Amateur uses Claude Code to find planet candidate; TESS checks
A 28-year-old product manager used Claude Code to find a possible planet. He ran 74 tests to disprove himself. Now NASA TESS is checking.
Research · Source: Gizmodo
What happened
Pavel Rabtsevich is a 28-year-old product manager living in Spain. He has no formal training in astronomy and does not own a telescope. Instead, he used AI coding agents to sift through public data from the Transiting Exoplanet Survey Satellite, a space telescope operated by MIT and NASA. He fed measurements from 126,000 stars into Anthropic's Claude Code, OpenAI's Codex, and TypeSafe's Jev. He prompted the AI systems to look for regular dips in starlight. These dips often indicate an orbiting exoplanet blocking the light.
The AI tools helped him narrow the search down to a single star called TIC 4206066. It is located roughly 116 light-years from Earth. The data showed a slight dimming every 3.18 Earth days. If a planet is causing this shadow, it would be about 1.4 times the size of Earth. Its surface temperature would be around 1,000 degrees Fahrenheit. Rabtsevich notes this is a candidate, not a confirmed planet. A nearby binary star system could also cause this exact signal.
Rabtsevich did not just accept the first output. He ran 74 separate tests with Claude Code to actively disprove his own hypothesis. In one crucial test, he hid a year of data from the model. He fitted the model to two years of observations and asked it to predict the dimming schedule for the hidden third year. It predicted the dips perfectly. Because of this rigorous approach, the TESS mission approved an official observation program. From October 31 to November 26, the telescope will measure the star every two minutes to check his predictions.
Key facts
- 126,000 — Number of stars in the initial TESS dataset analyzed by the AI tools
- 74 — Number of tests run with Claude Code to try and disprove the hypothesis
- 3.18 — Number of Earth days between each dip in the star's brightness
- 1.4 — Estimated size of the possible planet compared to Earth
- October 31 — Start date for the new TESS observations to check the predictions
Why it matters
This changes how we think about AI in scientific discovery. The AI did not just write a polite email or summarize a PDF document. Claude Code handled the heavy lifting of execution. It downloaded data archives, wrote analysis scripts, fitted possible transits, and generated plots. The human provided the judgment, the pass-fail criteria, and the skepticism. Rabtsevich used models like Opus 5.5 and Fable 5.1 to do the tedious work that used to require endless spreadsheets. This is a clear blueprint for AI-assisted research. You separate the execution from the scientific method.
The second-order effect is the total democratization of hard science. You no longer need a PhD or a massive lab budget to parse galactic quantities of data. A product manager with an internet connection processed 126,000 stars in two weeks. Legacy institutions will soon face a massive flood of amateur discoveries. They will need entirely new vetting systems to filter the noise from the actual breakthroughs. The gatekeepers of science are losing their monopoly on discovery.
For builders
Build rigorous validation loops
Do not trust the first output from any model. Rabtsevich ran 74 tests to break his own hypothesis. Founders who build automated cross-validation into their AI products will win the enterprise market. Those who rely on zero-shot answers will lose trust and churn users.
Target massive public datasets
Government agencies sit on mountains of unanalyzed data. Builders can use agents to find valuable signals in this noise. The cost of data parsing has dropped to zero. Bootstrapped founders can now compete with well-funded research labs.
Separate execution from judgment
Let the AI write the scripts and plot the graphs. Keep the human in charge of the scientific method and decision rules. Products that force this separation will earn user trust. Tools that try to automate the final judgment call will fail.
My take
Everyone talks about AI hallucinations, but this guy used AI to do rigorous science. He succeeded because he treated the model like a tireless intern, not an all-knowing oracle. If you want to build real products, stop asking AI for the right answer and start asking it to prove you wrong.
Original reporting: Gizmodo. This is my rewrite and opinion.