Google models hold all the world's books but can't read them, dev argues
gnukeith · x · 2026-10-31
Continuing his anti-scaling thread, gnukeith uses Google models as his example: they pack all the world's knowledge inside, yet don't truly understand it — like handing you every book in existence and asking if you know how to read, or even hold, one. What happens when a page is ripped?
He reiterates that the only theoretical benefit of a massive model is concentrated control: only the company with enough compute can host it, which serves no one but itself. Hence the industry should aim for small and mighty models.
Related event: Developer pushes back on scaling hypothesis: constraints drive innovation(2 posts)→
More from AGI Musings
- OpenAI Researcher: Anthropic and OpenAI Could Hit RSI as Early as Next Year — haider1 · 2026-10-31
- Krueger: analyzing petabytes of logs may beat mech interp for superhuman AI — DavidSKrueger · 2026-10-31
- Anthropic researcher ben kuhn: leverage, not raw talent, sets apart the most impactful engineers — benkuhn · 2026-10-31
- Mechanomorphism: when the metaphor flips from machines-like-us to us-like-machines — PierceLilholt · 2026-10-31
- 'Math Talent Edge Erased by US Models': A Debate on US-China AI Competitiveness — teortaxesTex · 2026-10-31
- Andrew Carr: strong priors and RL solve reasoning, math, games, 3D vision—everything but jokes — andrew_n_carr · 2026-10-31