Under 1T params already superhuman — AI models aren't underparametrized, argues researcher
teortaxesTex · x · 2026-11-03
teortaxesTex argues that models under 1T parameters already show human-level or sometimes superhuman performance, suggesting they aren't "underparametrized" in the most interesting ways. He thinks it would suck if human sample-efficiency requires models spanning hundreds of nodes, and doubts humans meaningfully have 100T parameters since synapses are heavily redundant.
More from AGI Musings
- 'Recursive self degeneration': AI that gets more powerful while making worse decisions — cocktailpeanut · 2026-11-03
- Very few people treat the future as construction rather than prediction — signulll · 2026-11-03
- Researcher: alignment will be solved the bitter-lesson way, not via human enhancement — tszzl · 2026-11-03
- Psychics push back as AI-generated readings flood apps and TikTok Live — Chris_Armstrong · 2026-11-03
- Bending Spoons CEO: AI now writes 94% of our code, up from 0% in two years — illscience · 2026-11-03
- Lightcone Commons donor-routing mechanism draws candid caveat from evaluator — NathanpmYoung · 2026-11-03