ai & weekendRead on Notion ↗
GODSPEED
The Jev Build Jam
In this guide
Description
A researcher behind ChatGPT spent two years building a different kind of AI. Come build with it.
Welcome to GODSPEED: four hours of building with Jev, TypeSafe AI’s new model for fast, structured decisions. We’ll unpack the research, make our first call and turn an idea into something the internet can play with.
The launch is fresh. The experiments are wide open. Bring your curiosity. Leave with a demo people want to try.
4 hours · Teams of 2–4 · Build + demo
Quick links
Meet Jev: the model and the company behind it.
Read the launch: the idea, the experiments and the results.
Get started: playground, API and Python setup.
Build through Vercel: the AI Gateway integration.
What really happened
On 15 September 2026, TypeSafe AI introduced Jev, its first “System One” model. The founders are Diogo Almeida, Sasha Sheng and Erik Gafni, with backgrounds spanning OpenAI, Google Brain, Meta/FAIR and production AI systems. Almeida worked on the instruction-following research behind ChatGPT. Meet the team
Their question: if your software needs a decision, why make a model write an answer first?
Jev produces structured decisions and probabilities. TypeSafe describes a new architecture and a training approach called Reinforcement Learning for Calibrated Decisions (RLCD), designed around decisions and uncertainty. The launch experiments include Doom and Wikipedia races, alongside evaluations of practical software tasks. Its reported speed and cost gains come from specific company-run comparisons; we’ll explore what happens in our own builds. Launch and evaluations
The early response has already spilled into real projects. TechCrunch reported that demand briefly overwhelmed the API and described developers trying Jev in existing workflows. This is a good moment to form your own opinion by building something. Launch coverage
Understand in five minutes
Give Jev a state, information about what is happening and ask small, specific questions. It returns answers your code can act on.
| Primitive | Try asking | What comes back |
|---|---|---|
| Choice | Should the character hide, follow or attack? | A selected option, probabilities and confidence |
| Score | How playful is this idea on our scale? | A score and uncertainty information |
| Noul | Does this message ask us to stop? | A value between 0 and 1 |
For a game, send a text description of the scene and the player’s instruction. Jev picks an action; your game engine carries it out. You can ask several independent questions about the same state in one call and combine the results in code. How the primitives work
A typed answer can still be a wrong answer. Test awkward examples, watch the confidence values and decide when your app should ask for help. That is part of the experiment.
A few ideas to get started with
Reflex Arena — give a character an instinct. Build a tiny browser game where visitors type instructions such as “protect the blue player.” Start with one character, a grid and three possible actions. Put the model’s response time beside the game.
Wikipedia Speedrun — race the model. Pick a starting page and a destination. Jev chooses the next link while a human tries to beat it. Start with a small prepared set of pages; compare routes, hops and time.
Vibe Atlas — make a map of meaning. Plot films, project ideas or fictional characters on axes such as “chaotic ↔ calm” and “practical ↔ strange.” Change the axes and watch your canvas reorganise. Start with 30 items and two scores.
The Agent’s Brake Pedal — build the decision to stop. Give a simulated agent four choices: continue, retry, ask or stop. Feed it ambiguous requests and failed tool results. Show the room when it keeps going—and when it knows to ask.
Bring your own idea, too. Aim for one interesting decision inside an interaction people can actually try.
Agenda
A four-hour build jam, from first experiment to live demo.
| Time from start | What we’re doing |
|---|---|
| 00:00–00:20 | Meet Jev: the research idea and a working example |
| 00:20–00:40 | Make a first call, find teammates and pick a build |
| 00:40–02:20 | Build the smallest playable or usable version |
| 02:20–02:35 | Take a break and try someone else’s project |
| 02:35–03:20 | Test edge cases, measure speed and polish the interaction |
| 03:20–04:00 | Demo time: show what you made and what surprised you |
Aim to finish with a working prototype and a two-minute demo. Show the input, the decision and the result. Include one example that failed—you’ll learn as much from that as from the perfect run.
What you need
- A laptop and charger, plus your usual coding environment.
- A working API connection. Follow the TypeSafe quick start or Vercel Gateway guide and try one request beforehand. Direct TypeSafe access may involve a waitlist.
- Basic Python or JavaScript experience, or a teammate comfortable making API calls. The Python SDK requires Python 3.10 or newer.
- A small API budget. Check the current pricing for your chosen route before building.
- One idea you’re curious about. Come solo or with friends; teams of 2–4 work well. Designers and non-coders can shape the interaction, create examples and test the results.
No GPU or model-training experience needed.
More links for curious ones
- TypeSafe’s manifesto — the bigger vision: intelligence as a component developers can build with.
- Confidence guide — understand uncertainty before making decisions with it.
- Implementation patterns — ways to combine the primitives.
- YouTube: Diogo Almeida on Jev and RLCD — a founder talk on AI Engineer.
- X: TypeSafe’s launch announcement — the launch conversation.
- X: Jev arrives on Vercel AI Gateway — an access update.
- Hacker News: Introducing System One Models and Jev — the enthusiasm, scepticism and technical debate.
- Show HN: Explore 2D semantic space with Jev — inspiration for a visual experiment of your own.