The video showcases the creator’s adaptation of Andrej Karpathy’s “autoresearch” framework to automate and improve an AI-powered Polymarket arbitrage trading bot, using a research loop that autonomously experiments, evaluates, and updates trading strategies. After successful dry runs, the bot is deployed with real funds, demonstrating profitable live trades and validating the effectiveness of the autoresearch approach.
The video documents the creator’s attempt to adapt Andrej Karpathy’s “autoresearch” project to their own AI-powered Polymarket trading bot. Inspired by Karpathy’s approach of using a small GPT model to automate research and code evolution, the creator sets up a similar loop for their arbitrage bot, which trades on the five-minute up/down Bitcoin markets on Polymarket. The goal is to see if this autoresearch framework can improve the bot’s trading strategy by autonomously running experiments, evaluating results, and updating the codebase.
The system is structured around a research loop: the bot operates within a GitHub repository, following instructions from a markdown file called “training program.” This file acts as a research playbook, outlining how experiments are chosen, run, and evaluated. The agent updates the strategy code, proposes new experiments, and tests them in a live (but initially dry-run) environment. If an experiment improves performance according to predefined metrics, it is kept; otherwise, it is discarded. The results are logged, providing context for future experiments.
The creator demonstrates the dashboard and experiment history, showing how the bot’s performance is tracked over time. Each experiment runs for about an hour, and the system automatically evaluates and commits changes based on the results. The dashboard displays stats like uptime, number of trades, fill rate, and win rate. The creator notes that, since the strategy is arbitrage-based, the win rate should be close to 100%, though early mistakes affected this initially.
After several dry runs, the creator switches to live trading with real funds. The bot executes a series of arbitrage trades, buying both sides of the market to lock in small, risk-free profits. The video shows the bot successfully completing five trades in a row, with the balance increasing from $150 to $152, demonstrating a $2 profit in about 20 minutes. The creator highlights the process of resolving trades, updating balances, and ensuring the system remains synchronized.
In conclusion, the experiment is deemed a success: the autoresearch loop effectively improves and tests the trading strategy, and the bot performs well in live conditions. The creator expresses enthusiasm for continuing to run experiments and potentially applying the autoresearch framework to other domains. Viewers are encouraged to check out Karpathy’s original project on GitHub and to subscribe for future updates and experiments. The video serves as both a technical walkthrough and an inspiration for others interested in automated research and trading systems.