---
title: "Notes on LLM Clock Face Limitations"
description: "LLMs struggle to generalise clock face generation beyond the standard 12 divisions, highlighting a specific gap in reasoning versus training data memorisation."
type: note
pubDate: 2025-11-18T00:00:00.000Z
topics: ["llms", "generative-ai", "failures", "generalisation"]
canonicalUrl: "https://actuallymaybe.com/blog/notes-on-llm-clock-face-limitations/"
---
> LLMs struggle to generalise clock face generation beyond the standard 12 divisions, highlighting a specific gap in reasoning versus training data memorisation.

An interesting thing I learnt, inspired by a comment on LLM clocks, is that LLMs struggle to generate clock faces with more than twelve divisions. Perhaps there is a type of prompt that would work, but when you ask in a straightforward way it simply does not. The model tries to cheat by changing the numbers rather than the number of divisions.[^1]

A direct approach does not work.

Using historical examples of other clock face divisions also did not work.

I am sure there is some prompt or instruction that would eventually work, but out of the box it is quite interesting how LLMs and diffusion models remain tied to this pattern. Clearly it is a training data issue, but at the same time it nicely shows a gap in generalisation. I am confident that if I gave similar instructions to an eight year old they could easily draw a rough version of what I am looking for.

[^1]: See [Hacker News discussion](https://news.ycombinator.com/item?id=45930664).

## Further reading
- [Notes on Visual Interfaces, AI Tooling, Fast LLMs and Decision Models](https://actuallymaybe.com/blog/notes-visual-interfaces-ai-tooling-fast-llms-decision-models/) — Why agentic work is stuck in text, and what canvases, fast models, branching version graphs, and decision models might add.
- [Building Explainers Random Thoughts - Drone 101](https://actuallymaybe.com/blog/building-explainers-random-thoughts-drone-101/) — An experiment in interactive explainers: drone control, Fourier, Bret Victor, philosophical prompts, accessibility, and learning with LLMs.