Developer says Jev decision model beat Pokémon Red in under a week

TypeSafe AI's Jev AI model recently achieved victory over the Elite Four and Champion in Pokémon Red, entering the game's Hall of Fame. Unlike other chatbots that took weeks or months to complete the task, Jev could only choose from a pre-designed list of options. Anthropic's Claude Opus 5 monitored the game log and adjusted the options and wording, effectively acting like a coach.

Developer Andrew Boyd announced the project on X, and it was livestreamed in a browser or terminal. Jev moderated the chat, and viewers sent tips that were incorporated into Jev's decision-making process. Using a traditional LLM, Claude Opus 5 reduced the text sent to Jev by about two-thirds and improved accuracy. The reliance on multiple models and human input demonstrates both the strengths and limitations of decision models.