Self Improving AI is getting wild

Quick Overview

The development of self-improving AI agents, exemplified by Jürgen Schmidhuber's team's Huxley-Gödel Machine (HGM), demonstrates recursive self-improvement works at scale, achieving human-level performance on coding benchmarks like SWE-Bench Lite by iteratively modifying its own code and using a custom metric (CMP) to guide improvements, which is significantly more efficient than previous benchmark-maxing methods.

Key Points: Jürgen Schmidhuber's team demonstrated recursive self-improvement in coding agents using the Huxley-Gödel Machine (HGM), achieving human-level performance. The HGM uses a custom metric called Clade-Metaproductivity (CMP) to estimate the potential of self-modifications, avoiding the pitfalls of only optimizing for immediate benchmark scores. The HGM achieved a 200% performance increase over its initial ancestor on the DGM Archive Tree lineage by iteratively improving its own code. The HGM agent surpasses the best human-designed agent built on GPT-5-mini on the SWE-Verified leaderboard by solving 61.4% of tasks. The HGM approach is significantly more efficient, requiring only 517 CPU-hours on SWE-Verified compared to 1231 hours for the DGM agent. The research highlights the importance of tracking agent 'family trees' (lineages) to find truly productive self-improvements, rather than just maximizing benchmark scores. Andrej Karpathy's tweet about LLM tokenization is used to illustrate why agents must learn language from scratch, as they see text as sequences of unique emojis (tokens) rather than letters.

Context: The video discusses recent breakthroughs in Artificial Intelligence research, specifically focusing on recursive self-improvement in coding agents, as highlighted by a tweet from Alex Duffy referencing work by Jürgen Schmidhuber's team. The core concept revolves around the Huxley-Gödel Machine (HGM), an agent capable of rewriting and improving its own code, contrasting this with older methods that only optimized for current benchmark scores.