@LysandreJik says decision models landed as a shared primitive in Hugging Face Transformers: DecisionConfig, pipeline("decision") and model.decide(...). He calls @typesafeai's Jev the OG: typed decisions, calibrated probabilities, one forward pass.
- 6.8k views
- 105 likes
- 70 bookmarks
- 20 reposts
About this case
Platform. Lysandre (@LysandreJik) says this could matter more than support for any single model. He contrasts Jev with @AutoTrustAI's JEV (open System One models around a learned decision block, with multimodal), @jaredpalmer's KEV (a pointer head over the options), @Cloudflare's CLEF (a joint schema head), @perplexity_ai's Decider (a separate 255-way readout on Qwen), @liquidai's d1 (small multimodal models down to the edge), and @XunzhuoLiu's Semantic Router team with System One and the D3 family (one typed interface for choice, score and yes/no, from sub-1B to 27B). Transformers adds a shared layer so engines, training tools, the Hub and evaluations can speak one interface without custom or remote code. The idea is not to make every model work the same internally. The post links to https://github.com/huggingface/transformers/pull/50054. Photo attached.
Original posthttps://x.com/LysandreJik/status/2120160642934157609