A reflex layer, in the creator’s words
ConvAI’s founder, Nandakishor Mukkunnoth, sat down on Front Page this week. The host asked how a decision model shipped from Kerala could trend on Hugging Face within days. This note is a map of what he said, pointed at the gallery.
He described Laya as a System 1 layer: a state goes in — an email, a ticket, a JSON blob — and typed questions come back as choice, score, and noul. There is no paragraph to parse. The model card is the spec. The interview is him explaining why that shape exists.
Why not generate the answer
The example he used was a hospital door. Two patients arrive. A doctor has to pick who goes to the ER first. An 80% from a model is not useful if 80% does not mean 80% in the real world. He walked through calibration as rewards when a probability matches the ground truth and penalties when it does not. That is why the card tells you to fit a temperature on your own data before you trust a score in production.
Speed is the other half. He kept coming back to on-device, offline, “zero egress”: healthcare, finance, anything that should not leave the machine. The card’s published figure for one multilingual question on a T4 is 32.8 ms. He also pointed at the ports people have already posted — MLX on Apple Silicon, games in the tens of milliseconds — as the reason a decision model is worth putting next to a larger chat model, not in place of it.
His routing picture is simple. Use Laya for the cheap, structured gate. Escalate when confidence is low, the request is unsafe, or the job actually needs a paragraph. That is the reflex-layer story. The catalog is already full of those experiments: MLX, Core ML, ONNX, browser Snake, and the typed-decisions fine-tune.
What he was careful about
When the host cited an independent test where a base Laya checkpoint lost on accuracy to larger classifiers while winning on speed, he did not argue the timing. He said a base checkpoint is not the production tool. Llama has instruction-tuned variants for the same reason. The published 0.766 figure on the authors’ 2,000-decision set belongs to laya-typed-decisions, not the zero-shot English or multilingual weights. There is a Kaggle notebook in the upstream repo for that job.
He also said the first public Laya drop was about 12 to 15 hours of work on a train, after years of papers and smaller on-device models.
Watch it, then add yours
The conversation is on YouTube. His longer write-up is at laya.convaiinnovations.com, and the weights are Apache 2.0 on the hub. If you have a public port, bench, or typed-decision loop, put it in the gallery. The question he left on the table is the one this catalog is for: which decisions should happen locally, quickly, and already typed.