Jevgpt — news.ycombinator.com

No, and the author says so. He describes it as a shower-thought demo he does not expect to improve anything, and points out the mechanism still samples tokens by a shallow heuristic, which is exactly what autoregressive models get for free.

Because a single request caps at 255 decisions. To cover a token vocabulary around a thousand, the author splits the options across several calls and adds a verifier pass that picks the best of the parallel answers.

The author suggests 255 squared in theory, with deeper verifier stacks pushing it further, while flagging that it gets slow. Every additional token is another round trip to the hosted API, so latency rather than correctness is the wall.

None published. A commenter asks how much the verifier pass actually improves output over simply enlarging the vocabulary, and the thread never answers with a measurement. Treat it as a curiosity, not as evidence about the model's capabilities.

A Show HN by kunggaochicken for giga-james/jevgpt, an autoregressive chatbot built by repeatedly asking Jev to pick the next symbol from a tiny vocabulary. It drew 26 points on 23 September 2026.

It samples tokens by a shallow heuristic to build the decision pool, around 1,000 of a 90,000-token tokenizer vocabulary, while standard autoregressive models see the entire vocabulary when predicting the next word.

At github.com/giga-james/jevgpt. It is a demo project rather than a TypeSafe product, and the author says the cost is impractical for real chat.