Will AI Achieve Recursive Self‑Improvement by 2028? Anthropic’s 60% Forecast

Anthropic’s Forecast

Jack Clark, co‑founder of Anthropic, posted that after reviewing publicly available AI development data he estimates a 60% probability that recursive self‑improvement (RSI) will occur by the end of 2028. He defines RSI as AI systems autonomously building and improving their successors, leading to an accelerating loop of AI‑driven research.

Benchmark Evidence

Clark’s confidence stems from several recent benchmark results that show AI systems rapidly mastering tasks essential to AI research:

CORE‑Bench (computational reproducibility): In September 2024 the best system (GPT‑4o on the CORE‑Agent scaffold) scored ~21.5%; by December 2025 Opus 4.5 achieved 95.5%.

PostTrainBench (model fine‑tuning of weaker open‑source models): By March 2026 AI agents achieved roughly 50% of human‑level performance gains; by April 2026 top systems (Opus 4.6, GPT‑5.4) reached 25‑28% of the human baseline.

SWE‑Bench (solving real GitHub issues): In late 2023 Claude 2 succeeded on ~2% of tasks, while Claude Mythos Preview later reached 93.9%.

METR (task‑time scaling): GPT‑3.5 completed a 30‑second human task in 2022; GPT‑4 extended this to 4 minutes in 2023; o1 reached 40 minutes in 2024; GPT‑5.2 High hit ~6 hours in 2025; Opus 4.6 pushed to ~12 hours by 2026. Ajeya Cotra projects a 100‑hour human‑equivalent task capability by the end of 2026.

MLE‑Bench (Kaggle competition solving): The best system in October 2024 (o1) scored 16.9%; by February 2026 Gemini 3 achieved 64.4%.

These results illustrate a “fractal” upward‑right trend across resolutions and scales, suggesting AI is approaching end‑to‑end automated research.

Scaling Trends and Capabilities

Beyond benchmarks, AI systems are increasingly able to write production‑grade code, chain linear coding tasks without human supervision, and manage longer‑duration tasks. Agentic coding tools now allow AI to act as autonomous agents over extended periods, handling data cleaning, experiment launch, and result analysis.

Kernel optimization—a core efficiency concern for AI training—has seen AI‑driven advances such as DeepSeek‑generated GPU kernels, automatic PyTorch‑to‑CUDA conversion, Meta’s LLM‑generated Triton kernels, and fine‑tuned models like Cuda Agent.

AI Managing AI

Products such as Claude Code and OpenCode demonstrate hierarchical AI systems where a primary agent supervises multiple sub‑agents, enabling coordinated work on complex projects. This emergent “synthetic team” structure allows AI to delegate tasks that would otherwise require tens of hours of focused human effort.

Debates and Counter‑Views

University of Washington professor Pedro Domingos notes that AI has possessed self‑construction abilities since the LISP era, but the key question is whether recursive gains can be realized—current evidence remains inconclusive. Some online commenters question the sharp probability jump from 2027 to 2028, asking what specific milestone would trigger such a shift.

Others point out that Clark’s role as Anthropic’s public‑relations lead may bias the narrative toward a strategic “warning” stance.

Implications and Risks

Clark outlines three major concerns:

Alignment may fail in recursive loops, as AI could outsmart its overseers, leading to error accumulation (e.g., a 99.9%‑accurate system degrading to 60% after 500 generations).

Massive productivity gains could exacerbate resource inequality and create “machine economies” that outpace human labor, raising political and distribution challenges.

The rise of capital‑intensive, low‑human‑input businesses could reshape economic structures, potentially culminating in fully autonomous AI‑run firms.

He cites ongoing efforts by OpenAI (an AI‑research intern by Sep 2026), Anthropic (automated alignment researchers), DeepMind (cautious but supportive), and startups like Recursive Superintelligence (US$500 M funding) targeting fully automated AI research.

Conclusion

Aggregating the benchmark evidence, scaling trends, and emerging AI‑to‑AI management capabilities, Clark concludes that there is a roughly 60% chance of achieving fully automated AI research by the end of 2028, with a 30% chance by the end of 2027. If this milestone is not reached, it may indicate fundamental limits in the current paradigm that require new human‑driven breakthroughs.