# Horowitz Says AI Self-Improvement Still Lacks Four Key Conditions By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/34181 Published: 2026-10-17T20:59:30.000Z Updated: 2026-10-17T20:59:30.000Z Section: Technology > As of August 2026, Claude led 26% of Anthropic’s measured research work, but none of that work was fully autonomous. Andreessen Horowitz co-founder Ben Horowitz said rapid AI self-improvement remains unlikely because four conditions for a cycle of models building better successors have not converged. “Zero of those four have been met,” Horowitz said in a podcast episode released Friday, Oct. 16, at 9:51 p.m. ET (01:51 UTC Saturday, Oct. 17). The conditions are lower computing needs, shorter training time, stronger capabilities and repeated gains in later generations. Horowitz also pointed to more demanding model training and limits on graphics processing units (GPUs) and electricity. Anthropic’s measured research and development work offers a narrower picture of AI’s role in building AI. As of August 2026, Claude was rated in the “lead” category for 26% of that work. More than 90% reached at least the “collaborates” level, while none of the measured work was fully autonomous. Full recursive self-improvement means a model independently building its successor. Buildout estimates put data center construction at about two years, new power plants at four years or longer, and transmission projects at seven to 10 years. Existing generation and flexible demand could still accommodate substantial near-term growth, so those timelines describe a buildout challenge without establishing that rapid AI self-improvement is impossible. The infrastructure question connects to the broader debate over [AI infrastructure spending](). ## Links in this article - [AI infrastructure spending](https://www.tokenpost.com/news/business/26128)