Ex-OpenAI's Keith: a 2B model could handle 75% of your tasks and beat 70B
gnukeith · x · 2026-11-11
Keith (@gnukeith) says much of next year will go into exploring "intelligence per task" rather than general intelligence: a 2B model can likely do 75% of your tasks. He argues the sub-9B market has huge potential with the right tooling—it just needs task-specific data wrapped around it first, a friction users won't do themselves. The model will look very dumb initially, but given enough time it can far exceed 70B models. A reply hopes for more intelligence per parameter, e.g. a new 12-14B model matching an older 27-32B one.
Related event: Ex-OpenAI Member Buys Framework Machines to Train Small Models(2 posts)→
More from Models
- User recreates an Ocarina of Time-style 3D adventure with Fable 5.2 — FinanceYF5 · 2026-11-11
- Fable 5.2 built a browser-playable 1v1 FPS you can share with a link — FinanceYF5 · 2026-11-11
- Fable 5.2 generates a samurai sword-fighting game with a real combat loop — FinanceYF5 · 2026-11-11
- Fable 5.2 lands and users already built 5 playable 3D games with it — FinanceYF5 · 2026-11-11
- GPT-7 could make prompting obsolete: just describe the outcome, says pundit — VraserX · 2026-11-11
- Qwen4 Max Preview quietly appears in Qwen Chat alongside Qwen3.7 and 3.8 models — scaling01 · 2026-11-11