The Daily Hallucination

SAN FRANCISCO — A well-funded startup founded by a co-inventor of ChatGPT this week launched a new kind of AI model that does not generate text at all but instead reads a situation, considers the options it is given, and returns a probability for each, an approach the company says is "what people wanted from AI all along: a very fast coin that has read the room."

By Monday afternoon the open-source community had produced Kev, Jeff, Nimble and OpenJev, the last of which was renamed SemIf within 48 hours following what its maintainer described as "a conversation about the word Jev."

"It runs in your browser. It takes between half a second and two seconds, which is not fast," said Tomasz Wierzbicki, who built a clone on a 600-million-parameter model over the weekend. "But it is not writing you a poem about it either."

The original model was trained on 100 percent synthetic data and launched without a paper, weights or a dataset, a combination one investor called "clean." At least one researcher has pointed out that he published the same idea in 2025 with weights, a dataset and a package, an approach industry observers describe as "the mistake."

"Classification models have existed for 60 years," said Ingrid Sallinen, a machine-learning engineer. "What is new is that it is 2026 and we are relieved. We have spent three years asking a model to write a paragraph and then parsing the paragraph for a yes. Now it just says 0.83. People wept."

Asked what the model is for, the company said "routing," "triage," and "decisions," and that its enterprise customers had used it to choose between three options 11 million times in the first week, none of which were "none of the above," because that option was not offered.

At press time, asked whether it was a breakthrough or a classifier, the model had returned 0.51 and 0.49.