How Codex Helps an Astrophysicist Build Black Hole Simulations

Astrophysicist Chi-kwan Chan is using OpenAI's Codex to help write and build code for black hole simulations. According to a post on the OpenAI blog, the tool lets Chan focus on the science rather than getting stuck on implementation details, supporting research into extreme physics and helping scientists test Einstein's theory of general relativity in computational environments.

What happened

Astrophysicist Chi-kwan Chan has been using OpenAI’s Codex as a coding assistant to build simulations of black holes. According to OpenAI, the goal is to model extreme physical environments that are otherwise impossible to study directly.

Black hole simulations are computationally complex. They require translating dense mathematical physics, much of it rooted in Einstein’s theory of general relativity, into working code. Chan’s use of Codex is aimed at bridging that gap: letting a domain expert describe what they need and having the model help produce the implementation.

Why it matters

This is a practical example of AI coding tools being used outside their most common context. Most coverage of Codex and similar tools focuses on web development, data pipelines, or software engineering workflows. Scientific simulation is a harder, less forgiving domain.

A few reasons this is worth paying attention to:

• Domain experts are not always strong programmers. Scientists like Chan may have deep knowledge of physics but limited time or background to write optimised simulation code from scratch. A tool that accelerates that process has real research value.
• General relativity is not a simple test case. If Codex can assist with the mathematics and logic behind black hole physics, the range of scientific fields where it could genuinely help is broad.
• It shifts what AI coding tools are for. The story OpenAI is telling here is not about replacing developers. It is about giving researchers a faster path from hypothesis to working model.

According to OpenAI, the simulations Chan builds help scientists study physics under extreme conditions. That kind of research feeds into a wider understanding of how gravity behaves at the limits described by general relativity.

Our take

OpenAI published this story, so treat it as a case study with a promotional purpose. That said, the underlying point is credible. We have seen similar patterns with clients who are experts in their field but not in code: Codex and tools like it genuinely reduce the friction between knowing what you want and producing something that runs.

The black hole angle is attention-grabbing, but the more transferable lesson is simpler. If a researcher working on some of the most mathematically demanding problems in physics finds value in an AI coding assistant, the same logic applies to a business analyst trying to automate a report, or a marketer trying to pull data from an API they have never touched before.

What we would want to know more about: how much of the output Codex produces does Chan actually use without modification? That number matters a lot. A tool that writes 80 percent of a simulation correctly is useful. One that writes 40 percent correctly and requires deep review may add as much work as it saves. OpenAI does not provide that detail here.

What to do about it

If you or your team includes people with strong domain knowledge but limited coding confidence, it is worth running a structured test with Codex or a comparable tool (GitHub Copilot, Cursor, Claude). Pick one real task. Time how long it takes with and without assistance. The result will tell you more than any benchmark.

The practical takeaway: AI coding assistants are most valuable when the person using them knows exactly what the output should do, even if they struggle to write it themselves.

Source: OpenAI Blog