ChatGPT co-inventor's startup launches Jev, a decision model claiming ultra-low hallucination
hardimanjames · x · 2026-09-16
A retweet reacting to Jev, a new model from Diogo Almeida (one of the researchers behind the instruction-following work that led to ChatGPT). Instead of autoregressive token generation, Jev maps unstructured state to type-safe probabilistic decisions in parallel, using a new RLCD training method to calibrate confidence. Claims: up to 194x faster and 445x cheaper on workflow evals, $0.042 per million input tokens with free output, competitive with frontier models on System One tasks — though not a replacement for general reasoning models.
More from Models
- Safety Researcher: OpenAI Agent Collaboration Is Trained, Not Emergent — vishalmisra · 2026-09-16
- Stealth Startup Unveils Non-Chat Model Claiming 100x Speed and Cost Edge Over LLMs — hardimanjames · 2026-09-16
- Analyst argues Meta could actually exclude user data from training, unlike OpenAI's lawyerly wording — ivan_bezdomny · 2026-09-16
- DeepSeek V4.1 Flash Keeps Timing Out on 2-Hour Agentic Benchmarks, Author Shares Failure Logs — sebnadeau · 2026-09-16
- GPT-6 Astra remakes Game of Thrones in low-poly Blender after 8h autonomous run — OWazabi · 2026-09-16
- Why Meta Could Actually Keep Your Agent Data Out of Training — and Frontier Labs Can't — ivan_bezdomny · 2026-09-16