Turn LLMs into Tunable Soft Classifiers with Logprobs
JnBrymn · x · 2026-09-19
Building on the classification chapter of Hands-On Large Language Models, the author extends the standard "hard classifier" (LLM outputs only positive/negative) into a "soft classifier": by reading logprobs, you get per-class probabilities instead of a single answer, letting you set thresholds tuned against a training set.
He notes he had a similar idea two years ago, and predicts OpenAI and Anthropic will bake this directly into models, obviating many external use cases. A 10-minute video summary is included.
Related event: Turning LLMs into Tunable Soft Classifiers via Logprobs(2 posts)→
More from coding & agent
- Fully Local, Private Computer-Use AI Arrives — Runs on a 3060 Ti 12GB — WolframRvnwlf · 2026-09-19
- ActObs: Supervising Observations During SFT Boosts Agent RL by up to 4.2pp — kastnerkyle · 2026-09-19
- Cloudflare open-sources agentic-inbox, a self-hosted AI email client running entirely on Workers — dinasaur_404 · 2026-09-19
- Pi hits Pareto frontier on both benchmarks with just four tools: read, write, edit, bash — arena · 2026-09-19
- "Let me check in with my agents" is now a daily phrase in this developer's marriage — josh_bickett · 2026-09-19
- "I gave jev eyes": devs celebrate gloriously pointless AI experiments — cocktailpeanut · 2026-09-19