An ex-OpenAI researcher just deleted language from the LLM...
Fireship introduces Jev's fixed answer types, contrasts them with generated text, and questions launch benchmarks and claims about how the model works.
Original by FireshipGetting startedOverview5 min 27 secPublished
Give the model context and a question with a fixed answer shape: a choice, a rating or a yes-or-no probability, rather than a written reply.
A correctly shaped answer can still be wrong. The report explicitly separates the allowed output format from factual correctness.
Treat the reported speed and cost multipliers, architecture claims and open-source comparisons as launch commentary, not independently tested results.
Worth knowing
Launch/news commentary, not an implementation walkthrough or an independent benchmark. The confidence explanation around 3:31 is misleading: Choice and Score confidence describe the returned probability distributions, not percent-correct or measured accuracy; Noul has no separate confidence field (TypeSafe documentation: https://docs.typesafe.ai/confidence). The closing Mux advertisement, from about 4:24, describes Mux features, not Jev capabilities. Notes are based on retrieved captions, not a playback review.