Jev Plays Mario Using NES Memory

Zach A demonstrates Jev, TypeSafe’s System One model, playing Mario by reading NES memory rather than pixels. For each move, it receives the game state and a typed question, then returns a choice with probabilities.

View on X timeabt 300ms each

Zach A@rherton𝕏

the obligatory AI plays Mario, Jev version Jev is TypeSafe's System One model. u send it state and a typed question, it sends back a choice with probabilities. so every move here is one question: run, hop, jump, wait or go back it reads the game from NES memory, no pixels. 193 calls to clear 1-1, abt 300ms each right side is the real request and answer for every decision. the game is paused while

Sep 18, 2026X postsView on X

The run cleared 1-1 in 193 calls. Each call took abt 300ms.

Also filed under Games & real time

Steer a 3D Box with Goals
The build steers a 3D box with plain-language goals. An LLM plans each instruction, while TypeSafe Jev makes move, turn, and jump decisions in real time.
Benchmarks Jev on Pokémon Red
The project benchmarks Jev, TypeSafe’s fast decision model, against Jev paired with a GPT-6 Sol planner for playing Pokémon Red.
Plain-English Cellular Ecosystem Simulator
LifePot is a cellular automaton-inspired ecosystem that users describe in plain English. Jev turns those inputs into species, feeding relationships, reproduction strategies, and environmental conditions.
Mapping Children's Behavioral Intent in Games
Mores Research used Jev to map children’s behavioral intent in games and examine how it correlates with real-world behavior. It reports mapping intents across 45+ real sessions involving 20 child users over a long horizon.