Jev AI Plays Minecraft, Subway Surfers, and Drives Cars: Demos Reviewed
A hands-on look at community demos of Jev, a fast AI model, controlling Minecraft, Subway Surfers, drones, and driving sims in real time.
What is Jev, and why does it matter for real-time AI?
Jev is a model released by Typesafe AI, a company founded by Diego Almeida, who previously worked on ChatGPT and RLHF at OpenAI. Typesafe pitches Jev as a “system one model,” a different architecture built for fast, automated decision-making rather than conversational reasoning. Instead of generating long chains of text, Jev outputs decisions with probabilities and confidence scores, which lets it respond in a fraction of the time a typical large language model needs. That speed is the whole point: it makes Jev usable inside real-time loops like games, drones, and driving simulators, where a system has milliseconds to decide what happens next.
TL;DR
- Jev is built for speed, not conversation, using a non-autoregressive approach that returns structured decisions instead of generated text, which Typesafe claims makes it roughly 100 times faster than standard LLMs.
- Pricing is aggressive, with input tokens listed at $42 per billion tokens and output tokens effectively free, since Jev’s outputs are short decisions rather than long text completions.
- A Minecraft bot built by a developer using Jev reportedly ran for two minutes on about 150,000 tokens for roughly one cent, reacting to threats like zombies and nightfall without hardcoded “if zombie, then flee” logic.
- A Subway Surfers demo showed Jev making split-second jump, duck, and lane-change decisions, the kind of fast-twitch gameplay that normally exposes latency problems in AI-controlled agents.
- A simplified self-driving demo was built inside a simulator in under an hour, with Jev choosing between accelerate, brake, maintain speed, or change direction based on live simulator state.
- A drone navigation demo inside a simulated obstacle course was built in about 15 minutes and cost around 10 cents, with Jev picking movement actions from structured data like position and distance to obstacles.
- None of these demos involve real-world hardware, cameras, or safety systems, so they show decision-making speed and adaptability in simulation, not production-ready robotics or autonomous vehicles.
How does Jev actually make decisions in these demos?
Every demo follows roughly the same loop. A game or simulator sends Jev a compact, structured snapshot of the current state: things like player health, enemy positions, time of day, drone speed, or distance to an obstacle. Jev is given a limited menu of allowed actions and picks one based on that snapshot. The program executes the chosen action, updates the state, and immediately asks Jev again. This cycle repeats continuously.
The key difference from a typical LLM-driven agent is what happens inside that loop. A conventional large language model would need to generate a written explanation or reasoning chain before landing on an action, which takes time and costs tokens. Jev skips that step. It returns a decision directly, more like a classifier or a lightweight control policy than a chatbot. That’s what allows the loop to run fast enough for something like Subway Surfers, where a half-second delay means a missed jump.
Is the Minecraft demo actually impressive?
Specialized Minecraft bots that navigate terrain, fight mobs, or farm resources already exist, and they’re generally faster and more reliable than any general-purpose model would be. So the interesting part of the Jev Minecraft demo isn’t that an AI can play the game. It’s that a general model, given only a live text description of the game state, adapted to changing conditions without anyone writing specific rules for each scenario.
According to the demo, when zombies appeared or night fell, the bot recognized that staying put was risky and chose to retreat, without a developer coding an explicit “flee at night” instruction. That’s a meaningful distinction: it points to general situational reasoning rather than a scripted response. The reported cost (about one cent for two minutes of play, using around 150,000 tokens) is also notable, because it suggests persistent AI-controlled game characters could run continuously without the cost or latency overhead of calling a full-sized LLM on every tick.
What do the drone and driving demos actually prove?
The self-driving demo was a simulated environment, not a real car on a real road. The developer built a simplified driving simulation, complete with a car, road, traffic, and a small set of controls (accelerate, brake, hold speed, change direction), in under an hour. Jev received the current state of that simulation and picked an action, and the loop ran fast enough to look like real-time driving.
The drone demo worked the same way: a simulated obstacle course, built in about 15 minutes, fed Jev structured data on the drone’s position, speed, and distance to obstacles and the target. Jev chose from a fixed set of moves (forward, turn, climb, descend, hover), and the simulator updated the drone’s position accordingly. The reported cost was around 10 cents for the demo.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Neither demo used cameras, real sensors, or safety-critical control software, and neither operated real hardware. What they do demonstrate is how quickly a developer can bolt reactive, language-model-driven decision-making onto an existing simulation, without collecting a specialized dataset or training a dedicated model from scratch. That’s a meaningfully lower bar than building a full self-driving stack or a certified drone autopilot, and it’s worth being precise about the difference.
Why does cost per decision matter more than raw intelligence here?
Most AI benchmarks focus on how smart a model is on a hard reasoning task. These demos highlight a different variable: how many decisions per second an AI can make, and how much each decision costs. In a real-time system, a model that’s slightly less “smart” but orders of magnitude faster and cheaper can be more useful than a slower, more capable one, because the loop only works if the AI can keep up with the game or simulator clock.
That’s why the pricing details matter as much as the demos themselves. Typesafe’s stated pricing (input tokens at $42 per billion, output tokens effectively free) is built around the assumption that Jev’s outputs are short structured decisions, not long generated text. If that pricing and speed hold up outside of launch demos, it changes the calculus for anyone building agents that need to act continuously rather than respond once per user prompt.
Frequently Asked Questions
What is a “system one model” and how is it different from a normal LLM?
Typesafe describes system one models as an architecture optimized for fast, automated decisions rather than conversation. Instead of generating text sequentially, Jev outputs decisions with associated probabilities, which removes the step-by-step token generation that makes standard LLMs slow for real-time control tasks.
Did Jev actually control a real drone or a real self-driving car?
No. Both demos ran inside simulators built specifically for the demonstration. Jev chose from a small set of predefined actions based on simulated sensor data; it did not process camera feeds, control real hardware, or operate in any safety-critical environment.
How much did the Minecraft and drone demos reportedly cost?
The Minecraft demo reportedly ran for about two minutes using around 150,000 tokens for roughly one cent. The drone obstacle-course demo reportedly cost about 10 cents. Both figures come from the developers who built the demos, not from independent benchmarking.
Why does Subway Surfers matter as a test case?
Subway Surfers requires very fast reactions (jumping, ducking, switching lanes) with little margin for latency. Using it as a demo stresses whether an AI model can make decisions quickly enough to keep pace with a fast-moving game, which is a harder real-time test than turn-based or slower-paced games.
Is Jev ready for production robotics or gaming products?
Based on the available demos, Jev shows promise as a fast decision-making layer for prototyping reactive behavior in simulations. It hasn’t been shown operating real hardware, handling safety-critical systems, or working at scale in production games, so treating these as early proofs of concept rather than finished products is the more accurate read.
