LLMs aren’t just regurgitating—they’re evolving provably better math proofs and GPU kernels. Auto-discovery just went general-purpose.
What if instead of prompting LLMs for code, they discovered entirely new algorithms better than human SOTA? That’s no longer sci-fi.
Enter knowledge-creating LLMs like AlphaEvolve (Novikov et al., 2025) and TTT-Discover (Yuksekgonul et al., 2026). These are general methods turning LLMs into optimization engines for any problem: they evolve novel, provably correct algorithms in math, CS, GPU kernels (2x faster), biology denoising, and more—surpassing prior specialized AI like AlphaFold.[1]
Devs, this supercharges your stack: auto-optimize models, invent better schedulers/codecs, or craft trading algos. Unlike narrow tools, these adapt quickly to your domain, slashing R&D cycles from months to hours.
AlphaFold was protein-specific; these are universal, generalizing across maths (Erdős problems), AtCoder contests, and single-cell bio. TTT-Discover SOTA’d nearly every benchmark it touched, verified by experts. The shift from ‘AI assists’ to ‘AI invents’ rewires the ecosystem.[1]
Check the papers, replicate on toy optimizations, or build verifiers for your use case. Imagine deploying self-improving agents—what breakthroughs will you unlock first?
Source: tecunningham.github.io