# Plato's Chair

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Roger Newman-Norlund](https://cerebralvalley.ai/u/rogiedodgie)
- **GitHub:** https://github.com/rnorlund/MetaLesionLab
- **Demo video:** https://www.youtube.com/watch?v=mbonHCdFUcg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1170

**Plato’s Chair**

What would happen if you gave the original armchair philosopher an actual AI laboratory?

Plato’s Chair turns questions about knowledge, imagination, and the mind into experiments with Meta’s AI and real people. Visitors can explore the results, contribute their own drawings, generate fresh AI examples, and question an AI Socrates whose answers are grounded in what the experiments found.

We started with a chair—and ended up investigating how AI imagines, learns, and breaks.

Does AI have an “ideal chair”? Ask Meta’s Muse Image to draw a chair 200 times, and it returns strikingly similar chairs that are far less varied than people’s drawings. We found the same pattern across most everyday objects we tested. But ask for specific kinds, and AI produces more variety than people. Its imagination appears broader than its defaults suggest.

That opens a laboratory of ancient questions: Can we recognize something from its shadow? Do names naturally fit the things they describe? Does writing preserve what we mean? What does AI picture when asked to draw something no one has seen? For God, our repeated generations converged on a familiar image: a bearded man with a halo.

The experiments also explore how a mind develops and how it breaks. A small AI trained on human chair drawings develops patterns resembling the shift from children’s drawings toward adults’. On the clock-drawing task used in memory clinics, Muse Image performs comparably to adults in an aging study—a provocative task-level comparison.

Using Meta’s open Muse Glimmer, we progressively switch off neurons to investigate what fails first. In our experiments, pictures deteriorated before words. An interactive “damage zoo” lets visitors compare how different open image models fall apart, bringing a central method of neuroscience into public AI exploration.

Our site contains results from over a dozen experiments, most of which were planned in advance, including the predictions that failed. The project combines philosophy, human behavioral data, image generation, and direct model interventions in a lab anyone can enter.

Draw a chair. Challenge a finding. Ask Socrates what the evidence supports.

**Plato supplied the questions. Meta supplies the models. Everyone gets to experiment.**

Try it live: [https://rogiedodgie-platos-chair.hf.space](https://rogiedodgie-platos-chair.hf.space)

## More from Meta Global AI Developer Hackathon

- [Second Look Desk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1167)
- [Sawa Sentinels](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1168)
- [Laisan86](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1169)
- [ManualMate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1171)
- [Latent Futures](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1172)
- [Fahaam Qureshi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1173)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1170. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt