Labonne demos one model playing multiple games in parallel from screenshots alone

On October 6, Maxime Labonne shared a series of hands-on experiments on playing and building games with LLMs. The centerpiece: a single model running multiple games simultaneously in parallel (e.g., Wordle, Quick, Draw!). He argued this matters far more than "using LLMs for text generation" and will have a much bigger impact on the video game industry.

Confirmed

A single model can play multiple games in parallel, with the input path being screenshots only: the model relies on built-in OCR to read the screen and make correct decisions, and its Wordle performance was "very stable."
Labonne found that adding text input provided no extra benefit—the pure vision path was sufficient.
In an experiment using an LLM to generate Flappy Bird, the first version turned out poorly; with better prompts and complete input information (such as game state), the model did very well.
He also compared image inputs and found that for Flappy Bird, images didn't help—the text alone already carried enough information.

Why it matters

Parallel multi-game play with screenshot-only input shows that a vision-reasoning loop can already run stably on games with no text interface, offering a viable path for LLMs to directly take over traditional game runtimes.
Both experiments point to the same conclusion: model performance often hinges not on the model itself but on whether the input information is complete—feed it full game state and results improve, a lesson directly relevant to game development and agent design.

2026-10-06 ~ 2026-10-06 · 5 related posts

Primary sources

One model plays many games in parallel from screenshots alone — text input adds nothing — maximelabonne · 2026-10-06 04:45 (UTC+08:00)
Prompting lesson: a Flappy Bird flop became great once the prompt supplied all needed inputs — maximelabonne · 2026-10-06 05:04 (UTC+08:00)

One model runs multiple games in parallel, researcher calls it bigger than LLM text generation — maximelabonne · 2026-10-06
[source] One model plays many games in parallel from screenshots alone — text input adds nothing — maximelabonne · 2026-10-06
Vision models beat Wordle from screenshots alone: OCR plus reasoning lands correct guesses — maximelabonne · 2026-10-06
Better prompting fixed my LLM-built Flappy Bird: full inputs beat images — maximelabonne · 2026-10-06
[source] Prompting lesson: a Flappy Bird flop became great once the prompt supplied all needed inputs — maximelabonne · 2026-10-06