Show HN: Pac-Bench – How well can models one-shot a Pac-Man game?
Pac-Bench evaluates how effectively various AI models and harnesses can one-shot generate a functional, faithful Pac-Man game from a single prompt.
summary is enough The benchmark site is a neat novelty, but the core concept and model comparisons are fully captured in the comments.
summary
Pac-Bench is a lightweight benchmark testing whether AI models can write a complete Pac-Man game in a single HTML page from a basic prompt. While top models produce surprisingly polished results, commenters debate whether this constitutes a valid test given that Pac-Man clones saturate training data. Critics note that while the visual output looks convincing, fine-grained details like authentic ghost pathfinding often fail on the first try.
key points
- Pac-Bench tests one-shot generation of a Pac-Man clone inside a single HTML page using a minimal prompt.
- Commenters note that models frequently botch subtle mechanics like ghost AI behavior despite generating working visuals.
- Some users argue the benchmark is flawed because Pac-Man is heavily overrepresented in training data, acting more like regurgitation than reasoning.
who should read it
AI engineers, developers interested in code-generation benchmarks, and hobbyists tracking LLM coding capabilities.
discussion positive
consensus
Commenters generally agree that while models like Claude Opus 5.5 show impressive progress in one-shotting games, Pac-Man is too common in training data to be a rigorous benchmark.
pushback
A minority argue that the exact initial output matters less than how easily models can correct specific mechanic flaws through quick prompt iterations.
Interesting, recently I am working on my own clone of Pac-Man. LLM implementations lose lots of details. They are not 1:1 replication of the original game. For example, the behavior of the ghost is not the same as the original. If I have not implemented the game myself, I can not tell the differences. What LLM produced look like the original game, but they are not.
It highlights that LLM game clones often look correct at a glance while failing authentic gameplay details that only a careful player would notice.
Generated by an LLM from the article and thread. Check the sources before relying on it.