: modify code, train 5 minutes, keep improvements, discard failures, repeat.
Results: initial overnight run — 83 experiments, 15 kept, val_bpb 1.000 → 0.975; a 2-day extension ran ~700 experiments with ~20 fully additive improvements that transferred to larger models; production impact reduced Time-to-GPT-2 from 2.02h to 1.80h (11% speedup). Shopify CEO Tobi Lutke's run: 37 experiments, 19% validated improvements, a 0.8B model beating a 1.6B model.
Its README contains a tongue-in-cheek prophecy about a future where research is done entirely by autonomous AI agent swarms on compute megastructures, with self-modifying code beyond human comprehension.

The machine economy

Virtuals Protocol Q1 2026: $479M agent GDP, 18,000+ deployed agents, 1.77M+ completed tasks. AI token total market cap: $14.17B (CoinGecko). McKinsey projects a $5T agent-payment market by 2030; Coinbase projects $20T in AI+Web3 GDP additions.
Stripe (Sessions 2026) is pushing agentic commerce; president John Collison predicts agent-as-buyer becomes mainstream in 12–18 months. Companies incorporated via Stripe Atlas in 2026 are reportedly earning 5× their year-ago counterparts, driven by AI coding tools lowering startup barriers.
Payment infrastructure: x402 (Coinbase, pay-per-request over HTTP 402), ACP v2 (Virtuals, on-chain escrow for agent commerce), Circle Gateway (~$0.00001 per transaction), USDC as the dominant settlement currency (~99% of agent payments).

Labor impact

Already/soon automated: literature review, experiment reproduction, code implementation, hyperparameter search, benchmarking reports.
Currently automating: training pipeline optimization, kernel/low-level optimization, data center scheduling, model evaluation.
Human moats: problem definition, theoretical intuition, cross-domain association, value judgments, physical-world operations.

Skeptical view

Andrew Trask (OpenMined) argues scaling laws remain the better predictor: capability growth requires synchronized growth in data, compute, and talent/algorithms. "AI writing its own JAX code doesn't necessarily change that" — RSI may amount to a skilled-labor layoff plus cheaper tokens, not a qualitative change in capability growth rates.
Physical ceilings remain: high-quality data exhaustion, diminishing compute cost reductions, power/chip/cooling constraints.
The plausible path is gradual recursive improvement (1% gains compounding over a decade), not a sci-fi hard takeoff.

Deeper questions

Crossing the Rubicon: from tool to agent, from human speed to machine speed, from designed behavior to emergent behavior.
The post's closing question: if AI can automate AI research, who decides what AI should research? — a value/politics question, not a technical one.
Practical advice: understand AI's capability boundaries, redefine human value (creativity, judgment, cross-domain insight, emotional connection — not just "physical-world barriers"), and participate in AI governance rather than waiting to be displaced.

Sources cited

Jack Clark, Import AI 455: Automating AI Research (2026-05-07)
Google DeepMind, AlphaEvolve annual report (2026-05)
Andrej Karpathy, autoresearch (GitHub, 2026-03-07)
Stripe Sessions 2026, Agentic Commerce
IEEE Spectrum, Recursive Self-Improvement Edges Closer (2026-05-08)
TechFastForward, The Bootstrapping Has Begun (2026-04-14)