Summary
Fireship’s Code Report dives into “Jev,” a new AI model from TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida after the company raised $40 million. Unlike conversational large language models such as GPT-6 or Claude, Jev strips out free-form language generation entirely, returning only structured “system 1” outputs — a choice, a score, or a null — inspired by Daniel Kahneman’s fast/slow thinking framework.
The video breaks down how Jev claims to be roughly 200 times faster and 400 times cheaper than mainstream chatbots for simple classification tasks, using a technique called RLCD (reinforcement learning for calibrated decisions) to produce confidence scores alongside each answer. Real-world use cases discussed include content moderation, NPC behavior in video games, and real-time AI calculators.
It also covers skepticism around Jev’s novelty, noting critics who compare it to older zero-shot classifiers and a competing open-source project, OpenJeb, that reportedly reproduces similar behavior using a frozen Qwen-4B model without any new training. The video offers a critical look at how a well-funded startup is positioning classification AI as an architectural shift, weighing bold claims against unverified benchmarks and prior art.
📺 Source: Fireship · Published September 21, 2026
🏷️ Format: News Analysis
