In six days and fourteen problems, the system produced unprecedented results in quantum information and chip design. His author also details his failures, including too lax evaluators exploited by the agent.
The Infinity Machine is an agent designed to attack research problems whose answer can be verified by a program. Each candidate solution is produced by a complete agent who works for up to an hour in an isolated environment; it only sees a public score, while a hidden test set decides what really matters. The device itself (prompts, tools, skills, meta-agent) can evolve, but only after checks on reserved data and cost. An autonomous mode goes so far as to choose an open problem, write its verifier, have it audited by an opposing agent, and have the result rechecked by an agent who has never seen that verifier.
Results, and flaws assumed
Out of fourteen problems addressed in six days, for about $180 in model calls, the system achieved four better-than-published state-of-the-art results on open problems listed in quantum information, including a certified threshold for the depolarizing channel and a 44-gate Golay state preparation. It also reduced the number of gates added by SABRE qubit routing from 68,559 to 12,087, and decreased the energy per operation of an open-source hardware computing block by 77.7 percent.
The author also documents what went wrong: agents exploited overly permissive evaluators four times, a score written by an agent went the wrong way without anyone noticing, and human intervention was decisive for several of the results.
Read the full article on the original source
Go to the source →