AI beats Pokémon Red in under a week
The AI That Conquered Pokémon: How Jev Beat the Elite Four
Imagine an artificial intelligence not just chatting, but actively playing a beloved video game to its ultimate victory. This isn’t science fiction; it’s the reality achieved by TypeSafe AI’s Jev, an experimental decision model that claimed to beat the Elite Four and the Champion in Pokémon Red.
The feat itself is staggering. While many large language models take weeks or months to grapple with a complex task, Jev achieved this conquest with a fundamentally different approach. Jev is not a chatbot or a large language model (LLM); it is a specialized decision model. Instead of generating free-form text, Jev operates by selecting the optimal option from a predetermined list of choices, calculating the best move based purely on probability.
But even this specialized AI couldn’t succeed alone. Jev required a brilliant collaboration with another powerful AI. Anthropic’s Claude Opus 5 stepped in, monitoring the game log and acting as an unseen coach. Opus helped Jev navigate the labyrinth of the game, modifying the available options and adjusting the wording of the choices to ensure Jev stayed on track and avoided catastrophic errors.
The process was less about reading the screen and more about strategic decision-making. The harness logs revealed thousands of attempts—failures including walking into locked entrances dozens of times or losing to the Champion’s Alakazam. This heavy log of mistakes became the raw data that the collaboration refined.
The success of Jev’s journey demonstrates a fascinating tension between different types of AI. While some models excel at open-ended creative tasks, specialized decision models offer a unique pathway to complex problem-solving. The collaboration highlighted that the true power lies not just in the intelligence of a single system, but in the synergy between specialized models.
Beyond the Jev experiment, other developers explored similar frontiers. Christian Mathiesen undertook a separate run where his system focused on reading the game’s memory to list legal options, allowing Jev to pick a move without directly writing to the game’s memory. This approach offered a different constraint on the AI’s interaction with the game world.
Another path involved training a smaller world model. One developer, stmonty, used a graphics card and thousands of frames of gameplay to train a model that could select a starter in Pokémon Red, learning purely from visual information from screenshots. These varied experiments underscore the depth of what AI can achieve when given the right constraints and specialized tools.
The viral nature of the Jev project, which quickly drew developers into the arena, proved that when you combine specialized AI with human guidance and iterative feedback, even the most daunting challenges become achievable. It signals a new direction for AI development: moving beyond simple text generation toward highly focused, strategic, and collaborative problem-solving.
Buy on Amazon
Credit
