# this EX-OPENAI RESEARCHER just released it... Source: https://www.youtube.com/watch?v=tUkD0oj92Qg Recap page: https://rapidrecap.app/video/tUkD0oj92Qg Generated: 2026-03-10T14:04:02.997+00:00 --- ## Quick Overview Andrej Karpathy released an open-source "autosearch" project that allows an AI agent to iteratively improve itself by running experiments, which he claims is the next phase of AI research, potentially leading to an intelligence explosion faster than previously expected. **Key Points:** - Andrej Karpathy, an ex-OpenAI researcher, released an open-source "autosearch" project designed to let an AI agent iteratively improve its own LLM training setup. - The agent modifies code, trains for 5 minutes, checks validation loss, keeps or discards changes, and repeats the process autonomously overnight. - Karpathy confirmed that tuning the 'nanochat' model using this method resulted in a 11% improvement in its 'Time to GPT-2' leaderboard score, dropping from 2.02 hours to 1.80 hours. - The project is intentionally kept small, relying on three main files: 'prepare.py' (data prep/utilities), 'train.py' (the single file the agent edits), and 'program.md' (baseline instructions for the agent). - The author notes that this automated research approach is a first for him, contrasting it with his previous manual, iterative optimization of neural network training over two decades. - Karpathy suggests this method of autonomous, self-improving research, which he calls 'the new meta,' could trigger an intelligence explosion sooner than anticipated. - The results of the auto-research process are visualized in a graph showing validation loss decreasing across approximately 700 autonomously generated experiments. ![Screenshot at 00:56: The 'Scenario: Intelligence Explosion' graph illustrates the projected growth of 'Effective Compute' normalized to GPT-4, showing where 'Automated AI Research' intersects with human-level and potentially superintelligent capabilities, framing the context for Karpathy's work.](https://ss.rapidrecap.app/screens/tUkD0oj92Qg/00-00-56.jpg) **Context:** The video discusses the release of an open-source AI research project by Andrej Karpathy, formerly of OpenAI. This project, dubbed "autosearch," is presented as a method for achieving recursive self-improvement in Large Language Models (LLMs) without constant human intervention. Karpathy compares this automated approach to his prior manual research methods and highlights significant performance gains achieved rapidly by the AI agent. ## Detailed Analysis The speaker details the release of Andrej Karpathy's "autosearch" project, which aims to automate the research process for improving AI models. Karpathy previously used this system to tune the 'nanochat' model, achieving significant improvements in just two days across about 20 changes that lowered the validation loss. These changes were additive and transferable to larger models (depth=24). The primary result was a drop in the leaderboard's 'Time to GPT-2' metric from 2.02 hours to 1.80 hours, an 11% improvement, which Karpathy considers a real breakthrough. The project's core idea is to give an AI agent a small LLM training setup and let it experiment autonomously overnight, modifying the code in 'train.py' based on results, while humans only interact with the instruction file 'program.md'. The speaker notes that this automated, iterative research process, which generated approximately 700 changes autonomously, is unlike the manual, paper-reading process he is accustomed to, suggesting it could accelerate the path to Artificial Superintelligence (ASI). The speaker plans to further explore this by kicking off "round 2" of auto-research, focusing on how multiple agents can collaborate to unlock parallelism. ### Karpathy's Auto-Research Project - Andrej Karpathy released an open-source "autosearch" project allowing an AI agent to conduct iterative research autonomously - The agent modifies 'train.py' based on experiment results (5-minute training cycles) and updates 'program.md' - The goal is to find the research code that produces improvements on nanochat the fastest. ### Quantifiable Results - Auto-research tuning on nanochat for ~2 days on a depth=12 model yielded ~20 changes that improved validation loss - Measured improvement: 'Time to GPT-2' dropped from 2.02 hours to 1.80 hours (~11% improvement) - This result is considered a new leaderboard entry. ### Project Structure - The repo is deliberately small, containing three main files: 'prepare.py' (constants/data prep), 'train.py' (the file the agent edits), and 'program.md' (baseline instructions edited by the human). ### The New Meta - Karpathy suggests this automated, self-improving research is the 'new meta' and could potentially trigger an intelligence explosion faster than anticipated, contrasting sharply with the slow, human-driven research he has done for two decades. ### Future Work - Karpathy is kicking off 'round 2' of auto-research to investigate how multiple agents can collaborate to unlock parallelism in this research process. ![Screenshot at 00:00: The speaker, Andrej Karpathy, begins the discussion on his latest AI project.](https://ss.rapidrecap.app/screens/tUkD0oj92Qg/00-00-00.jpg) ![Screenshot at 00:56: A graph titled 'Scenario: Intelligence Explosion' projecting compute growth and potential timelines for reaching superintelligence.](https://ss.rapidrecap.app/screens/tUkD0oj92Qg/00-00-56.jpg) ![Screenshot at 02:28: A screenshot of Andrej Karpathy's tweet detailing the 'autosearch' project and its minimal codebase.](https://ss.rapidrecap.app/screens/tUkD0oj92Qg/00-02-28.jpg) ![Screenshot at 08:04: A section of the project's README file detailing the training parameters for the small GPT model, including context size, layers, and training time on an A100 GPU.](https://ss.rapidrecap.app/screens/tUkD0oj92Qg/00-08-04.jpg)