Diogo Almeida’s new transformer, Jev, produces calibrated probabilities instead of text, promising cheaper, faster, and hallucination‑free automation.
Diogo Almeida, one of the co‑founders behind OpenAI’s ChatGPT, has unveiled a novel transformer model called Jev that outputs calibrated probability distributions instead of raw text, a shift that could make AI‑driven automation cheaper, faster, and far less prone to hallucinations.
Why calibrated probabilities matter
Traditional language models generate token sequences and then rely on post‑processing tricks to estimate confidence, a process that often yields over‑confident or nonsensical outputs. Jev’s architecture is built to predict a probability vector for every possible token, giving developers a direct measure of uncertainty that can be used to filter or rank results before they reach end users.
Performance gains and cost savings
Because Jev does not need to produce full text strings during inference, it can skip the expensive decoding step that dominates compute budgets for large models. Early benchmarks shared by Almeida’s team show up to a 30% reduction in GPU hours for comparable tasks, translating into lower cloud bills for enterprises.
The model also scales more gracefully: when the probability output is sufficient—for example, in classification or ranking scenarios—developers can avoid a second pass of a full‑text generator, cutting latency by half in many real‑time applications.
Implications for hallucination‑free AI
Hallucinations often stem from a model’s confidence in a token that looks plausible but is factually incorrect. By exposing the raw probability distribution, Jev lets downstream systems detect low‑confidence regions and either request clarification or fall back to rule‑based checks, dramatically reducing the risk of misinformation.
- Improved uncertainty quantification for safer decision‑making
- Lower inference costs for large‑scale deployments
- Faster response times in classification‑heavy pipelines
- Reduced need for extensive prompt engineering
Developers who have experimented with Jev report that the model’s output is immediately useful for tasks like intent detection, sentiment analysis, and recommendation ranking, where the exact wording of a response is less important than a reliable confidence score.
Jev feels like a missing piece in the AI toolbox – it gives us the statistical rigor we’ve been craving while keeping the flexibility of modern transformers.
Almeida plans to open‑source the core training code later this year, inviting the community to build specialized heads on top of Jev for domains ranging from medical diagnostics to financial forecasting.
For a deeper look at Jev’s architecture and early performance numbers, see the TechCrunch coverage of Jev’s debut.