Short answer: On October 18, Pietro Miotti was traveling in Peru and reading about Hiram Bingham, the explorer who reached Machu Picchu in 1911. He wondered what Bingham would have done with today's tools. That night he pointed a group of AI agents at public NASA lidar data and went to bed. By October 21 he had an official reply from the Mexican government about a Maya site hidden under the jungle.
Then, on November 11, he added an update to the top of his post: archaeologists had already mapped the place. The story still climbed to the top of Hacker News with more than 1,000 points. Both halves deserve attention: what the agents pulled off, and the one step where they failed.
What happened in those 72 hours?
Miotti is not an archaeologist, and he says so. He asked Gemini, ChatGPT, Claude and Grok how a modern Bingham would look for ruins, and all of them gave the same answer: lidar. From a satellite, a ruin in the jungle is just trees. Lidar fires laser pulses through the gaps in the canopy and maps the ground underneath.
Peru has little public lidar coverage, so he switched to the Maya region, which has much more. He asked the agents to analyze it overnight. In the morning the reports listed villages and other regular shapes that looked like structures. He asked which of them had not been discovered yet, and after a search of the literature and the archives the answer came back: one major village did not seem to be registered.
From an idea on October 18 to an official reply from Mexico's INAH on October 21, according to Pietro Miotti's diary of the project.
The agents also told him what to do next: write a report and send it to INAH, Mexico's National Institute of Anthropology and History. They supplied the reporting steps and the email addresses, and Claude suggested translating the report into Spanish, a language Miotti does not speak. INAH replied that no registered archaeological site was identified at the location in his report.
How did 48 agents search 1,200 km² of jungle?
With no special tooling. The lidar strips, about 1,200 km² of forest, were turned into images, and 48 agents went through them one by one, looking for pyramids, platforms and plazas while filtering out natural hills, sinkholes and roads. In Miotti's words: "There was no special model or classifier involved: the agents literally looked at the pictures, following a precise guide."
48 agents went through images of about 1,200 km² of lidar data one by one, following a written guide.
That is the part worth copying. A huge search became thousands of small, identical visual checks, each one run against the same written rules. No single check needs a genius. It needs patience, and patience is what a fleet of agents has.
There was a second pass, too. The agents went back to the raw point cloud to make sure the shapes were really on the ground and not created by the trees. Under 10 to 17 m of forest, the ground points traced the same plaza, pyramid and courtyard.
What did they find under the trees?
A hilltop on the edge of a cliff with straight edges and sharp corners. As Miotti puts it, "Nature doesn't really do straight lines."
Lidar pulses pass through gaps in the canopy and map the ground below. Illustration, not the real site.
Zoomed in, the lidar shows a levelled plaza, an oval pyramid at least 8 to 12 m tall, a raised courtyard attached to it and a flat-bottomed hollow that might have been a rainwater reservoir. Miotti marks the interpretation as speculation from a long chat with the agents: probably a small Maya civic centre from the Classic period, roughly AD 250 to 900. About the reservoir he adds: "Or it's just a hole."
One careful detail: the exact coordinates went to INAH in a confidential annex and are not in the post.
Why is it a rediscovery and not a discovery?
Because the site is already in the scientific literature. According to Miotti's update, it was mapped in Estrada-Belli et al. 2025, a study of settlements across the Maya lowlands. The agents missed that paper during their literature check. INAH's register did not have the site yet, so both statements are true at once: it was absent from the official register, and it was not new.
The site was already mapped in Estrada-Belli et al. 2025. The agents missed the paper in their literature check.
Here is the irony. The step that sounds impossible, spotting a pyramid under 17 m of forest somewhere in 1,200 km², went fine. The step that sounds routine, checking whether somebody got there first, is where the agents slipped. Anyone who has asked an agent "has this been done before?" should remember this case. When an agent reports that it found nothing, that is a statement about its search, not about the world.
The method itself is the one professionals use. Miotti points to Auld-Thomas et al. 2024, where archaeologists found a whole Maya city in Mexico in lidar that had been collected for environmental research. His own conclusion is generous to both sides: "an outsider with agents ended up in a similar place to the experts."
What did Hacker News argue about?
The thread passed 370 comments, and it was not all applause. Someone coined "vibe archeology" within hours. Skeptics said they would wait for an archaeologist to verify the claim, because agents are good at producing findings that sound groundbreaking and collapse on review. The November 11 update proved that instinct half right.
Others worried about the receiving end. Open-source maintainers already complain about floods of AI-written reports, and archaeologists could be next. The most serious point was about capacity: finding a site is now cheaper than protecting it. If agents surface hundreds of candidates, there will be more than archaeologists can excavate or guard, and looters do not wait for permits. That makes the confidential annex look like the smartest decision in the whole story.
Can you run a search like this yourself?
The recipe is simple enough to reuse far outside archaeology. Take a large public dataset that can be rendered as images. Write a precise guide: what counts as a hit and what the usual false positives look like. Fan the tiles out to many agents in parallel. Send every hit to a second pass against the raw data. Then treat the "is this new?" answer as a lead, and put it in front of a human expert before you call it a discovery.
The bill is mostly image input, so the model choice matters. A cheap vision model can do the first sweep over every tile, and a stronger one can recheck only the tiles that were flagged. On Tokely, vision models from different developers sit behind one OpenAI-compatible key, so a two-tier pass like this is a change of model name in the request. The current list is on the models page.
Miotti ends his post with a proposal: "we should probably lidar the whole planet and let the agents discover the rest." Maybe. For now the site is with INAH and with the archaeologists who mapped it first, which is exactly where a find like this belongs.
Originally published at tokely.me.
