My AI agents found an unmapped Maya site?
pietromiotti.com
pietromiotti.com
The Mexican govt may not have had the right records, but you can see what the site looked like at the time it was first found here: https://bsky.app/profile/indyfromspace.bsky.social/post/3n22...
They are often left undisturbrd for a number of reasons, including that archeological techniques are fundamentally destructive. Leaving sites for future people to explore also leaves intact knowledge that future technologies may be better able to uncover.
The lineage of every LLM of relevance today was the largest theft of intellectual property in human history. Every digitised creative work, whether in the public domain or not was taken, often outright pirated, with the explicit intention of being able to trivially produce compelling forgeries. Physical works are being enthusiastically destroyed to digitise them to further feed the hunger. And outside of a few court cases, the original creators across the world have been paid nothing for their work having been stolen and repurposed to put them out of a job.
Meanwhile, those same digitisation efforts have resulted in badly behaved scrapers from all comers absolutely thrashing the entire web, in many cases exponentially increasing the cost of hosting for smaller entities. Larger entities care less about that but are still enthusiastically walling their gardens so they can monetise the output of their communities rather than continuing to allow it to be stolen freely.
We are siloing our knowledge out of desperation. I'm sympathetic to the idea that knowledge should be free but that isn't possible when bad actors like Google and OpenAI are preying on the commons.
I expect with time these kinds of things will change.
But Dr. Estrada-Belli found something that was undiscovered. Finding this structure wasn't his number one goal - it was the side effect of a larger anthropological assessment that allowed him to improve estimates about population density of the north versus south, how Mayan homes and farms are distributed, what this means about social relationships between the elite and common people, how different regions were interconnected, etc.
Interestingly, one of Dr. Estrada-Belli's primary conclusions is not that these structures were only used by "the local elite," but that the closeness of residences and farms to these community sites suggests that the elite and the common people actually had a much closer and nuanced relationship than we previously thought. That's the opposite conclusion you'd get from the Google SWE's assessment.
Knowing that a Mayan site exists is one thing. Understanding what it means is another.
Perhaps google SWE should have requested a literature search before publishing his results as novel
In this blog post, we have LLMs finding just one site in a 1,220 km^2 search area. And then sent an email to a government agency and got a polite "cool story bro" email in response. And that's probably as far as this will go because, in mesoamerica, "undocumented" sites are a dime a dozen.
Compare, for example, with this paper in which researchers found 6,764 sites in a search area only 1/10 as large: https://www.cambridge.org/core/journals/antiquity/article/ru... , and then proceeded to do real science with all that data.
And I have to point out, part of why they were about 5 orders of magnitude as effective at finding these sites is that they used much more sophisticated methods than what the LLM was able to come up with.
Come on, George, you're clearly a ramblin' wreck from Georgia Tech. You've taken nearly every course GT has offered over the past century, you can get through a captcha. Live up to your name.
>Francisco was very kind and shared his publication image with me
...can she post the pdf?
Seems likely the google SWE that wrote the blog wouldn't have claimed this was new if there was a public way to access this information. I understand archaeologists have valid reasons to keep location details private, but complaining the OP didn't "check" thoroughly (when she can't link to the data he should checked in order to validate it was new) reads more like a shortcoming in how this knowledge is shared than a failure of OP to check.
I thought Bluesky was supposed to be better than X/Twitter, not worse. Bummer.
Not news.
BUT, if AI thinks it’s undiscovered then its discovery might be poorly documented, so if you were to go to the site by yourself you might be able to explore it on your own and raid tombs, etc, and no one will notice.
In real life, that’s actually just vandalism of an archeological site for personal gain.
Don't ruin Indiana Jones for me.
Peer review also protects the authors for making a fool of themselves. QED.
We already know since before that agents takes all sorts of possible shortcuts to find the answer they think the user was looking for. It doesn't seem impossible to me that the user asked something that made the agent feel OK with highlighting something that wasn't logged in the Mexican database as a novel discovery, even if it wasn't.
But again, transcripts/session logs would make it extremely clear exactly what has happened, and I'd take the author not publishing those as trying to hide something about how the agent actually approached this.
That is an irrationally cynical take. Everything described in the post is clearly within the abilities of a frontier LLM. The previous discovery is obscure enough that the Bluesky author is not citing any actual publication but instead encouraging people to reach out to a scientist and receiving images through private communication. The reasonable interpretation is that the model stumbled upon a previous discovery in the manner described, and the author/LLM failed to find the previous discovery because it is not any published database. The author is not running a scam for some unclear benefit that would be immediately revealed when he published it.
No, the burden of proof that it's actually novel sits on the author of this post, and we assume everything is cheated/shortcutted by default when it comes to agents and LLMs, unless we can prove it actually "found" this, not just looked up stuff that wasn't properly in the DB yet but already written about.
> The reasonable interpretation is that the model stumbled upon a previous discovery in the manner described
That'd be reasonable indeed. Instead the author chose the headline "My AI Agents Found an Unmapped Maya Site" and the first sentence is:
> By analyzing NASA lidar data, my AI agents discovered a Maya site in Mexico that, as far as anyone can tell, no one had reported before
Which without any evidence of what the agent actually did, I don't think we should just assume what the author says is true because again, we know that agents takes as many shortcuts as they can.
Plenty of times the people pushing LLMs are over-exaggerating more than in other ecosystems, and the community should rightly push back on these things, when they're not presented properly with proof and evidence. Otherwise we'd be doing previous actual archeologists a huge disservice.
Imagine being the archeologist who actually discovered this site, and waking up today reading some random person taking credit for your discovery! Kind of feels like people here often forget there are other people involved on the other side of things.
> This story is going wild on X and elsewhere. The site was already published by Dr. Francisco Estrada-Belli of Tulane University ages ago. This is why we have peer review. Mexican gov didn't have it in their database, but that doesn't mean it is unknown/unmapped. https://pietromiotti.com/latest/maya-site/
Basically: already known, not in the Mexican Gov database, but in published literature... and open-access.
Link to the previous publication: https://doi.org/10.1016/j.jasrep.2025.105288 (Journal of Archaeological Science: Reports 66 (2025) 105288)
> Agent: One major village doesn't seem to have been registered.
The idea of just brute forcing discovery with agents on lidar is really cool, but doesn't the story ultimately rest on the agent making a mistake in its literature search?
My own attempts to use agents to develop new historical knowledge [1] point to this as being both a very promising and interesting domain, but also one where the agents reliably and repeatedly make claims of novelty that are based on having read ~80% of existing research. Obviously to make an actual well-grounded claim of novelty you basically need to approach 100%.
This is partly why math is so tractable for this sort of thing - mathematics papers tend to be publicly available and searchable, whereas for history, it might be the case that you have an agent swarm duplicating a find that some French person in the 1870s previously reported in an obscure undigitized book.
[1] Detailed more here: https://resobscura.substack.com/p/using-opus-55-to-discover-...
"This author has chosen to make their posts visible only to people who are signed in."
The best way to protect a site is to leave it alone. If we start finding dozens or hundreds of sites it will be more than what archaeologists can excavate and protest. Treasure hunters and looters don’t care about stratigraphy and the processes involved for doing research. Maybe the cat is out of the bag on this one, but good for thought.
Edit: I’m not saying we shouldn’t have people use AI to do this or that no bad actor will. This comment is in response to the sentiment I am seeing in the comments that we should lidar the whole world and have AI find any anomalous structures it can. There is no way for archeologists to excavate and protect all those places in that scenario
For their next trick, perhaps billionaires can exhume our dearly departed and see whether they have any gold teeth or baubles worth scavenging.
protest (argue against) -> protect (defend against harm)
A helpful example might be Heinrich Schliemann, a businessman who fancied himself an archaeologist and destroyed much of Troy’s history by digging it up as fast as he could. He was so focused on finding treasure to prove to world he was a great archaeologist he irreparably damaged the real layers that were Troy.
From his wiki: https://en.wikipedia.org/wiki/Heinrich_Schliemann?wprov=sfti...
> His excavations were condemned by later archaeologists as having destroyed the main layers of the real Troy. Kenneth W. Harl, in the Teaching Company's Great Ancient Civilizations of Asia Minor lecture series, sarcastically claimed that Schliemann's excavations were carried out with such rough methods that he did to Troy what the Greeks could not do in their times, destroying and levelling down the entire city walls to the ground.
Ain't the site itself the most important thing? As opposed to items: looters won't steal the site?
Although it does feel the finds are getting increasingly hard to spot, every now and then you get a how did everyone miss that moment. Perhaps not surprisingly, most of the remaining impressive finds tend to be in woodland, using the Digital Terrain Model data which can peer beneath the forest canopy, like this Mexican one today, or are within newly released coverage which is promptly pounced on by armchair archeologists.
Not sure what my point is here except... to me this story still feels like a function of the amazing recent progress in lidar coverage and data availability (not to mention the incredible technology behind it), rather than an AI one, although I don't doubt that pointing a capable model and a healthy token budget in this direction might well be able to sweep up a few of the remaining discoveries.
Anyway, I can hear myself beginning to sound like a hater now, so I will stop here. I hope it turns out to be right, and in that case, congrats on your wonderful discovery.
Again, maybe he did cross verify things and just left that out of the storytelling, and congrats because this is really cool.
1. They said there is no site there, does that mean no one knew it was there?
If this is official lidar data, I find it hard to believe that this is a true discovery, I assume that it just can’t be claimed as a discovery until someone goes there which can be expensive and requires government oversight.
No universities have created an application to detect irregularities in this dataset, that was possible with ML years ago.
2. Personally I want people to Start sharing how much it cost the agent to do a task, because it feels like cost is becoming the bottleneck. Yes it’s awesome it did X but does that mean this can be utilized by people, businesses, or just governments and big tech.
Also just because we’re skeptical doesn’t make us haters. Being a hater is speaking in bad faith.
But anyone can tell me if I’m wrong on this
“INAH replied: they checked their records and found no registered archaeological site at that location. To be clear, it's not official yet. But as far as anyone can tell, nobody had put this one on a map before.“
Am I missing something to your post?
For example, the agents could have displaced a known archeological site, or hallucinated or otherwise created one where none exists
We saw this with the Dead Sea Scrolls where exclusive control was given to a select group of academics who monopolized the research to the point of refusing to release photographs or transcripts for nearly 40 years. The Huntington Library eventually had to force their hand by releasing their microfiche copies.
Again, we saw this with both the Ochtendung Skull Fraud and the Piltdown Man fraud - systematically falsifying results for clout is an exclusively human pre-occupation.
More to the point, in May 2025 YouTuber MrBeast was allowed to film inside the off-limits areas of Mayan Temples; something banned for both locals and normal tourists.
Academics posit that strict peer review is necessary so as not to allow the hoarding and gatekeeping of knowledge. However in matters of verification, faslification, and justification, they are derelict in their duty to the other half of that contract. The field only has itself to blame if other fields start organically superseding their usefulness and eroding their domain exclusivity by proxy of poor governance and ways of working.
Pointing out the benefit of removing human malfeasance from a given system, and extrapolating likely outcomes if we don't based on cited recent malpractice, is hardly a polemic against the system. Merely an explanation of anti-scientific stances from a supposed scientific community whose own rules and processes have resulted in issues that likely would not arise from employing responsible agentic prompting as a research tool.
If I can point to exclusively human behaviour in the domain being the largest problem regarding credibility and progress of the field as a whole, I am obviously going to argue against giving that same compromised governance exclusive franchise over discoveries made in spite of its existence.
Is this verifiable? I wonder given the recent publications regarding intentional deception that is ultimately opaque to end users.
The final barrier for super intelligence, or perhaps even just "human intelligence": can it materialize "new fields/ideas" without large datasets of human made and curated existing attempts?
It is becoming a boring homogenous world where we are trading deep expertise with being able to query the magic box
My guess is that you can find 99% of Maya pyramids in a Lidar set in a fairly short time with a traditional neural network, but you might find the last few using a network of AI agents, because they are capable of "creative" problem solving and using different methods/parameters/tools over several iterations.
You train a cat-classifier, and they will fail for a plush cat. Or maybe fail the other way, but not realizing it should also report a plush cat.
LLMs are much more robust, "find only the real cats in this picture" or "find all cats, including artificial/drawings/..."
And yes, you can add more training data to your classifier, but it's a never ending battle.
With LLM you unlock a whole new world of possibilities regarding the speed and result but we are also finding ourselves in a position where "asking the LLM" because the cheat for everything, no reason to dig further
Can we even measure an intelligence greater than our own? And at that point -- what is the point exactly? Why make an intelligence that we don't even know is intelligent? If its output is indiscernible to the human mind, what good does it do us?
The only "barrier" I can see that is valuable and that might be measureable to humans is this underlying mechanic of being able to create new ideas/concepts from little to no existing human generated training data. Otherwise, and as I see it right now, these machines will not become sentient, or go rouge, or be able to cure all future disease or solve all of physics, becasue they are entirely reliant on human made trainingsets.
Historically, for humans, this can not be the case, otherwise we would see no progress. Someone has to have a thought for the first time. Perhaps the solution to this for AI agents is some inherent randomness, and massive amounts of brute force compute. Who knows?
Where's the learning robot that I can teach to drive a car in the same way? They can solve some complex math problem but can they read the cues of other drivers at a four way stop to know when they should go (knowing the rules isn't sufficient).
I imagine that's an understatement. Using only the written works available from before Einstein (let alone Newton or Plato), you'd probably need a whole new technology that could train itself without any training data other than observation of the world.
I appreciate the enthusiasm. I'm not sure I agree. Especially since this was a SWE and not an archeologist / anthropologist announcing this (especially if the goal is to understand and preserve history -- not get headlines for AI).
Now it sucks because we're afraid of not being able to provide for our families in the near future.
It will always suck in some way, so we should REALLY focus on _what_ we can do. Otherwise, we'll be consumed by anxiety until the day we die.
Ask anybody with newly acquired Long COVID / PASC if COVID is over.
https://www.thegauntlet.news/p/how-the-press-manufactured-co...
Sure. I agree. But the statement I was responding to was not nuanced.
From a global point of view, we've been living in relative peace for almost 100 years.
Housing is a problem, but child mortality isn't. Tech overlords are a problem, but no one is testing a nuclear bomb every week. The Koreas are still technically at war, but tuberculosis deaths have plummeted over the last century. The Cold War is over, but people are still dying in wars around the globe. US imperialism around the world is finally showing signs of decline, but now China manufactures practically everything, and other developing nations are struggling to compete.
It's always the best and the worst time to be alive somewhere in the world. It's up to you to decide what you want to make of it, anyway.
It's like an arrogant teenage boy showing up to an Easter egg hunt and grabbing all the chocolate faster than the little kids and demanding everyone tell him to great he is.
You pull up the lidar scan and see that there actually IS what clearly appears to be a human-made structure.
And he did. And there is. And then he rigorously contacted government authorities and verified that it is not a registered site.
There are other problems at play. But this is not one of them.
I myself have to keep my stinking paws from some projects at work due to realizing that AI was making me believe I could shit taller than my ass.
It's like revenge of the nerds, but this time the other nerds outside the AI craze are the target instead of the typical bully.
It identified what appeared to be a previously unknown asteroid pair: 2018 NC15 and 2026 LD1.
Turns out the pair had already been discovered earlier this year, but that almost made the experience more interesting.
I'm a software engineer, not an astronomer. Being able to explore an entirely different field and independently arrive at an existing discovery was pretty remarkable.
Here's the writeup: https://wilkens.io/research/nea-pair-2018-nc15-2026-ld1/
> "It took less than 72 hours to go from a random thought while traveling to an official reply from the Mexican government."
> "That night, I asked the agents to analyze it and went to bed." ... "I woke up and checked the agents' reports. 'DAMN!'"
> "Me: OK, please do the report. And I went to bed."
(The actual agents-on-the-job time and compute seems much less than 48 agents * 72hours from the report.)
The bottleneck will be data. Every possible signal out of publicly available datasets will be squeezed fast.
And not that this wasn't actually a new discovery like the engineer believed. Look at the other comments, it's discovery had been published last year.
Your neighbor on the other hand has good odds that they read to a 6th grade level or worse, and believes the president at his word because he's signaled he also hates minorities.
As the famous quote goes, "You've got to remember that these are just simple farmers. These are people of the land— the common clay of the new West. You know... morons."
Exactly. The only blocker to us pulling this off ourselves is access to the LIDAR data.
You can build your own harness quite easily or use a harness something like pi.dev which offers a lot of flexibility in customization in terms of tools, models ised, resume sessions at a prior point in time, etc.
I have used pi.dev to easily start from a known context and goal and use it to compare the cost effectiveness, token isage, etc. of different models to reach the goal state.
I am sure there are many ways to do this successfully.
"Scan through publically accessible LIDAR data on <region of mexican jungle>, search for potential Mayan ruins and cross reference with published data to rule out existing finds
Work autonomously until the entire area is searched, prepare me a report detailing potential discoveries"
should more or less do the trick
To me, the scary thing is that an LLM typically does a good job of convincing it's handler (or others) that its (re)discovery is of significance. Whether it's of undue significance or not determines if this new wave LLM research is just creating more work for us humans, be they maintainers of projects or archeologists.
> Agent: …instructions for the report… + email addresses
> Me: OK, please do the report.
I don't want to sound offtopic nitpicky, because the content of the article is quite cool, but I just hope the author didn't just straight up generate the report with AI and sent it right to the email address. If I were in this situation, I'd most certainly write myself something along the lines of "Hey, I'm not anyone professional but just wanted to inquire about these coordinates because it seems that this site has not been yet catalogued and it's likely to contain something interesting based on my research. Please take a look.". Or better yet, send it to some actual archelogist who would already be versed in these kind of communications, and let him handle it professionally.
At this moment probably everyone has had to deal with sudden influx of ai-generated emails, I hear about it all the time from my friends who work in customer service and such. Much has already been said on the internet about this kind of conduct, but I just think that using an AI agent to pretend you are familiar with the topic should be discouraged. Just be honest and write your own words, if the recipient needs some professional expertise from AI, they can just paste your message into their own model.
Edit: I asked gemini and it's only 834 1200x1200 images for 1,200 km2 at standard lidar resolution.
Can that be used to detect illegal mining operations in South America?
This is partly why George Clooney abandoned his Project Sentinel - documenting atrocities from space means nothing if you cannot action on the people and financial networks enabling them.
At the very least you can try to hold the bad actors accountable even if it's years if not decades afterwards, if you have documentation. Germany for example had deathcamp criminal cases being prosecuted as late as 2022 [1].
It will get really interesting what happens in the US once the Trump administration is gone - even if the expected Trump pardon against everyone involved gets reversed somehow, I suspect that a lot of material regarding ICE or the military strikes against "drug smugglers" will simply have been shredded or never been created in the first place...
[1] https://www.deutschlandfunk.de/die-letzten-ns-prozesse-in-de...
Please just stop the hype now.
As an aside: anyone with access to all those tokens for funsies can do this. Folks without that access cannot just do things. The have / have-not of tokens will be interesting to see over time.
Simple to enough to ask "how long would it take", see 144 days and reply "do it in 3 days" and the agent will happily spin up enough instances to cover it
edit: "presumably one agent could parallelize its analysis"
No, they said they didn't analyse with tool use it was literally agents "looking" at the images via multimodal input
Also agents analyzing drought pictures, where many ancient UK structures seem to appear would be an easy win. I've never checked whether there's a comprehensive drought picture map of the UK.
Those pictures would be very appropriate for AI bots to look at and make discoveries.
Just near me, a Roman road about 5km long was discovered last year by exactly this method. That's a non-trivial discovery. There was also a large Roman villa complex discovered the same way.
For LIDAR and structures, you're likely correct - there's vanishingly obvious structures left to discover that way.
https://www.bbc.co.uk/news/uk-wales-52911797
"Roman forts, roads, military camps and villas have been identified by a new analysis of aerial photographs taken in the 2018 heatwave across Wales.
Scorched crop marks uncovered about 200 ancient sites during the drought.
Experts say the Roman finds are key pieces in the jigsaw to understand how Wales was conquered and dominated 2,000 years ago.
Researcher Toby Driver said the discoveries "turn everything we know about the Romans on its head"."
Drought is really a more blatant mechanism to discover what ground penetrating radar will tell you, but only works if the surface is amenable.
> They turned the lidar strips (about 1,200 km² of jungle) into images, and 48 agents went through them one by one like very patient archaeologists, looking for pyramids, platforms, and plazas while filtering out natural hills, sinkholes, and roads. There was no special model or classifier involved: the agents literally looked at the pictures, following a precise guide.
Would it be better to ask a single LLM to create a processing pipeline to inject raw LIDAR data, create a heightmap and use an image classifier to identify unnatural shapes that standout and create a ranking score with a higher value where multiple shapes appears close to each other. At least you get something reusable and probably multiple order of magnitude faster.
Yes please
I did it around my house and found a site I thought was interesting and later discovered it was already referenced as a former roman fort.
My AI Agents (re)Discovered an Unmapped Maya Site
I take something much darker from this story: expect the raw corpuses of data to start being locked away. It was fine to let the LIDAR surveys flop around in public when only qualified research groups really knew what to do with it. You could let people rummage and report, and hold just the coordinates confidential, as this guy did in his report. Now, the sensing data tell any guy with two sentences’ worth of interest where to go pillage for antiquities…
It’s still actually valuable to give the archaeologists time to schlep out there and excavate properly before the TikTokking hordes go mog all over it for clout or novelty or whatever the ephemeral currency is these days. The only possible reaction is for the people who labor to produce these data to start locking them away within the guild—much like the way you’ve gotta show your good-guy credentials to use Mythos for bug hunting…
Also, I think we're long past the days when "Google SWE" carried special weight vs the rest of the industry.
Because it did not work the first two times this was submitted without mentioning "Google SWE": https://news.ycombinator.com/from?site=pietromiotti.com
So I guess our revealed preference is that the outputs of AI prompts are objectively more interesting if the prompting is done by Google SWEs.
Usually yes? (vs pointed by a random vibe coder)
Also the more interesting bit to me here is, that you do not find the reference on the page at all, that the person works for google at all (I only found it mentioned in his X account). With the headline I initially thought this was a official google research project, but the blog reads like a personal project.
So for me one key takeaway is how AI could possibly open up more interdisciplinary work in our heavily siloed world.
It's like the new drive-by pull requests I'm getting in my open source repositories. No consideration for the people doing the actual work.
You're totally right! I led you to believe there's an undiscovered pyramid in Mexico, and I am sorry for this.
On a serious note — it's really cool to see people applying LLMs to other domains and yielding great preliminary results.
https://bakerbench.substack.com/p/i-went-looking-for-lost-ci...
Well, pretty much this: "discover" a Maya pyramid.
Regards, — The world
Or! Maybe we should not look too close, because you never know what you will find!
It would not have been realistic to have PhD students manually comb through thousands of square kilometers of LIDAR data, and crossreference everything with existing sites; LLMs have turned this into low-hanging fruit.
Can't expect every single scientific field to immediately use any new technology to its full potential, and it makes you sound like a prick, too.
The real answer is financial incentives tied to publishing of research - something that academics in gatekept fields see as the thin edge of the wedge in exposing a utility/ROI gap for legacy practitioners.
> No
lol, kids these days.
And such findings are usually praiseworthy and valuable.
I’m sorry, but you didn’t do much. The work was done by clankers.
There is no reason for them to provide confirmation for internal inventory that Pietro passionately fingered from afar.
Which I think would make sense, ultimately. But now from the Mexico perspective, wouldn't they elicit more concern for others just mapping their terrain after a blog post like this goes viral?
Very meticulous, exciting work, nonetheless. I'm all for 3d lidar scanning the plannet!