A 'new type of mind' trained without natural language writes kernels at up to 66x Triton speed
repligate · x · 2026-09-27
A demo of a system described as a 'new type of mind' — trained on neither natural language nor natural-language reasoning — went viral: it produced kernels that passed all tests 28% of the time and achieved speedups of up to 66x compared to Triton kernels. The result sparked discussion about whether non-language-trained systems (e.g. search/evolutionary approaches) hold unique advantages for kernel and code generation.
Related event: Language-Free 'New Mind' Writes Working GPU Kernels(2 posts)→
More from coding & agent
- Dev caches hide 264 CVE-laden packages and 71.7 GiB no one audits, warns Cache Commander dev — julsimon · 2026-09-27
- WebMCP could become the HTML/API layer of the agentic web — Thionne_WTZ · 2026-09-27
- Stanford/Together AI paper: agent teams hit 66.7% vs 48.8% for single agents — mark_k · 2026-09-27
- antirez regrets not finding time to hand-write small poetry-style programs — antirez · 2026-09-27
- mitsuhiko on Astra: expensive, narrow edits, and a frustrating experience to review — mitsuhiko · 2026-09-27
- François Fleuret asks: are there programming languages tailored for LLMs? — francoisfleuret · 2026-09-27