Creating black hole simulations with Codex
Summary
OpenAI's Codex reduced black hole simulation from 10 days to minutes by automatically generating and testing 10 different numerical algorithms, with the best performing 1000x faster than standard methods. This demonstrates AI's potential to accelerate scientific discovery by automating algorithmic exploration.
Key Takeaways
- Use AI agents to automate algorithm exploration: Codex generated 10 different numerical schemes and automatically evaluated which performed best, reducing 10-day iteration cycles to minutes.
- Identify computationally intractable problems as AI opportunities: Black hole plasma simulation was impossible with standard approaches until Codex found novel algorithms that made previously impossible simulations possible.
- Leverage AI for unstable numerical scheme optimization: Rather than manually debugging unstable standard methods, use AI to systematically explore alternative schemes that maintain stability and performance.
- Measure AI agent effectiveness through iteration speed: The 1000x performance improvement was secondary to the primary win—reducing manual algorithm testing from 10 days to minutes through automated exploration.
- Apply AI-driven discovery to frontier scientific problems: Codex excels when domain experts don't know the optimal solution beforehand, making it ideal for exploring uncharted problem spaces with high uncertainty.
Related topics
Transcript Excerpt
The Event Horizon Telescope project has been around for many years now. In 2017, we created the first image of a black hole. And this year is very exciting, because we are making the first video of a black hole. The only reason we can see black holes is because of this very hot plasma falling onto them. In order to simulate this plasma around black holes correctly we need to follow electrons and ions around the magnetic field. And this is just a computationally intractable problem. With Codex, we are now able to find new algorithms that can simulate them. So they make previously impossible simulations possible. CK was part of the Event Horizon team that captured the first still image of a black hole. Around the department, he's known as the black hole simulation guy. This thing here, this …