Kristian Fagerlie builds a research loopAn autonomous research loop is a workflow in which an AI system repeatedly proposes, runs, evaluates, and updates research with limited human intervention. inspired by Andrej Karpathy's autoresearch, with a separate evaluator that the coding agent cannot modifyAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository.. Proposed changes must survive multiple seeds, day-level bootstrap testsBootstrap resampling estimates uncertainty by repeatedly drawing replacement samples from observed data and recalculating a statistic. and a held-out confirmation periodHeld-out evaluation checks an AI system or proposed change on examples that were not used to choose or tune that change., and several apparent improvements are rejected.
The Hyperliquid demonstration shows a voice interface placing and closing orders, including missed fills and losing trades. Kristian Fagerlie presents this as a software-control experiment, not evidence of a profitable trading strategy.
A Kalshi example examines contract semantics, stale quotes, settlement rules and actual fills. The creator reports large price changes, but also shows low liquidity and states that profit after fees for the automated research system is not proven. The video's results are attributed experiments, not guaranteed returns or investment advice.
Watch on YouTube
