Inference compute will dwarf training compute — yet we never use it to improve models
LChoshen · x · 2026-11-12
Akshat Gupta points out a long-standing thesis that inference compute will far outweigh training compute globally — yet none of that inference compute is ever used to actually improve the model, which he argues seems like a waste, raising questions about recycling inference into training.
More from AGI Musings
- Can transformers test AI consciousness? Functional equivalence keeps the person silent — tallmetommy · 2026-11-12
- Stanford psychologist Jamil Zaki warns AI sycophancy reinforces biases — CurieuxExplorer · 2026-11-12
- Beyond 'open-source AI': the case for permissionless compute and decentralized training — const_reborn · 2026-11-12
- Reddit user argues AI-generated images, videos and music are art — Sensitive_Judge_5502 · 2026-11-12
- Remote Labor Index:AI 远程工作自动化率一年翻倍至 41.3% — Puzzleheaded-King584 · 2026-11-12
- Gary Marcus: AI coding agents still need human cleanup in ~75% of tasks — GaryMarcus · 2026-11-12