On October 18, 2026, while traveling in Peru, Google researcher Pietro Miotti had an idea: If Hiram Bingham, the explorer who sought Machu Picchu, had today’s AI tools, what would he do? Three days later, 48 AI agents scanned roughly 1,200 square kilometers of NASA LiDAR data and sent a Spanish-language report to Mexico’s National Institute of Anthropology and History (INAH), flagging a suspected Maya site that had not previously been in the official registry. From idea to a response from a government agency, the whole process took no more than 72 hours.
The story then spread through tech media. In a blog update, Miotti said the site had in fact been documented in a 2025 archaeology paper by Estrada-Belli et al., but the AI agents had missed it while reviewing the academic literature. The site was not registered in INAH’s official archives, but the archaeological community already knew it existed.
The Methodology Is the Real News
The technical significance of this episode lies not in the word “discovery” but in the replicable way of working it demonstrated. Miotti did not use any specially trained archaeological model or custom classifier. His approach was to convert raw NASA LiDAR point-cloud data into images, then have 48 AI agents browse them image by image like a team of “extremely patient archaeologists,” using a detailed guide to pick out pyramids, platforms, and plazas while ruling out natural hills, karst caves, and roads.
During the conceptual stage, Miotti consulted Gemini, ChatGPT, Claude, and Grok at the same time, and all four systems pointed to the same path: LiDAR. LiDAR works by firing laser pulses at the ground, penetrating dense forest canopy to map the terrain directly, revealing the outlines of sites buried under vegetation through differences in elevation—something satellite optical imagery cannot do at all.
The suspected site features ultimately flagged included an elliptical pyramid estimated at 8–12 meters or more in height, a leveled plaza, a raised courtyard, and a possible reservoir. In his blog, Miotti stated clearly that these judgments came from “long conversations” with the agents and were inferences rather than field confirmation, and that the coordinates were submitted separately to INAH as a confidential attachment.
What the AI Missed
The agents performed well at image recognition but missed something in literature search. The 2025 paper by Estrada-Belli et al. already contained a record of the site, but the 48 agents missed it while filtering the academic literature, leading Miotti to believe for a time that he had made a “brand-new discovery.”
Spatial pattern recognition and literature synthesis are two fundamentally different cognitive tasks. The former involves looking for specific geometric features in structured visual data, where the signal-to-noise ratio is high, agents can divide the work in parallel, and errors can be caught through cross-comparison. The latter requires judging relevance in unstructured text, handling different naming conventions, and understanding the geographic attribution of sites—tasks that pose a greater challenge to language models and make errors harder to detect.
The site was not included in INAH’s archives, indicating a time lag between discoveries in academia and official registration, and that the circulation of archaeological knowledge is not always complete or timely. In this sense, the AI system exposed not only its own limitations but also the gaps in the existing knowledge-management system.
Historical Precedent: LiDAR Is Itself a Revolution
In recent years, LiDAR has repeatedly redrawn the boundaries of knowledge in archaeology. In 2018, an airborne LiDAR scan of a bounded region in northern Guatemala revealed more than 60,000 previously unknown Maya structures at once, thoroughly overturning outside estimates of the scale of Maya civilization. The scan covered about 2,100 square kilometers, compressing work that would have taken decades of manual surveying into a few flights. In May 2025, drone LiDAR discovered the Los Abuelos ceremonial center in Guatemala’s Petén region, covering about 16 square kilometers, with pyramids reaching 33 meters at their tallest.
But those discoveries all depended on systematic deployment by professional archaeological teams, requiring permits, funding, and time. What set Miotti’s experiment apart was that it was a personal project by one person, with a few API calls, completed within 72 hours, at a total cost that was almost negligible. The barrier to scaling has dropped sharply, meaning this kind of work no longer needs to wait for institutional approval.
Signals from “72 Hours”
At the end of his blog, Miotti wrote: “I think we are on the edge of a revolution in all scientific fields, not just programming, physics, and mathematics. We should now LiDAR-scan the entire Earth and let agents help us find everything else that remains.”
NASA and space agencies in several countries have already accumulated LiDAR datasets covering millions of square kilometers, much of which has never been systematically analyzed. Under traditional methods, the manpower and time needed to process this data are nearly impossible to mobilize. Miotti’s experiment shows that image-level analysis using parallel agents is entirely feasible: 1,200 square kilometers, 48 agents, less than three days.
Feasible does not mean reliable. The experiment’s “discovery” was ultimately downgraded to a “rediscovery,” and the core lesson is that the efficiency of AI agents in processing visual-spatial data has not yet extended to literature cross-comparison of comparable quality. With field archaeological confirmation not yet completed, any flagged result is only a candidate, not a conclusion.
The Capability Limits of AI Agents
The real value of this experiment lies not in the catchy headline “AI discovers a Maya pyramid,” but in how precisely it maps the current capability limits of AI agents: in well-structured perception tasks, agents can perform work that would otherwise require substantial specialized manpower at extremely low cost; in tasks requiring cross-domain reasoning, literature linkage, and knowledge deduplication, agents still make costly errors.
Missing a 2025 paper was just a minor episode. The real questions are: as the area covered by agents expands from 1,200 square kilometers to 1 million square kilometers, how will the error rate change, how will omissions and false positives be detected, and who will be responsible for oversight? Miotti’s experiment is a highly instructive prototype—it tells us both what AI can do and what needs to be seriously addressed next.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接