Every 3.18 days, a star sitting 116 light-years from Earth loses a tiny fraction of its brightness. The person who says he caught that pattern is Pavel Rabtsevich, a 28-year-old product manager based in Spain. He has no astronomy training and owns no telescope. His tools were AI coding agents and NASA data that anyone can download. Within a few weeks, a space telescope will check whether he got it right.
That is why the story deserves a closer look. Nobody is making a loose claim that a chatbot “discovered a planet.” Rabtsevich has made a precise prediction that can be proven wrong, with dates attached, and he has published those dates for anyone to see.
“If the star dims on time in November, this gets a lot more serious,” Rabtsevich wrote in an X post laying out his results. “If it doesn’t, I’ll post that too.”
The star and the signal
The star is TIC 4206066, which is slightly smaller than the Sun. The signal is a small, steady dip in its brightness every 3.18 Earth days. That is the kind of drop you would expect if an object were passing in front of the star.
Rabtsevich estimates that the candidate planet would be a little more than 1.4 times the size of Earth. A planet whose year lasts just over three Earth days would be very hot, and he puts its surface temperature at roughly 1,000 degrees Fahrenheit. That is well above Mercury’s temperature and hot enough to melt lead.
He also found a second, fainter signal. The AI-assisted search picked up another dimming pattern on the same star, repeating every 11.13 Earth days, which could mean a second exoplanet that nobody has spotted before. Rabtsevich says the case for this one is weaker, and he deserves credit for saying so at the outset.
Three AI agents, one person in charge
His raw data came from the Transiting Exoplanet Survey Satellite (TESS). The space telescope is run by MIT astronomers and was launched by NASA in 2018. TESS looks for planets beyond our solar system by measuring how bright stars are over time. When a star dims at regular intervals, a planet may be crossing in front of it and blocking part of its light.
Rabtsevich used only a small part of that archive, recorded between November of last year and early January. He said it covered more than 126,000 stars. He gave the data to three agents: TypeSafe’s Jev, Anthropic’s Claude Code and OpenAI’s Codex. He prompted them to search for dimming patterns, and with each round the list of candidates got shorter.
He made it clear that he, not the models, was making the decisions.
“I didn’t take the models’ first answers,” Rabtsevich said. “I pushed them, round after round, to improve their own methods, and checked what they gave me against other models. “I decided which tests to run, judged whether a result was strong enough, and chose what to keep.”
If any AI-driven workflow deserves trust, it is one like this. Checking one model’s output against another’s is the cheapest protection against an answer that sounds confident but is wrong, and most people don’t bother to do it.
74 tests designed to prove him wrong
After TIC 4206066 emerged as a candidate, Rabtsevich checked older records. He pulled the star’s TESS observations from 2020 and 2018. Both years showed the same dips at the same frequency.
He then used Claude Code to run 74 tests, each one meant to knock down his own hypothesis. In one test, he gave the agent two of the TESS datasets and asked it to predict when the dimming would show up in the third. It predicted correctly in all three cases.
Rabtsevich sent the report to MIT and asked for TESS to observe TIC 4206066 the next time the telescope covers that part of the sky. MIT approved the request. Between October 31 and November 26, the telescope will measure the star’s brightness every two minutes.
Earlier this week, he published the exact times the star should dim during that period if a planet really is transiting it on a regular schedule.
A dip in brightness is not yet a planet
Some context helps here. Astronomers have already catalogued more than 6,400 exoplanets. Some orbit stars. Others, known as “rogue planets,” drift through space with nothing holding them in place, which fits the Greek root of “planet,” meaning “to wander.” Billions more are thought to exist. Over eight years of observations, TESS has found more than 1,000 exoplanet candidates that astronomers later confirmed.
The transit method also has well-known pitfalls. A binary star system, for example, can easily look like a planet crossing in front of a star.
“The biggest challenge with confirming planets is that we have to rule out false-positive scenarios,” said Kevin Hardegree-Ullman, a research scientist at the NASA Exoplanet Science Institute (NExScI). “This is difficult because: (1) there are just so many candidates… (2) we have limited telescopes… and we are generally competing with all other astronomers for time on these… and (3) some targets are not amenable to follow-up measurements with current telescopes (e.g., they are too faint to get a good signal other than a transit-like event).”
Rabtsevich isn’t overselling his result. He points out that correlation does not necessarily mean causation and that something other than a planet may turn out to cause the pattern.
What AI is good for here, and what it isn’t
AI does well at finding candidates. Confirming them is a separate job.
NASA deployed an AI model a little under a year ago to sort through TESS data, and it has already identified around 7,000 exoplanets. Tools like Claude Code and Codex are available to the public, so they could also let amateurs like Rabtsevich join the search from well outside the scientific establishment.
“AI tools won’t necessarily be helpful in the process of confirming planets,” Hardegree-Ullman said, “but [they] will likely continue to be used to string together publicly available code to help search for planets and run them through the basic vetting processes… ”
For Rabtsevich, the main benefit was speed.
“Without these tools I couldn’t have processed a data set this size and carried it through to a result,” he said. “Even a couple of years ago my job meant building endless spreadsheets and grinding through data sets, and that ate up a huge amount of time. Here, a search I planned went through 126,000 stars, and then dozens of tests ran on a single one, in about two weeks.”
The flood of amateur candidates he’s worried about
His post has already pulled other amateurs into planet hunting. Rabtsevich knows that is not entirely a good thing.
“Since my [X] post, I’ve seen a lot of people start searching TESS data for planets themselves,” he said. “I hope that enthusiasm comes with careful checks, so professional astronomers aren’t left sorting through poorly vetted signals.”
Hardegree-Ullman’s point about limited telescope time makes that warning more serious. A surge of poorly checked AI candidates would use up the resource professionals have least of.
Some readers are less cautious. In a Claude subreddit, one Redditor welcomed an AI story that wasn’t about slop: “There is an awful lot of absolute shite made with the use of AI,” the user wrote. “But this… this is fucking incredible. Well done.”
It is too early to celebrate. If TIC 4206066 dims on the schedule Rabtsevich published for October 31 to November 26, his method will have earned the praise. If it doesn’t, he has promised to say so publicly, and that promise is the most credible part of the story.