Researcher: guidance is underrated and works for weak models — in any space
churchkey · x · 2026-11-18
The author argues that people in the language-model world don't understand guidance, yet it's a powerful tool for weaker models — and guidance can be applied in any space, not just text. This echoes recent research directions using generative-model priors for guided search.
More from Models
- Ethan Mollick: GPT-6.2 Is OpenAI's First Frontier Release Since GPT-4.5 That Isn't a Clear Step Forward — emollick · 2026-11-18
- banteg benchmarks GPT 6.2 Sol and Astra debut on his puzzle eval — banteg · 2026-11-18
- Apple Publishes Siri AI Daily Usage Limits for Every iCloud+ Tier on iOS 17.3 Dev Docs — rudrank · 2026-11-18
- DavidKPiano Taunts: Wait Until They Release Something Better Than Opus 5.5 — DavidKPiano · 2026-11-18
- Microsoft ships MAI Code 2 and Code 2 Flash coding models into GitHub Copilot — marlene_zw · 2026-11-18
- Dev altryne: vibes for the new ChatGPT 6.2 models are off the charts — altryne · 2026-11-18