Could a processor run Karpathy's 'algorithms of thinking' directly, no giant LLM needed?
McDonaghMatthew · x · 2026-09-04
Reposting @bpcostello's riff on @karpathy's idea that today's models contain an extractable "cognitive machinery": if that machinery could be embedded directly into a processor, AI could run the algorithms of thinking natively in hardware without a giant LLM at the center — a change the author argues would reshape the entire field. The post remains speculative, with no concrete implementation details.
More from AGI Musings
- Blogger: only ~0.1% of people are still worth hiring in the AI agent era — scaling01 · 2026-09-04
- After Harvard shut its writing center, dean Deming now urges AI use in writing courses — soumitrashukla9 · 2026-09-04
- Six months of 'fully capable' models, yet software feels no better, critic says — MoonL88537 · 2026-09-04
- NYC K-8 AI moratorium is misguided: students need guided AI use, critics argue — castrotech · 2026-09-04
- David Lynch explains why AI is just like a pencil — UrAWeaselBro · 2026-09-04
- Investor cites OpenAI execs: multiple reserve models, scaling laws still holding, long compute — thedealdirector · 2026-09-04