Models will stall near human-level since pretraining data holds nothing smarter, researcher argues
burny_tech · x · 2026-09-21
AI commentator Drachs1978 offers a blunt take: models will converge on human intelligence because there is nothing in the pretraining data smarter than humans — and once we get near human-level, progress will become hard.
The claim cuts at a core controversy of the current scaling paradigm: whether the ceiling of pretraining data caps model capability. It stands in contrast to the mainstream optimism around synthetic data and RL breaking through data bottlenecks, making it a useful contrarian datapoint.
More from AGI Musings
- AI debate: commenter sides with Terence Tao on the possibility of AI getting banned — teortaxesTex · 2026-09-21
- AI Is Compressing the Whole Company, Not Just Cutting Startup Costs — alexmacgregor__ · 2026-09-21
- Weighting others by relatedness isn't new: paper models agents learning relational value — xuanalogue · 2026-09-21
- Four motivations of mathematicians explain their split reactions to AI — littmath · 2026-09-21
- Binary Bits: post-AGI per-capita incomes likely cap below 10x US median if humans stay in charge — binarybits · 2026-09-21
- Debate: could post-human AI economies push GDP per capita far past 10x? — binarybits · 2026-09-21