Plinz: claims of hard limits on LLMs are nearly impossible to prove
burny_tech · x · 2026-09-28
Plinz argues that LLMs are not limited by training data or any given training approach, and that conjectures about hard capability limits are very risky and nearly impossible to prove.
His take: better approaches must exist, but AI labs' bets on getting LLMs beyond human AI research ability are not unreasonable.
Related event: Researcher Plinz: LLMs Have No Hard Capability Ceiling(2 posts)→
More from AGI Musings
- Classically trained artist with 3D roots since the 90s pushes back on the 'pick up a pencil' AI art critics — dreamwieber · 2026-09-28
- Beff Jezos Launches Kardashev Research, an e/acc Non-Profit, With @mjdramstead — beffjezos · 2026-09-28
- Critic says AI doom rhetoric is an aesthetic choice, not persuasion — trevposts · 2026-09-28
- Inner and outer alignment have shifted from theory to live empirical science — jam3scampbell · 2026-09-28
- François Chollet: AI industry should build tools for humans, not a 'successor species' — fchollet · 2026-09-28
- Next Token Podcast #004: Opus 5.5 Enters Video, Muse Goes Viral, and China's AI PM Debate — oran_ge · 2026-09-28