Summary
Matthew Berman breaks down Jev, a new AI model from a co-inventor of ChatGPT that departs entirely from traditional large language model architecture. Rather than a chat model, Jev is described as a generalized “decision model” trained with a new method called RLCD (reinforcement learning for calibrated decisions), positioned as an alternative to the standard RLHF approach. The video walks through the creator’s claims that Jev is up to 200 times faster and 400 times cheaper than comparable models, with output tokens offered free and input tokens costing fractions of a penny.
Berman demonstrates Jev’s speed through several live comparisons, including a real-time Wikipedia “wiki race” against Gemini 3 Terra, Haiku 4.5, and Sonnet 5, where Jev completes multiple hops in a fraction of a second. Other demos include Jev playing Doom in real time by making in-loop decisions, routing customer support tickets, and powering dozens of AI-driven characters in a simulated town who react individually to prompts within seconds.
The video also covers Jev’s claimed benchmark performance against models like Sonnet 5 and Opus 5, and its makers’ assertion of zero hallucinations, positioning it for high-stakes, latency-sensitive use cases like fraud detection, logistics, and automated business decisions. Viewers get a clear picture of what Jev is built for, its tradeoffs compared to conventional LLMs, and how it might fit into existing automation stacks like Zapier.
📺 Source: Matthew Berman · Published September 18, 2026
🏷️ Format: News Analysis
