Nandakishor Mukkunnoth published a reinforcement-learning model that predicted sales-conversion trajectories from conversations in March 2025 — arXiv paper, weights on Hugging Face, PyPI package. A year later a funded lab, TypeSafe AI, launched Jev, a non-autoregressive “System 1” decision model its author called a breakthrough, with no papers, no weights and no datasets. Mukkunnoth’s post is part grievance about that, and part argument about when a generative model is the wrong tool.

The argument first. Most decisions inside an AI pipeline are reflexes, not reasoning:

  • Which queue owns this ticket? Is this email phishing? Is this prompt a jailbreak attempt? How urgent is this on a 0-to-3 scale? Is this retrieved passage relevant?
  • Calling an 8B-plus generative model for those means waiting 500–2,000 ms, paying per call, and then parsing free text back into a clean label.
  • And the confidence an LLM reports is not a probability: as he puts it, when an LLM outputs confidence: 0.95, “it is just predicting tokens that sound confident.”

His answer, Laya, answers three typed questions in one forward pass — pick one from a list, rate on a scale, or return a yes/no probability. The output space is only numbers, so it cannot generate text, cannot hallucinate, and cannot return malformed structure.

The published numbers: 32.8 ms per question on one GPU (7.2 ms each when batched), 7.8x faster than Jev’s published latency, and a calibration error of 0.081 against Jev’s 0.246 — meaning its stated probabilities are closer to the frequencies they actually describe.

The most useful finding is not on the benchmark board. Running an English-only checkpoint across 51 languages, accuracy collapsed on non-Latin scripts while confidence stayed high:

  • Khmer: 0.000 accuracy at 0.952 mean confidence
  • Armenian: 0.050 accuracy at 0.885 confidence
  • Hebrew: 0.060 accuracy at 0.964 confidence

Across the whole sweep the English checkpoint’s confidence never dropped below 0.885, whether its accuracy was 82% or zero. That removes the usual safety net — you cannot rely on the model to tell you when to route elsewhere, because the routing has to happen before the forward pass.

He also publishes the limits, which is rarer than it should be: pick-one questions degrade past roughly 20 options (0.425 accuracy at 77 labels), a scalar temperature has to be fitted per question type on your own data, and the headline score of 0.766 comes from fine-tuning on the benchmark’s train split — out of the box the base model sits near random at about 0.35.

That last caveat is what decides whether Laya is a replacement for Jev, and it is where the discussion went. The 268-comment thread on Hacker News.

What the thread adds

  • johnfn — the diagnosis that reframes the post: “the full definition of ‘product’ INCLUDES being able to coherently communicate it. In some sense the branding is just as much the ‘breakthrough’ as the model.” He contrasts Jev’s page, legible in ten seconds, with Mukkunnoth’s single Reddit announcement, titled “Predicting sales conversion probability from conversations using pure Reinforcement Learning.” calebkaiser adds the ML-specific version of that dynamic, where near-identical work keeps getting reinvented in parallel.
  • Oras — tried both tools and gives the reductionist verdict from someone who trained NLP models before LLMs: “it’s just BERT with more data.” Their softer conclusion is the practical one: “a wake up call for the tech community to go back to basics for most tasks instead of relying solely on generic LLMs.” prometheus1992, dwa3592, pdp and kamranjon (who points to the earlier GLiNER papers) arrive at the same place from different directions.
  • soerxpso, yojo, ianbutler, wild_egg and cjalmeida — five commenters making the same counter-argument in five different wordings: the zero-shot property is the product. soerxpso: “I don’t want to spend a week to create a dataset (for a problem I might not already have data for), finetune a model, and set up infrastructure to run the model, every time I have a small routing or classification problem.” ianbutler: “I do just want a reasonably good general classifier served to me with a great devex and calibrated confidence scores to help me figure out when to fallback to another model.”
  • cjalmeida — a correction aimed at the article’s own comparisons: “This should be way up in the article. Fine tuning is a pain, requiring it for good results put Laya in a whole different category vs Jev.”
  • fernandezpablo — a one-line independent test on their own labelled support-question set: “jev gets 95% answers correct. Layla 48%.” No method or dataset is posted, so read it as a claim, not a result.
  • baobabKoodaa — disputes the “open source version of Jev” framing, citing the small context window and the fact that the 2025 paper solved one narrow business problem rather than building a generalist classifier. hbrn answers by disputing Jev’s own claimed capability instead, listing failures they say they observed. Both sides are unverified claims from opposite directions.
  • dcow — on the grievance itself: “I can understand why the author feels bitter but it still feels juvenile to me … Diogo decided to build a product out of the concept. The author didn’t.” jwpapi pushes back on the pile-on: “With everybody bashing OP here, how is he supposed to even make money?”

The question the thread leaves open

jerpint asks the thing this run of the argument never resolves: “Am I understanding correctly that the field has gone full circle and we are back to specialized classification models for domain specific tasks?” pdp makes the same observation without the question mark — Jev went viral because it reminded people that non-LLM forms of AI still work, and “we tend to forget that.” Whether that is a correction to the LLM era or just a niche that was always there is exactly what the comments disagree about.

Hacker News handles are pseudonymous and the site publishes no per-comment scores, so the ordering above is HN’s own ranking, not a vote. This is a slice of a 268-comment thread, not a consensus.