November 6, 2025
Sofa, So Good? Math meets meltdown
Mathematical Exploration and Discovery at Scale
AI hunts math breakthroughs — hype, confusion, and 'RIP mathematicians'
TLDR: DeepMind’s AlphaEvolve has an AI that writes and evolves code to search for math results across 67 problems, with data and prompts shared. Commenters swung from doom ("AI will replace mathematicians") to confusion (is the sofa solved?) to curiosity ("pls explain"), making this a milestone and a mood check.
Four mathematicians just dropped a mega-report on “AlphaEvolve,” a Google DeepMind tool that uses an LLM (a large language model) to write and remix code that searches for better solutions to math problems. Instead of tweaking numbers by hand, it evolves tiny programs, tests them, keeps the winners, and tosses the duds. They ran it on 67 problems and say it can match expert-crafted search tricks and scale fast across related puzzles, with shared prompts and scripts. There’s even a public trove of tasks and data; one reader helpfully dropped the link to the problem repository.
Reactions were internet energy. piker begged for a plain-English recap, calling the post “dense but nuanced,” while iNic casually claimed the legendary “sofa problem” was solved and linked an arXiv preprint — cue raised eyebrows and frantic fact-checking. The biggest spark: muldvarp’s doom-flavored mic drop that there’s “zero reason” to study anything but trades because AI will beat “almost all mathematicians” soon. Some cheered, others rolled eyes, and a few just said “very nice～” and moved on. The vibe: part awe, part panic, part meme. AlphaEvolve might scale discovery, but the comments are already stress-testing human morale as hard as the tool stress-tests math.
Key Points
- •Authors uploaded an expanded arXiv paper detailing experiments with Google DeepMind’s AlphaEvolve.
- •AlphaEvolve uses an LLM to evolve code that generates inputs, rather than directly optimizing inputs.
- •The tool was tested on 67 mathematical problems in analysis, combinatorics, and geometry, with mixed outcomes reported.
- •AlphaEvolve achieved results comparable to expert traditional tools and showed advantages in scalability and robustness.
- •Similar tools (OpenEvolve, ShinkaEvolve, DeepEvolve) have emerged since AlphaEvolve’s initial release.