I turned Jev into a (lousy) chatbot

Mechanically yes, and the thread's verdict is that it is expensive and bad. One comment notes the loop is trivially available by running an autoregressive pass over next-character predictions, which removes the novelty without needing the model at all.

Nobody measured it, but the most concrete claim is calibration: commenters suggest what separates the hosted model from cheap logit-trick projects is that it claims well-calibrated probabilities, a property general LLMs do not give you without special training.

The author concedes his own survey shipped while he was still writing it and that two better benchmark sets appeared in the meantime. He also notes several projects predated his chatbot gag, which is useful context on how crowded this space already is.

Open, per the thread. One comment grants that the hosted model has some secret sauce over the dumbest possible version of an open model, but says it is unclear how durable that is once a frontier lab does minimal tuning itself.

Asked to 'write me a short story', it answers simply 'a story'. A commenter called it the first time they knowingly laughed at a model's joke, and others compared it to early text-davinci and Llama completions.

petesergeant says Jev clearly has some secret sauce compared to the dumbest single-token approach with Qwen, but is not sure how durable that advantage is if OpenAI wires up Luna-5.6 with minimal tweaking.

It went full circle to 'BERT has a mouth and it must speak', a paper reference, alongside the author's own write-up at sgnt.ai/p/jev, which cataloged earlier single-token-prediction projects and their weak benchmarks.