LLMs are trained to predict next tokens, not make decisions
helloiamleonie · x · 2026-11-06
The naive pipeline: prompt, one forward pass, take last-position logits, keep option-level tokens, softmax into a decision. The root problem: LLMs are trained to predict the next token, not to make decisions, so this usage is inherently unstable.
More from coding & agent
- Meta engineer to speak at DevFest Essex 2026 on AI agents entering Kaggle competitions — AndLukyane · 2026-11-06
- Coding agents install deprecated MCP servers from stale training data, research finds — Arindam_1729 · 2026-11-06
- Ground Truth MCP connector keeps AI agent answers fresh with monitoring and alerts — modelcontextprotocol · 2026-11-06
- Reddit MCP Server lets AI agents read and interact with Reddit via OAuth2 — modelcontextprotocol · 2026-11-06
- playmcp turns any track into a playable music game via AI agents — WhaleSong2077 · 2026-11-06
- Supabase November update: Grok Bot integration lets teammates request changes in chat via Cursor — dshukertjr · 2026-11-06