Compute is the lagging factor: 1000x inference budget makes AI smarter without new models
_AustinCalvert_ · x · 2026-11-19
Boosting @DGartifact's rebuttal to compute pessimism: compute is the lagging factor — it just takes time to come online and can in theory be added almost endlessly. When you use an LLM or agent you're rationed a small slice of compute; give a task 10x/100x/1000x and the AI gets smarter without the model changing. Models solve new math problems precisely because massive compute is allocated there, which isn't cost-feasible for consumers. Stack on models improving themselves and designing better chips, and the intelligence pool keeps expanding — it's just diluted by skyrocketing demand.
More from AGI Musings
- Philosopher Thomas Nagel singles out Anthropic's approach to AI as 'abominable' — GaryMarcus · 2026-11-19
- Zvi's AI #195: Anthropic IPO day, Claude Fable 5.2 vs GPT-6.2, Google's new models — TheZvi · 2026-11-19
- Zvi's AI #195: Anthropic IPO day, Fable 5.2 vs GPT-6.2, and rising acceleration worries — Don't Worry About the Vase (Zvi) · 2026-11-19
- Bob McGrew on No Priors: superintelligence needs the physical world, hard takeoff is wrong — No Priors · 2026-11-19
- 68% of French see AI as job threat, only 14% make it a vote priority — poll — emmanuelvivier · 2026-11-19
- Gary Marcus: OpenAI's research-acceleration data beats Anthropic's and shows no runaway superintelligence — GaryMarcus · 2026-11-19