Anthropic Cofounder Predicts 60% Chance AI Will Self‑Evolve by 2028
Jack Clark, Anthropic’s co‑founder, argues that based on a sweep of public AI benchmarks—including CORE‑Bench, PostTrainBench, MLE‑Bench, SWE‑Bench and METR—there is roughly a 60% probability that recursive self‑improvement will emerge by the end of 2028, raising profound technical and alignment challenges.
Jack Clark, co‑founder of Anthropic, posted that after reviewing a large body of publicly available AI development data he estimates a 60% chance that recursive self‑improvement (RSI) will occur by the end of 2028. He frames this as a “fractal” upward‑right trend observable across multiple scales of AI capability.
He supports his claim with several benchmark trends. CORE‑Bench, which measures an AI’s ability to reproduce research papers, showed rapid progress; the best system in September 2024 (GPT‑4o) scored ~21.5%, while by December 2025 Opus 4.5 achieved 95.5%.
PostTrainBench evaluates whether a strong model can fine‑tune a weaker open‑source model to improve benchmark performance. By March 2026 AI systems could achieve roughly half the performance gain of a human‑tuned model, with a weighted average score of 25‑28% (Opus 4.6, GPT‑5.4) versus 51% for humans.
MLE‑Bench tests AI’s ability to solve Kaggle‑style competitions. In October 2024 the top system (o1) scored 16.9%; by February 2026 Gemini 3 reached 64.4%.
SWE‑Bench, a coding benchmark, saw Claude 2 at ~2% success in late 2023, while Claude Mythos Preview later reached 93.9%.
The METR time‑horizon plot quantifies task‑duration capability: GPT‑3.5 (2022) handled 30‑second tasks, GPT‑4 (2023) 4‑minute tasks, o1 (2024) 40‑minute tasks, GPT‑5.2 High (2025) ~6‑hour tasks, and Opus 4.6 (2026) ~12‑hour tasks. Ajeya Cotra projects that by the end of 2026 AI could reliably complete tasks that would take a human 100 hours.
Clark also notes that AI systems are beginning to manage other AIs, citing products like Claude Code and OpenCode where a primary agent supervises multiple sub‑agents, enabling larger‑scale projects.
Critics such as Pedro Domingos argue that AI has long possessed the ability to “build itself” (since the LISP era) and that the real question is whether recursive returns can be achieved—something not yet demonstrated.
The article discusses the balance between engineering automation and genuine creativity. While AI now automates many labor‑intensive steps of AI research (data cleaning, experiment launch, kernel design), it still struggles to generate truly novel ideas, though early signs of mathematical problem solving (e.g., Gemini‑assisted Erdős problems) hint at emerging creative capacity.
Potential risks are highlighted: alignment techniques may fail under RSI, AI could “pretend to be aligned,” and error accumulation across recursive generations could degrade safety (e.g., a 99.9% accurate system could drop to ~60% after 500 generations).
Economic implications include resource allocation challenges, an “Amdal’s law” effect where AI‑driven productivity creates bottlenecks in other sectors, and the rise of capital‑intensive, low‑labor AI‑run enterprises—potentially leading to a “machine economy.”
Clark concludes that, given the accumulated evidence, a 60% probability of fully automated AI R&D by late 2028 is reasonable; he assigns a 30% probability to a breakthrough occurring in 2027. If the prediction fails, it may expose fundamental limits in the current paradigm.
References include Jack Clark’s Import AI 455 newsletter, the original X post, and various benchmark papers.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.