The #1 Mistake People Make Right Now
Everyone sees “My AI agents found an unmapped Maya site?” and assumes you need a custom detector, a Mesoamerican PhD, or a private lidar plane. Wrong direction.
Pietro Miotti’s own write-up (AI researcher, not archaeologist) shows the opposite stack: public NASA G-LiHT strips, plain-English vision agents, a literature pass, and a polite note to Mexico’s INAH. Idea in Peru → official reply in under 72 hours. The catch is ugly and useful: agents and human both treated the hilltop as unregistered while a 2025 journal paper had already mapped it.
Skip the hype. Here is a scaled beginner workflow – data URLs, prompts, and the failure modes that turned “discovery” into “rediscovery.”
Quick Context: What Just Dropped
Public post landed 10 November 2026. Per Miotti’s diary, 48 agents had gone through roughly 1,200 km² of NASA lidar-derived images, flagged a cliff-edge hilltop (levelled plaza, oval mound ~8-12 m, raised courtyard, possible aguada), helped draft Spanish text, and got an INAH reply on 21 Oct that nothing was registered at those coordinates. Next-day update: features matched Estrada-Belli et al. 2025. Method still works. Register lag is real.
Lidar is not a magic map. Pulses slip through canopy gaps; bare-earth returns show edges and corners trees hide from satellites. That is the only reason a general vision model can score “platform-like” geometry at all.
Hands-On: Replicate the AI Agents Maya Lidar Hunt
Do not spin 48 agents on night one. Four to eight vision-capable chats and a few free tiles beat a heroic fail. Flag shapes → cross-check literature → decide if anything deserves a quiet human report.
1. Grab the data
Start at the NASA G-LiHT portal (2013 AMIGA-Carb Yucatan campaign) or the reprocessed bare-earth products on OpenTopography (DOI 10.5069/G9TH8JX9). Pull Cloud-Optimized GeoTIFF DTM or hillshade tiles at 0.33-1 m. Keep each file under a few hundred MB so image uploads do not die.
2. Turn elevation into agent-readable images
QGIS or rasterio + matplotlib: multi-directional hillshade or openness. No GIS day? Dump a DTM screenshot into one chat: “Make a clear hillshade PNG from this elevation grid; exaggerate vertical scale ~1.5×; set up to show platforms and mounds clearly.” Save 8-20 overlapping tiles.
3. Spin the agents with a tight visual checklist
Exact system prompts and models for the original 48 agents were never published – CellCog and others flag that gap – so this is reverse-engineering from the public description only. Same instructions in every parallel chat:
You are a careful remote-sensing assistant, not an archaeologist.
Examine the attached lidar hillshade tile.
Look ONLY for: levelled plazas; rectangular/oval platforms or mounds >3 m high;
raised courtyards; straight edges/sharp corners; possible aguadas (flat-bottomed hollows).
Reject: natural hills, sinkholes (cenotes), modern roads, drainage, tree-line artifacts.
For each candidate: shape, estimated height if possible, position inside the tile,
confidence (low/med/high). JSON list only. If nothing qualifies, say so.
One or two tiles per agent. A second “reviewer” chat merges duplicates and ranks by confidence plus geometric regularity. Force pixel or relative coordinates. “Middle left” is worthless later.
Pro tip: free-tier image quotas and context windows choke on full flight strips. Tile first or the run dies before lunch.
4. Literature and register cross-check
This is the step that failed in the original loop. Fresh web-enabled agent: “Search Google Scholar, ScienceDirect, and public Maya settlement catalogs for published structures matching these features and region. Check INAH-style registers nearby. List DOIs.” Run the same ask on a second model. Open the PDFs yourself. The site sits in Estrada-Belli et al. 2025 (Journal of Archaeological Science: Reports) even though INAH had no register hit when Miotti asked – as of that October 2026 query window.
5. Optional point-cloud sanity check and draft letter
If you hold LAS/LAZ, ask for ground-vs-canopy returns on a slice through the candidate. Then: “Draft a short factual Spanish note in an INAH-style tone. Methods, image descriptions, confidential coordinates annex. Do not claim discovery – say possible unregistered features.” You remain the human who sends nothing, or something, after you read the paper trail.
Common Pitfalls That Kill the Run
- One shallow Scholar pass – agents pattern-match; they do not exhaust recent regional papers.
- Whole strips instead of tiles – vision context and free image limits fold.
- “No register hit” treated as “unknown to archaeology.”
- Posting exact coordinates or “new pyramid” before any authority review.
- Ignoring that lidar catalogs already list huge numbers of known-but-unexcavated mounds (community point on the HN thread).
Token burn is the quiet killer. Parallel images empty a free tier fast. Start absurdly small.
What Results Actually Look Like
Expect noise. Original overnight pass surfaced several village-like clusters; one hilltop won on geometry. Human + agent iteration wrote the Spanish letter; INAH answered in about a day; wall-clock under 72 hours. Your first dozen tiles: maybe 5-20 false positives and at least one “already published” hit. That is the filter working, not a broken run.
When NOT to Use This Approach
Restricted or non-public data: stop. Low-confidence blobs emailed to agencies: stop. Agent JSON is not a dig permit – Mexico (and most countries) protect sites by law; fieldwork needs formal authorization. Clout-only runs: skip. Method beats the headline.
Is the bottleneck now curiosity plus a free Earthdata login, or do domain experts still need a veto on every claim? The November 2026 episode leaves that open.
FAQ
Did the AI agents actually discover a new Maya site?
No. Rediscovery of features already in Estrada-Belli et al. 2025, missing from the INAH register at query time. Ground truth stays with professionals.
What free data and tools do I need this weekend?
G-LiHT or the OpenTopography Yucatan reprocessing (hillshades preferred), two vision LLMs with image upload as of November 2026, QGIS or one conversion prompt, Google Scholar. Enough for a few-tile proof. Paid tiers only if you want more parallel agents.
Why did the literature check fail even though the paper existed?
Retrieval gap under time pressure. Very recent or region-specific papers slip unless you force multi-model queries, DOI hunting, and manual figure checks. Miotti later cited the same 2025 paper in his own references – the agents (and the first human pass) simply did not surface it when it mattered. Second independent pass. Open the PDF. Match shapes yourself.
Pick one OpenTopography G-LiHT Yucatan tile, feed it to two vision agents with the prompt above, and compare JSON lists. That is the next concrete move.