Dwarkesh Patel: AI's bottleneck isn't compute — it's geometry
Dwarkesh Patel · youtube · 2026-10-23
Dwarkesh Patel released a new video arguing that AI's real bottleneck isn't compute but "geometry" — the physical layout of processors and high-bandwidth memory, and the cost of shuttling data back and forth between chips. The accompanying visual shows processor-HBM stacking, framing data movement rather than raw FLOPs as the core constraint. A contrarian take on the usual compute-scaling narrative.
More from AGI Musings
- Analyst: DeepSeek is more ideological than Anthropic — that's why it will win — teortaxesTex · 2026-10-24
- Rachel Thomas revives classic 'Ironies of Automation': designers leave humans the tasks they can't automate — math_rachel · 2026-10-24
- tszzl predicts stunning biology breakthroughs within years, cancer included — tszzl · 2026-10-24
- Thread untangles Hinton's AI risk claims: real harms yes, conscious doom machine unproven — AlexTensor · 2026-10-24
- Berkeley-led report argues a hardwired pause of frontier AI training is feasible, challenging Hinton — wfithian · 2026-10-24
- Investor tszzl: we're in an extraordinary renaissance, epicentered in San Francisco — tszzl · 2026-10-24