Autoresearch claims 2718 Elo after 70 experiments on a Rust chess engine

• Developer Deedy Das says Karpathy-style Autoresearch pushed a "vibecoded Rust chess engine" from "expert" strength to a reported 2718 Elo after running "over 70 experiments" and hill-climbing for score results thread.
• The underlying engine is conventional search, not a newly trained model: Deedy's technical breakdown says it uses negamax alpha-beta search with pruning, iterative deepening, opening books, and a transposition table, all tested at a 500 ms per-move limit.
• The practical engineering story is the loop, not chess. The community reaction around Autoresearch frames it as a reusable pattern for "everything with a measurable metric," where the agent proposes code changes and keeps what improves the target.
• That pattern may extend beyond fully verifiable tasks. In a practitioner take, Shreya Shankar says she is "very optimistic" about combining these search loops with qualitative evaluators for more subjective coding-style work.

The claim is specific: the agent ran more than 70 experiments on its own, searching for code changes that improved Elo, and the best run landed at 2718 according to Deedy's results thread. He also describes the end result as "a top 50 grandmaster" and a "#311 chess engine," which gives engineers a concrete target variable rather than a vague quality claim.

The engine itself is not exotic. Deedy's technical breakdown says it uses "negamax alpha-beta tree search with pruning and iterative deepening," plus standard opening books and a transposition table to cache moves. He adds that every test used a 500 ms per-move budget and that there is "no offline computation or training element," so this was an automated experiment loop over ordinary engine code, not a training pipeline or a one-off tuned checkpoint. The most obvious next lever, by his account, would be replacing the static evaluation function with efficiently updatable NNUEs technical breakdown.

The broader takeaway is that Autoresearch looks most convincing on problems with a hard metric and a fast evaluation harness. The community reaction summarizes the pattern as the community applying it to "everything with a measurable metric," which is exactly why chess is a useful demo: Elo is legible, regression is cheap to detect, and the search space is mostly code and parameter changes.

Engineers are already testing how far that loop generalizes. Shreya Shankar's subjective tasks argues that "auto research-style search loops" can be paired with qualitative evaluators for "non-verifiable" tasks, suggesting a bridge from benchmark optimization to coding workflows where the score is fuzzier but still rankable. Another practitioner is already trying the approach with RL infrastructure, as an RL experiment puts it: "Maybe it works, maybe it doesn't." That is where this story lands for engineering teams: an agentic optimization loop appears able to improve real code when the objective is measurable and the eval cycle is short enough to run dozens of times.