Show HN: Jev Plays Pokémon Red
An AI demo shows a lightweight “system one” model called Jev playing through Pokémon Red via a heavily engineered harness that handles pathfinding and high-level goals while Jev makes fast, cheap decisions at key points. Viewers find the result fun but note that Jev often makes poor or looping choices, sparking debate over how intelligent it really is versus how much the scripted scaffolding carries the experience. The project is framed as an early glimpse of ultra-low-latency, low-cost decision models for automation, with many suggesting that combining such models with slower, more capable LLMs could be a more powerful and general approach.
Demo setup and harness design
- Jev controls Pokémon Red via a TypeScript “harness” connected directly to the ROM, not just raw button presses.
- The harness provides higher‑level actions (e.g., move to coordinates, follow path, select from menus) and a series of ~33 high‑level goals plus sub‑goals to complete the game.
- Much logic is hard‑coded: pathfinding, textual milestones, selective context, and some menu/item handling. Several commenters argue this makes it closer to a scripted walkthrough with Jev as a stochastic decision source.
- The author explicitly acknowledges the heavy guidance and has open‑sourced the code.
Performance, latency, and cost
- The run has reached at least four badges and deep into the Elite Four for under ~$0.50–$2 in tokens, depending on final outcome.
- Jev is praised for being fast and cheap relative to frontier LLMs, enabling “interaction-speed” decisions.
Quality of decisions and limitations
- Viewers find it fun and “chill” to watch but note frequent poor choices: getting stuck in loops (doors, menus, Rocket hideout), wasting time on TMs/HMs, and questionable move choices (e.g., forgetting key moves).
- It struggles with complex, multi-step plans (e.g., properly preparing for and resourcing the Elite Four).
- A more minimal setup without pathfinding or extra context reportedly failed to even reach the starter Pokémon.
Model role: “System one” classifier
- Jev is framed as a fast, cheap, single‑shot decision/classification model rather than a full reasoning agent or RL system.
- Some see it as a promising “system one” to be paired with slower “system two” LLMs for planning, escalation, and recovery.
- Others argue its capabilities are overhyped, likening it to fancy RNG, pre‑LLM game AIs, or traditional classifiers/zero‑shot methods.
Broader ideas and future directions
- Suggestions include combining Jev with frontier LLMs for high‑level strategy, using vision inputs for generic game control, and applying similar decision models to domains like security alert triage or manufacturing SPC.
- Some commenters are excited about the broader concept of fast, general‑purpose, no‑training decision models; others stress that current intelligence levels are still far from replacing domain‑specific or frontier models.