The Jev idea was published a year ago — open, and unnoticed
An independent researcher says he built the thing TypeSafe raised $40 million for, in March 2025, and posted the paper, the weights and the dataset. The receipts are the story.
Nandakishor M, who says he now runs a small AI company called Convai Innovations, published a reinforcement-learning model that skips autoregressive token generation entirely and outputs probability distributions in a single pass — in March 2025. On Wednesday he reposted the arXiv paper, the Hugging Face model and the training dataset, and laid them beside TypeSafe's Jev launch from last week. His pitch is narrow but exact: a policy model trained with PPO over sequence embeddings that emits turn-by-turn conversion probabilities between 0.0 and 1.0, executed without decoding one token at a time. He says it hit 96.7% prediction accuracy against 62% for GPT-4 at the time, at 85 milliseconds per conversation turn versus 3,450 — both his own figures, from his own paper, not independently reproduced.
The architectural overlap with Jev is real but not identical, and he says so. His model is vertical — sales conversations — while Jev generalizes across domains, and it samples in parallel using reinforcement learning from calibrated decisions where his used PPO. A second paper he published in September 2025, on confidence-aware routing to catch hallucinations before generation, is closer to the System One framing TypeSafe is selling: structured outputs, typed decisions, no prose.
The comment thread did not hand him a victory, and that is the most useful part of the post. A LocalLLaMA regular pointed out that non-autoregressive models are not new — encoder architectures like BERT predate GPT, and natural language inference framed as typed decisions is older than both; Jev's makers were vague about what their model actually is, but a matching input/output interface is not theft. The author agreed: "It was never stolen work. Just the architecture similarities." What he is contesting is the word "breakthrough" attached to a design that shipped in public, with weights, fifteen months earlier.
That argument only bites because the artifacts exist. TypeSafe's Jev arrived with a demo video, a pricing page and a claim of 40 to 400 times cheaper and 20 to 200 times faster; there is still no technical paper and no open weights. Nandakishor's are on arXiv and Hugging Face, citable and downloadable, and his complaint is the oldest one in open source — that distribution, not construction, decides who gets credit.
What to watch: whether TypeSafe publishes a technical report that distinguishes its training method from the 2025 work, and whether the open version gets evaluated independently now that it has attention. We covered the TypeSafe launch on September 15 — TypeSafe exits stealth with $40M for models that don't chat — and this is the follow-on question that post raised: what is actually new here.
Would you rather have the frontier lab's closed model or a year-old open one that does most of the job? Tell us in the comments.
