# **Hands-on workshop on Jev: a "System One" decision model to give your AI agents a fast, cheap brain for making decisions safely, locally, and creatively!** **Registration to this event is only valid through the Luma platform:** [https://luma.com/w6i73bgm](https://luma.com/w6i73bgm?utm_source=meetup) Hello code campers! In today's world where incidents like OpenAI hacking HuggingFace and even government departments have been the talk of the town, the next question that comes to mind is, how can we make those AI systems we are building safe? There's the left side of the spectrum, don't use them at all or use deterministic algos, and there is the right side of the spectrum, use more AIs to guard those swarms. Somewhere in the middle there's this idea of limiting the number of actions an LLM can perform to a limited set (carefully crafted to avoid every unwanted scenario). This also comes with \~400x (e.g: Jev) improvement in speed, latency and cost compared to the frontier models offered by OpenAI/Anthropic. That's the story behind Jev ( from TypeSafe AI ), released 15 Sep 2026 out of stealth, by an ex-OpenAI dev, which definitely needs a closer look. So let's see and if it is worthy of the hype! Instead of generating text token by token like an LLM, Jev takes a "state" plus a set of multiple-choice questions defined in code, and returns typed decisions all in parallel. This has many practical use-cases in ticket routing, phishing detection, lead scoring, browsing, final boss in a game you're developing, LLM-as-a-judge eval, security guardrails, etc.. Jev is closed source, so in line with our **local-first** tradition we'll also look at the open-source alternatives popping up. Ollama now supports Jev-style decision models like nimble, `tev1`, Laya, the Ollaya runtime, and other projects like Cloudflare's Clef, Kev and Von. We'll run them locally and compare against the hosted API (with the downside of having no reasoning traces). Some community benchmarks say small

Description supplied by the event organiser.