48 agents found a Maya site in 72 hours. Archaeologists found it first
Less than 72 hours from a stray thought on a trip to an official reply from the Mexican government. That is the headline of Pietro Miotti's post (https://pietromiotti.com/latest/maya-site), the top Hacker News story on Wednesday at 883 points. Miotti, a Google researcher and openly not an archaeologist, had agents pull public NASA lidar over the Maya lowlands, turn about 1,200 kmΒ² of strips into images, and send 48 agents through them one by one looking for pyramids, platforms and plazas while filtering out hills, sinkholes and roads. No special model, no classifier. The agents looked at pictures and followed a guide.
One hilltop stood out: a levelled plaza, an oval pyramid at least 8 to 12 meters tall, a raised courtyard and a possible water reservoir, all under 10 to 17 meters of forest. The agents checked the raw point cloud to make sure the shapes were ground and not trees, drafted a report, translated it into Spanish, and found the right contacts at INAH, Mexico's national heritage institute. INAH replied on October 21: no registered site at that location.
Then the twist. Archaeologist Sarah Parcak pointed out on Bluesky that Tulane's Francisco Estrada-Belli had already published it, as The Verge reported (https://www.theverge.com/ai-artificial-intelligence/1031498/an-ai-agent-swarm-may-have-unearthed-an-unmapped-maya-site). Miotti updated the post the same day: the site is in Estrada-Belli et al. 2025, "the agents missed it while going through the literature, so they rediscovered it." It just was not in INAH's register yet.
Look at where the pipeline broke. The hard, physical part worked: 48 agents doing the patient survey work, point-cloud checks, a clean report. The step that failed was the cheap one, the literature search. "Has anyone done this?" is the question agents are worst at answering confidently, because a missing result and a missed result look identical from the inside. An official register that lags the journals made it worse, since the agents had a confident no from a real authority.
The lesson is not that outsiders should stay out. The survey speed is real and the agents landed exactly where the experts had. It is that novelty claims need their own verifier, and right now nobody ships one. Anyone running agents on discovery tasks, in science or in patents or in security, should treat "this is new" as the claim that needs the most checking, not the least.
← Back to all articles
One hilltop stood out: a levelled plaza, an oval pyramid at least 8 to 12 meters tall, a raised courtyard and a possible water reservoir, all under 10 to 17 meters of forest. The agents checked the raw point cloud to make sure the shapes were ground and not trees, drafted a report, translated it into Spanish, and found the right contacts at INAH, Mexico's national heritage institute. INAH replied on October 21: no registered site at that location.
Then the twist. Archaeologist Sarah Parcak pointed out on Bluesky that Tulane's Francisco Estrada-Belli had already published it, as The Verge reported (https://www.theverge.com/ai-artificial-intelligence/1031498/an-ai-agent-swarm-may-have-unearthed-an-unmapped-maya-site). Miotti updated the post the same day: the site is in Estrada-Belli et al. 2025, "the agents missed it while going through the literature, so they rediscovered it." It just was not in INAH's register yet.
Look at where the pipeline broke. The hard, physical part worked: 48 agents doing the patient survey work, point-cloud checks, a clean report. The step that failed was the cheap one, the literature search. "Has anyone done this?" is the question agents are worst at answering confidently, because a missing result and a missed result look identical from the inside. An official register that lags the journals made it worse, since the agents had a confident no from a real authority.
The lesson is not that outsiders should stay out. The survey speed is real and the agents landed exactly where the experts had. It is that novelty claims need their own verifier, and right now nobody ships one. Anyone running agents on discovery tasks, in science or in patents or in security, should treat "this is new" as the claim that needs the most checking, not the least.
Comments