Show HN: Jev Plays Pokémon Red
An AI agent named Jev successfully completed Pokémon Red in 38 hours for a total token cost of just $1.65.
summary is enough The project is a fun proof-of-concept for cheap LLM decision-making, though commenters thoroughly dissect the heavy scaffolding required to make it work.
summary
A developer created an AI agent named Jev to play Pokémon Red using low-cost token decisions paired with a structured game harness. Over 37 hours and 16,150 decisions, Jev finished the game for $1.65. Commenters noted that while impressive for its cost, the agent relied heavily on predefined milestones and pathfinding assistance, frequently getting stuck in loops and making questionable strategic choices.
key points
- Jev completed Pokémon Red in 37 hours across 16,150 decisions, costing a total of $1.65 in API tokens.
- The project utilized a structured game harness with predefined high-level goals and pathfinding tools rather than raw end-to-end vision.
- Discussion focused on the trade-offs between cheap, fast decision models and heavy scaffolding versus full frontier models or traditional RL.
who should read it
AI developers and hobbyists interested in low-cost LLM agents and game automation benchmarks.
discussion positive
consensus
Commenters are amused and impressed by the low cost and real-time execution, but agree that the heavy harness and preset goals do most of the heavy lifting.
pushback
Some critics argue that traditional scripts or simpler rule-based approaches could accomplish the same feat without incurring API costs.
I've been watching GPT_Plays_Games on twitch and with Astra now it can beat pretty much all of them with very few mistakes, with vision only. One downside is that when pathfinding it can only plan high level paths, wait for result, then try again if it fails. Its a ~10-30 second loop.
Has anyone tried a combined LLM + Jev? So the LLM directs the high level goal ("reach the door of the pokemon center while avoiding NPCs") then allow it to instruct Jev to do the actual movements? That seems like a good balance between the high-level, slow planning of the LLM to the fast but limited Jev. That …
It proposes a practical architectural pattern of combining slow, high-level LLM planners with fast, low-level execution models like Jev.
Generated by an LLM from the article and thread. Check the sources before relying on it.