Rethinking classifiers: conversation+tool templates as special cases of inference-time specs
austinvhuang · x · 2026-09-21
- Austin Huang proposes flipping the framing: rather than treating classification as an optimized special case of LLM usage, the conversation+tool-template paradigm may itself be a special case of general models that accept arbitrary inference-time specs.
- He cites Vik Paruchuri's Lift model direction as convincing prior art: an encoder reads the state (vision in Lift's case), an inference-time JSON spec is passed as input, and the model's output fills the spec's slots.
- The take has direct implications for dspy, XML-structured-output approaches, and whether task-specific classifiers survive at all.
More from coding & agent
- e2b to demo agents forking live VMs to explore solutions in parallel — badphilosopher · 2026-09-21
- User drops Astra after bad week, says Codex needs a real orchestrator — koltregaskes · 2026-09-21
- Agent evals' cold start is smaller than you think: start ugly, iterate — sarahcat21 · 2026-09-21
- Models Don't Have Agency. Systems Do.: Structured Output Is What Makes Agents Work — sethjuarez · 2026-09-21
- LukeW: Using the humble ALT tag to fix irrelevant image retrieval in AI replies — LukeW · 2026-09-21
- Dev shares AGENTS.md rule: skip unit tests after coding, rely on E2E only — alexcovo_eth · 2026-09-21