rsi.watch

Harish Santhanalakshmi Ganesan · 2026-10-26

Infinity Machine: a self-improving AI agent for research

“Across fourteen problems in six days, for about $180 of model usage, the system produced four results beyond the published state of the art on catalogued open problems in quantum information”
Who
Harish Santhanalakshmi Ganesan
Date
2026-10-26
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ indirect measure. The results are credited to the system, but for the fourteen problems the operator chose and vetted each problem, wrote the verifier, seed and notes, and rewrote the strategy arms for math and ML. The author adds that several problems were chosen because they looked search-shaped, "so the success rate is not representative." “And the operator’s hand-made changes to the harness were decisive for several results.”
  • ↓ imprecise. One to three sessions per problem with a single proposer model, and one session sequence reported per problem. “There are no repeated seeds, ablations or matched-compute comparisons.”
how we rate evidence
Replication
No rerun found. The paper says the Zenodo deposits let anyone re-check the published results without the harness. No code link for the harness is given. We found no re-check or rerun by another group. Searched up to 2026-10-28. all claims
Checked
2026-10-28, against the source page
Links
Source Paper

Also on the page

  • The harness (prompts, tools, skills, playbook and the meta agent's procedure) is a genome the system may edit. A kernel the agent cannot write to accepts an edit only if neither held-in nor held-out tasks drop by more than 0.03 and the cost is at most 1.5 times the incumbent's.
  • The one self-derived generation, four playbook bullets, scored 0.944 against 0.927 on held-in tasks and 1.000 against 1.000 on held-out tasks, at 2.02 times the cost, and was rejected. The author: "Every discovery result in Section 3 was therefore produced by the initial genome."
  • For the fourteen problems, the operator chose the problems, wrote the verifiers and replaced the strategy arms by hand for math and ML. Besides the quantum results, the system cut SABRE's 68,559 added gates to 12,087 (9,312 against the best published 13,044 on the 15 circuits shared with later work); the author notes "The SABRE comparison is not compute-matched." Three fully autonomous runs all ended in ties with the public record, and "in runs 2 and 3 the operator named the table to search."
  • The author re-verified every result on the host, and says "The quantum results are lower bounds or explicit constructions with stated claim boundaries, not optimality results." the qubit threshold point was certified with the record authors' own certifier. The author: "We do not yet have evidence that the Infinity Machine improves itself on these problems."

In the self-improving AI catalog

  • Infinity Machine: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that Infinity Machine cites, oldest first. Under each: where that work sits on the map, then where Infinity Machine sits (changes in bold), and what Infinity Machine's own paper says about it. Its whole family on the citation map

  1. 2023 STOP sets itself apart

    Tools, own code · becomes the next version · improves the improver → does not improve the improver

    “However, unlike STOP [3], we did not audit traces systematically for such attempts, so this is an absence of observed incidents, not a measured rate.”
    “STOP [3] improved its own scaffolding program.”

  2. 2026 Hyperagents uses it

    Tools, own code · becomes the next version · improves the improver → does not improve the improver

    “The meta agent (Claude Opus 5.5), whose prompt and procedure are themselves genome files [5], reads an evidence bundle of recent runs and mines weaknesses.”
    “Hyperagents [5] made the meta agent editable.”

  3. 2026 Infinity Machine

    Tools, own code · becomes the next version · does not improve the improver

Also cites Good, the intelligence explosion 1965, Reflexion 2023, The AI Scientist 2024, AlphaEvolve 2025, Darwin Gödel Machine 2025, Weco AIDE² 2026

← Ledger Report an error