An amateur found a planet. Another found ruins. What will you find?
This year the labs cracked problems that had been open for decades. What interests me more is what people outside them are doing with the same models.
At a glance
2026 has brought a run of AI results in science. Levent Alpöge, using the publicly available Claude Fable 5, announced a counterexample to the Jacobian conjecture (1939) in three or more dimensions. GPT-6 Astra improved an estimate of large gaps between primes that had stood for over 80 years. In early September, OpenAI published an internal model's solution to the Navier-Stokes problem. In September, Anthropic unveiled a biology lab where around 950 agents found a CRISPR-like enzyme system in 21 hours. Outside the labs, Dominik Peters and colleagues proved with Astra that the core always exists in committee elections. Gamow Labs used an AI agent to find the cause of an infant's death. Data analyst Paul Conyngham used ChatGPT and AlphaFold to design a vaccine against his dog's cancer. Pavel Rabtsevich found a planet candidate in TESS data, with follow-up observations from 31 October; it is still only a candidate. Nate found at least 26 mounds in NASA lidar data in Mexico, not yet assessed by archaeologists.
Pavel Rabtsevich is a 28-year-old product manager living in Spain. He never studied astronomy and he doesn’t own a telescope. He has an internet connection and subscriptions to a few AI tools.
In September he set Claude Code, Codex and Jev loose on public data from the TESS space telescope. More than 126,000 stars. In one of them, TIC 4206066, he found a regular dip in brightness every 3.18 days. Then he found the same dip in data from 2018 and 2020. Something is may orbiting that star. By his calculation, a planet about 1.4 times the size of Earth. Or a binary star that is only pretending to be a planet. We’ll know in November.
Because for now it’s a candidate. Rabtsevich used Claude Code to run 74 tests against his own hypothesis, then wrote to MIT and asked for new observations. They said yes. From 31 October, TESS will look at the star every two minutes, and Rabtsevich has published in advance the times when it should dim. “If the star dims on time in November, this gets a lot more serious,” he wrote. “If it doesn’t, I’ll post that too.”
A few days later, Nate read his post on Reddit. He started wondering where else public data might be sitting with something in it that nobody had written down. In a few hours, and for a few hundred dollars spent on Claude, he found a cluster of at least 26 earthen mounds and a raised platform in NASA laser data from the Gulf of Mexico coast. The maps from 1947, 1982 and 2015 show nothing there. Archaeologists have not assessed the find yet, so for now it is only a pattern in the data.
Read that again. Two people, no lab, one month. And tools you could pay for tonight.
Labs first
This summer was strange. News that used to come once a decade was coming once a week.
This summer, the mathematician Levent Alpöge announced that, with Claude Fable 5, he had found a counterexample to the Jacobian conjecture. Ott-Heinrich Keller formulated it in its multidimensional form in 1939, and for 87 years nobody managed to prove or disprove it. Alpöge’s counterexample fits in a single post on X, so other mathematicians could check it easily. It holds for three or more dimensions. The two-dimensional version is still open. Alpöge works at Anthropic, but by then Fable 5 had been publicly available for a few weeks.
Then came GPT-6 Astra from OpenAI. It scored 98% on the hardest tier of the FrontierMath benchmark. On top of that, two new results on gaps between primes. In one of them it improved a term in an estimate that nobody had shifted in over 80 years.
At the start of September, OpenAI published a solution to the Navier-Stokes equations, one of the seven Millennium Prize Problems for which the Clay Institute has put up a million dollars. The proof is formalised in Lean, so even a computer can check it. It was the work of an internal model, considerably stronger than Astra, and a group of agents numbering in the order of ten thousand. OpenAI doesn’t want to claim the prize.
At the end of September, Anthropic unveiled its own biology lab. Around 950 agents running Claude went through more than 200,000 enzymes in 21 hours, and one of them noticed something that, according to Anthropic, nobody had described before: an enzyme system with an array of repeating DNA segments that resembles CRISPR. In other words, the principle today’s gene-editing treatments are built on. Nobody knows yet what the new system does. Scientists are testing it in the lab.
Labs with billion-dollar budgets and their own models you can’t get anywhere near. Fine, that was to be expected.
And yet what interests me most is something else.
People with the same tools
The idea came from the machine
Dominik Peters works on voting theory at CNRS in Paris. With colleagues from Oxford and Munich, he took on a question the field had been chewing on since 2017: when you elect a committee, can you always find an outcome that no sufficiently large group of voters can object to? The key idea of the proof came from Astra, over a long piece of joint work. The answer is yes. Such a committee always exists.
Notice what didn’t happen. Astra didn’t replace him. Given just the bare problem, it wouldn’t have worked this out on its own. It needed people to steer it. But the idea the field had spent nine years looking for came from Astra.
The map nobody had time for
Brice Ménard teaches astrophysics at Johns Hopkins and also does research at Anthropic, so he isn’t exactly an outsider. His story belongs here anyway, because almost every field has work like this sitting in a drawer.
Whenever he covered ultraviolet light with his students, he had to apologise. The only UV map of the sky was full of holes. The GALEX satellite deliberately avoided the brightest regions so it wouldn’t burn out its detectors, and about a third of the sky had never been photographed in UV at all.
You can fill the holes in by calculation. But that means weeks of fiddly work, pixel by pixel, so nobody did it.
This summer Ménard did it with Claude in a few days. The result is the first complete UV map of the whole sky, and every pixel is labelled as either measured or calculated, with an estimate of the uncertainty.
How many “when I get the time” jobs do you have?
An answer for one family
This is the story that stopped me.
Three years ago, the Frost family lost their infant son. An unknown lung disease. His genome was sequenced and went through a clinical lab, and finally through one of the best experts on rare developmental lung disorders. No answer came.
So the parents didn’t know the most important thing. Was it chance, or is it hidden in their DNA, and could it happen again?
Researchers from Gamow Labs and Baylor College of Medicine ran an AI agent, built partly on the Gemini 4 Argon model, on the same data. The agent compared the raw genome with the scientific literature and found a chromosomal rearrangement that standard software had missed. An annotation error had marked it as common, so it was filtered out. The lab confirmed the finding with further sequencing.
“It didn’t undo our loss, but it gave my wife and I closure and the confidence to move forward in growing our family,” said the father, Randall Frost.
A vaccine for Rosie
Paul Conyngham is a data analyst from Australia. He never studied biology. Rosie, the dog he adopted from a shelter, had a mast cell tumour the size of a tennis ball on her leg. Neither surgery nor chemotherapy helped, and the vets gave her a few months.
Conyngham paid a few thousand Australian dollars out of his own pocket to have the tumour’s DNA sequenced. Then he used ChatGPT to search the data for mutations that could help the immune system recognise the tumour, and AlphaFold to model what the relevant proteins would look like. He took the finished design to the University of New South Wales. Its RNA institute turned it into an mRNA vaccine in under two months. The only one in the world, made just for Rosie.
“That’s the really special part of the story, is that part was done by [Paul], someone had no background in biology or medicine or chemistry,” said Pall Thordarson, whose team made the vaccine.
According to Conyngham, within a month Rosie was jumping the fence after a rabbit. Her largest tumours shrank significantly. She isn’t cured: she had the vaccine alongside immunotherapy, so nobody knows exactly what worked, and another dose is planned for the tumours that didn’t respond. One dog, no control group. But not long ago, something like this would have taken a pharmaceutical company’s programme.
A planet, some mounds and a small telescope
You’ve already met Rabtsevich and Nate. Nate, by the way, had separate AI “sceptics” attack every find before he sent it to INAH, Mexico’s heritage authority, and to the team that found the Maya city of Valeriana in 2024. Now he’s waiting to hear what they make of it.
Another amateur pointed a 30 mm Seestar smart telescope at the known exoplanet HD 189733 b. According to his Reddit post, with help from Claude Code he measured the star dimming by 2.50% as the planet passed in front of it. Large telescopes put it at around 2.45%.
And at Stanford they built Paper2Agent. Out of a hundred computational biology papers, they turned 74 into working AI agents that run the paper’s method on your data. It was published in Nature, and the code is public.
What they have in common
Felix Rieseberg works on Claude at Anthropic, and in October he wrote a guide to being more ambitious with AI. He splits ambition into four kinds. The first is speed: the same work in less time. He reckons anyone can manage that, so he skips it.
The other three:
- Altitude. Don’t ask AI for a task, ask for the outcome behind it. Rieseberg stopped asking Claude to fix a specific app crash. He asked it to keep the apps from crashing.
- Depth. Work the way the best person in the field would: with raw data, real methods and someone from outside who checks the result. Exactly what Nate did with his sceptics and Rabtsevich did with his 74 tests.
- Range. Bring in fields that wouldn’t normally go near the problem. Nate combined laser data, archaeology, old army maps and a database of oil wells.
None of these people had a secret tool. They asked about something that looked impossible, and then checked it honestly.
The Bannister effect
For decades, people chased the four-minute mile. In May 1954, Roger Bannister ran it. He had it to himself for just 46 days before someone else did it too.
Human physiology didn’t change in those 46 days. What people believed did.
Rabtsevich found a planet candidate. Nate read about it and within days he was looking for mounds.
Now you’re reading about it.
So what will you try?
Sources: Felix Rieseberg, How to be more ambitious; Gizmodo: Rabtsevich; Nate: I went looking for lost cities with Claude; ScienceDaily: Fable 5 and the Jacobian conjecture; OpenAI: GPT-6 Astra; OpenAI: Navier-Stokes; Anthropic: Claude discovers a novel enzyme system; arXiv: Existence of the Core in Approval-Based Committee Elections; Anthropic: The missing map of the sky; Gamow Labs: Frost family; The Scientist: a vaccine for Rosie; The Conversation: an oncologist on Rosie’s vaccine; Skycruncher: HD 189733 b; Nature: Paper2Agent.