Claude Code Finds Efficient AI Scaling Algorithm #Shorts

Añadido: 2026-05-26

[00:00:03]AI Watch Claude code was used to search for better test time scaling rules, and the result reportedly used far less compute.

[00:00:12]A research team from UMD, UVA, Washington University in St. Louis, UNC, Google, and Meta used a coding agent to search for better test time scaling algorithms.

[00:00:23]Test time scaling is the idea that a model can perform better when it spends more compute on a problem, for example, by running several solution paths or extending reasoning.

[00:00:33]Traditionally, humans write the rules for when to open a new path, continue one, or stop it.

[00:00:40]In the Auto TTS work described by the decoder, humans instead built an offline environment where an agent could test control algorithms against pre-generated model outputs.

[00:00:50]Claude code then wrote and refined candidate algorithms over multiple rounds.

[00:00:55]The reported result was better accuracy per unit of compute on math benchmarks, with a lean setting cutting token use by about 70% compared with standard self-consistency while maintaining accuracy.

[00:01:07]The full discovery run reportedly cost about $40 and took 160 minutes.

[00:01:14]The interesting part is the role shift.

[00:01:16]Humans define the search space and evaluation setup, while the agent proposes the actual control logic.

[00:01:23]The caveat is that the current version focuses on width and depth tradeoffs, not more complex tree search structures.

[00:01:29]According to the decoder, the key question is shifting from whether humans can hand-design every reasoning strategy to whether they can build reliable search spaces for agents to improve them.

[00:01:41]Sources linked below.

[00:01:43]Follow for AI, automation, and security briefs.