48 AI Agents Scan NASA Lidar, Rediscover Maya Pyramid in 72 Hours

A software engineer turned 48 AI agents loose on public lidar data and found a Maya civic center in three days.

Tools · Source: Hacker News

What happened

A software engineer was traveling in Peru. He thought about Hiram Bingham, the explorer who found Machu Picchu in 1911. He wondered what early explorers could do with modern AI. He spun up 48 AI agents using Gemini, Claude, ChatGPT, and Grok. He pointed them at 1,200 square kilometers of public NASA lidar data covering the Maya region in Mexico. Lidar fires lasers through the jungle canopy to map the bare ground underneath. The agents did not use a custom classifier or special model. They literally looked at images of the lidar data one by one like patient archaeologists.

They followed a precise guide to filter out natural hills, sinkholes, and roads. Overnight, they flagged a hilltop hidden under ten to seventeen meters of forest. It sat on the edge of a cliff with straight edges and sharp corners. The site featured a leveled plaza, a raised courtyard, and an oval pyramid at least eight to twelve meters tall. The agents speculated it was a Maya civic center from the Classic period.

The agents drafted a report in Spanish and sent it to Mexico's National Institute of Anthropology and History. The government replied that the site was not in their official register. A later update revealed the agents missed a 2025 paper by Estrada-Belli that had already mapped it. The agents failed their literature review. Still, they successfully rediscovered a lost city in under 72 hours.

Key facts

Why it matters

This proves agentic AI is ready for brute force visual analysis at scale. You do not need to train a custom computer vision model anymore. You can just prompt off the shelf multimodal models to look at thousands of images and follow a precise guide. The barrier to entry for analyzing massive visual datasets just dropped to zero. Anyone with an API key can now process satellite imagery, medical scans, or industrial sensor data using natural language instructions.

The second order effect is the death of gatekeeping in specialized scientific fields. An outsider with zero archaeology background did what used to require grants, permits, and years of physical fieldwork. When compute is cheap and agents are tireless, naive outsiders will start solving old problems across all scientific disciplines. The only bottlenecks left are curiosity, imagination, and compute. The experts who rely on manual data processing are about to be outpaced by solo developers.

For builders

Off the shelf models beat custom classifiers

Stop wasting time training custom computer vision models for basic pattern recognition. Multimodal agents can just look at images and follow instructions. Builders who leverage raw agent vision will ship faster and cheaper than those stuck labeling training data.

Public data is an untapped goldmine

Governments sit on petabytes of unstructured data like NASA lidar scans. The data is free but unanalyzed because human labor is expensive. Founders who point AI agents at these public datasets can extract massive proprietary value while legacy institutions sleep.

Agent hallucinations require human oversight

The agents failed to find the 2025 paper proving the site was already discovered. They hallucinated a completely novel discovery because their literature review was flawed. Startups building autonomous research tools must build robust verification loops or they will lose credibility.

My take

I love this story because it proves my core thesis. You do not need permission to build and you do not need a PhD to do science. You just need curiosity and cheap compute. The legacy experts will laugh until your agents do their entire life's work over a long weekend.

Original reporting: Hacker News. This is my rewrite and opinion.

More AI news for builders