Build your own decision model
Points and comments are a snapshot, not live.
Build a zero-shot decision model by constraining LLM outputs to fixed options.
The article shows how to build a "System 1" decision model that selects from fixed options in a single forward pass, unlike standard LLMs that generate tokens step-by-step. Using Qwen3-1.7B and constrained decoding, it outputs a probability per option. Testing on CommonsenseQA (1221 samples) yields 59.4% accuracy. The model is overconfident; temperature scaling (T≈3.8) improves calibration. A GitHub repo provides scripts for building, evaluating, fine-tuning, and calibrating such models.
What commenters are saying
Commenters split between excitement for the convenience and frustration that this is just old-school classification rebranded. Some point out that zero-shot classifiers have been viable with plain LLMs for years, and Jev's hype feels like Zoom displacing Skype. Defenders argue that this is dramatically cheaper and faster, opening use cases previously not justifiable.
A specific critique: calibration via temperature scaling on the test set is essentially p-hacking, because adjusting any part of the system against the test set overfits to that benchmark. Others note that the appeal is the switch-statement ergonomics and speed (~200-300ms per input) that make it embeddable in games, code hooks, and agent workflows.