A play experiment of "Pokémon Red" using the decision-making AI model "Jev" has resulted in reaching the Hall of Fame in 37 hours and 40 minutes. The total cost of using the AI was kept to $1.65 (approximately 260 yen).
PLUS ULTRAModel ReleasesJevTypeSafe AI
Decision-making AI "Jev" Conquers Pokémon Red, Achieving Low Cost and High Speed with a "System 1" Architecture that Skips Text Generation
PLUS ULTRA by Amenoyomi
This article is a translation. Read the Japanese original
While Jev is a Transformer-based model, it differs from Large Language Models (LLMs) that perform text generation. Rather than "System 2" thinking, which involves deliberate and detailed output generation, it employs a "System 1" approach specialized for rapid, intuitive judgment, designed to perform no text generation whatsoever.
In its specific processing, the model takes the current state and a typed query as input and returns typed judgments—"Choice," "Score," and "Null"—along with their probabilities. By eliminating the computational load of autoregressive models that generate text sequentially, response times are extremely short, ranging from 70 milliseconds to 500 milliseconds.
This architecture enables a low price of $0.042 per million input tokens, with output being free of charge. The developer, TypeSafe AI, claims that in their own evaluations, Jev achieved speeds and cost reductions hundreds of times greater than LLMs such as GPT-5.6 Terra.
However, because Jev specializes in high-level decision-making, it cannot directly operate game button inputs on its own; in early experiments, it was unable to leave Pallet Town. Consequently, the setup relies on auxiliary systems, such as dedicated programs to read information from RAM, the A* algorithm to calculate paths to destinations, and walkthrough guides.
Sources
- GIGAZINE
- marktechpost.com「TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text」