# Self Improving AI is getting wild Source: https://www.youtube.com/watch?v=TCDpDXjpgPI Recap page: https://rapidrecap.app/video/TCDpDXjpgPI Generated: 2025-10-28T04:02:27.692+00:00 --- ## 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. ![Screenshot at 1:04: The video transitions to a plot titled "Scenario: Intelligence Explosion" projecting compute requirements leading to Superintelligence, highlighting the HGM's lineage line crossing the threshold for automated AI research.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-01-04.png) **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. ## Detailed Analysis The video explains the concept of recursive self-improvement in AI agents, focusing on a paper by Jürgen Schmidhuber's team introducing the Huxley-Gödel Machine (HGM). The HGM overcomes the limitation of previous methods that only maximized benchmark scores by employing a Clade-Metaproductivity (CMP) metric, inspired by Julian Huxley's concept of clades, to estimate the long-term potential of self-modifications. This metric allows the agent to guide its evolution down productive lineages, avoiding stagnation where agents might score high locally but fail to generalize or improve further. The DGM Archive Tree diagram visually represents this process, showing how the final best agent achieved a 200% improvement along its lineage compared to the initial 100% performance baseline. The HGM achieved superior results, solving 61.4% of tasks on SWE-Verified, surpassing human-designed agents built on GPT-5-mini, and did so with significantly less compute (517 CPU-hours vs. 1231 hours for the previous DGM agent). The speaker also touches upon Andrej Karpathy's analogy illustrating how LLMs tokenize language into unique symbols (like emojis), emphasizing that these models must learn meaning from scratch, thus requiring empathy for their complex internal processing. Finally, the video promotes 10Web's AI Website Builder, which utilizes AI to generate functional, customizable WordPress sites instantly via chat prompts, offering full WordPress backend control and white-labeling capabilities. ### Huxley-Gödel Machine (HGM) Paper Abstract - Self-improvement occurs through agents editing their own codebases, growing a tree of self-modifications; CMP metric aggregates performance of descendants to estimate potential; HGM introduced to simulate Gödel Machine behavior under certain assumptions; HGM outperforms prior methods on SWE-bench using less wall-clock time. ### DGM Archive Tree Analysis - The tree illustrates lineage improvement, with the best agent showing a 200% improvement relative to the starting point (100%); Agents that die off (red/yellow nodes) indicate unproductive paths, contrasting with the successful green lineage. ### LLM Tokenization Analogy (Karpathy Tweet) - LLMs process text as unique tokens (like emojis), requiring them to learn meaning from scratch based on training data statistics, urging empathy when querying them. ### 10Web AI Website Builder Promotion - Demonstrates an AI website builder that generates sites instantly from a prompt, offering a fully integrated WordPress solution with hosting, SEO tools, user management, and white-label options for agencies. ![Screenshot at 0:04: Initial stick figure climbing exponentially increasing stairs, illustrating the concept of accelerating progress.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-00-04.png) ![Screenshot at 0:32: Chart titled "Scenario: Intelligence Explosion" projecting compute growth towards Superintelligence, marking the HGM's lineage path.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-00-32.png) ![Screenshot at 1:06: Jürgen Schmidhuber's tweet discussing recursive self-improvement and Gödel Machines, crediting earlier concepts.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-01-06.png) ![Screenshot at 2:58: The DGM Archive Tree diagram showing the branching evolution of coding agents, color-coded by task success \(red=10 tasks, yellow=60 tasks, green=200 tasks\).](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-02-58.png) ![Screenshot at 4:41: Advertisement slides for 10Web, highlighting its open-source nature, powering 43% of the web, and featuring SEO/User Management capabilities.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-04-41.png) ![Screenshot at 5:58: Demonstration of editing the website structure via chat interface: "Add a hover animation to the hero section".](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-05-58.png) ![Screenshot at 7:35: The abstract of the Huxley-Gödel Machine paper, highlighting key terms like 'self-improvement', 'clade', and 'Meta-productivity-Performance Mismatch'.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-07-35.png) ![Screenshot at 9:36: DGM Progress chart showing the 'Lineage to Final Best Agent' spiking dramatically after implementing 'History-aware patch generations and ranking' around iteration 60.](https://ss.rapidrecap.app/screens/TCDpDXjpgPI/00-09-36.png)