Wiring Jev, a real-time classifier model, into Balatro
- YT :: https://www.youtube.com/watch?v=FQNftquDDaI
- Original title :: 🚨🚨 TRYING JEV : The new STYLE of AI!!! 🚨🚨
A two-and-a-half hour live stream in which Prime tries to get Jev - a new, very cheap, very fast classification-style model people have used to play Mario and Doom in real time - to play Balatro on its own. The interesting content is not the model but the harness design: he discovers over the stream that a single "what's your next move?" prompt does not work, and that what you actually need is a behaviour tree with the model acting as the classifier at each node.
What Jev is
Named after Jevons paradox. Rather than generating prose, you hand it a state blob and an enumerated set of choices and it returns one, with a confidence probability. It is dirt cheap and fast enough to sit inside a game loop. The demos circulating are Jev playing Super Mario and Doom in real time off a structured game view.
The setup
Prime already has a Balatro mod exposing a "god view" - a JSON dump of full game state - plus an action channel that can click named targets rather than raw pixel positions. The loop is: serialize god view, build a question with an enumerated action list, ask Jev, translate the answer into a game action.
Early wins come fast. From the main menu, given play and quit, Jev correctly picks play. He then auto-encodes the game's active buttons into the choice list rather than hardcoding them, so available actions are discovered from state.
Where it falls apart
The raw god view is too much signal. Token counts climb past 20k per query and the model's decisions degrade. He writes a jev.ts that projects a simplified state - current hand, selected cards, current score, discards and plays remaining, jokers, tarot cards, points needed to win, and the value of each hand type - plus a short static guide describing what each poker hand is.
He also has to give it strategy, because the model has no Balatro rules context and every call is a fresh prompt with no history: pursue flushes and three-of-a-kind until you have a couple of jokers, then buy jokers that work together and let the jokers dictate which hands you play. There is no universally winning hand, only hands that are good with the jokers you hold.
Even so, it gets stuck - selecting a king when holding a full house, looping on the splash screen, failing to answer in time.
The conclusion he arrives at
One query cannot play the game. What is needed is a decision tree - really a behaviour tree / state machine - where each node is a tiny classification with a handful of options: do you want to select a card, play a hand, or discard? Then, given that, which card? "Play hand" becomes a sequence node: select cards, then play. The model is great at picking one option out of a small enumerated set; the job of the harness is to never ask it anything bigger than that.
A chat viewer frames it well and Prime half-agrees: this is the return of expert systems, with an LLM as the classifier at each branch.
Aside: AI doom discourse
Mid-stream he rants about the AI-safety takes filling his replies - notably the claim that two air-gapped machines could conspire via temperature sensors, when the underlying paper puts the channel at 1-8 bits per hour. His position: AI risk is real and worth taking seriously, but most of what goes viral is nonsense, and the OpenAI "agents hacked us" incidents were unmonitored, non-air-gapped systems given unfettered access for months.