Applied Compute's Ari Autoresearch lifts Qwen3.5-9B chess reward 5.3x in a 16-hour autonomous run

ypatil125 · x · 2026-11-05

Applied Compute launched Ari Autoresearch, letting its Ari agent run post-training experiments autonomously on the AC2 platform without step-by-step human guidance.

In a single 16-hour episode, Ari raised Qwen3.5-9B's held-out reward on Lichess chess puzzles from 0.118 to 0.629, a 5.3x improvement.
Building on Karpathy's autoresearch line of work, it targets enterprise-scale post-training: the agent can tune hyperparameters, improve task harnesses, build curricula, combine learning algorithms, annotate traces with judges, and diagnose failure modes.
The company also announced a $100M run rate and a $350M raise.
Quoted context shows a companion practice: the author turned 3,500 of his review comments across 2,000 PRs into a code-review bot, then had Ari autoresearch and hill-climb its eval overnight to ease their biggest bottleneck.

Original post →

Applied Compute's Ari Autoresearch lifts Qwen3.5-9B chess reward 5.3x in a 16-hour autonomous run

More from coding & agent