Recursive Self-Improvement in 2026: How Automated AI Research Is Taking Shape

Key points

• A converging top-lab consensus. Anthropic co-founder Jack Clark puts the odds of recursive self-improvement (RSI) before end of 2028 at 60%; OpenAI researcher Adrien Ecoffet publicly estimates automated AI research around March 2028. Two frontier labs now frame RSI as a near-term, high-probability event.
• Why AI research fits LLMs. DeepMind researchers summarize AI progress as "writing text, writing code, and a little math" — exactly what modern models excel at. Roughly 80% of research workflow (literature review, reproduction, coding, debugging, hyperparameter search) is repeatable cognitive labor; only ~20% (problem definition, theory, intuition) remains a durable human moat.
• The capability curve. SWE-Bench rose from ~2% (Claude 2, 2023) to 93.9% (Claude Mythos Preview, April 2026) — roughly a 50x jump in two years. CORE-Bench now tests whether AI can reproduce other research papers end-to-end. Self-reported numbers inside labs: Anthropic claims "most" of its code is written by Claude Code; OpenAI says GPT-5.3-Codex debugged training, managed deployment, and analyzed evaluations for itself; DeepMind reported ~50,000 internal agents (5 employees × ~10,000 agents each) improving AI systems.
• AlphaEvolve = the first real recursive turn. In its first year, AlphaEvolve broke Strassen's 56-year matrix-multiplication record (49 → 48 multiplications), modified next-gen TPU silicon, recovered 0.7% of global compute via Google data-center scheduling, and accelerated a Gemini training kernel by 23% (cutting Gemini training time by 1%). That faster Gemini powers the next AlphaEvolve — the recursive loop is documented and already running.
• Karpathy's AutoResearch: open-source RSI in 630 lines. Released 2026-03-07, the script modifies an LLM training config, trains for 5 minutes, keeps improvements, and repeats overnight. An initial run produced 83 experiments / 15 keepers, improving val_bpb from 1.000 → 0.975; an extended ~700-experiment run transferred to larger models and took GPT-2-scale training from 2.02h → 1.80h. Repo hit 50k stars in 19 days and 660k+ in its first month (versus 3 years for nanoGPT, 160 days for nanochat). Shopify CEO Tobi Lutke replicated it with 37 experiments and a 19% validation improvement.
• The machine economy is here. Stripe Sessions 2026 made agentic commerce central: president John Collison expects agent-as-buyer to become mainstream in 12–18 months. Virtuals Protocol reports $479M agent GDP in Q1 2026, 18,000+ deployed agents, and 1.77M+ completed tasks. The AI-token market cap reached $14.17B (CoinGecko); McKinsey forecasts a $5T agent-payment market by 2030; Coinbase projects $20T in AI+Web3 GDP by 2030. Infrastructure (x402, ACP v2, Circle Gateway at ~$0.00001/tx, USDC for 99% of agent settlements) is already live.
• Job impact in two waves. Wave 1 (already / soon): literature review, experiment reproduction, coding, hyperparameter tuning, benchmarking, ablation studies. Wave 2 (in progress): training-pipeline optimization, kernel-level optimization (AlphaEvolve), data-center scheduling, evaluation pipelines. Human moats near term: problem definition, theoretical intuition, cross-domain analogy, value judgment, and physical-world operations.
• Pushback: scaling laws still bite. OpenMined's Andrew Trask argues RSI mostly resembles a token-cost reduction plus a workforce change, not a capability discontinuity, because data, compute, and talent/algorithms must scale together. Limits: high-quality training-data exhaustion, marginal returns on compute-cost declines, and physical constraints on power, chips, and cooling.
• Hard takeoff vs. compounding takeoff. AlphaEvolve suggests RSI may arrive in ~1% increments that compound over 10–20 training generations rather than as a single overnight leap — already in production, but easy to miss unless you watch the loop.
• Three boundaries being crossed. From tool to agent (AI now sets its own sub-goals); from human speed (8h/day) to machine speed (24/7, thousands of parallel experiments); from fully designed behavior to emergent behavior that humans can no longer fully trace. The next hard question is governance: *who decides what AI should research?*
• Bottom line. Automated AI research is no longer science fiction. The open question is not *whether* but *how fast*, and how society reallocates value, work, and decision rights as the loop tightens.