Meta is on track to become the largest digital advertising company on Earth this year, overtaking Google for the first time in the history of the history of the industry. It posted 56.3 billion in revenue in Q1 2026 alone, which is up 33% yearonear, which in Silicon Valley terms apparently means congratulations, you may now behave like a municipal hazard. It has 4 billion users across Facebook, Instagram, and WhatsApp. That means Meta controls roughly half the planet's remaining attention span. Three billion chaotic family group chats and every single aunt currently sharing a minion meme about wine:00. It is currently sitting on what is arguably the most valuable data set in the history of commerce. Decades of behavioral signals showing what people look at, what they click, what they want, what they buy. And in the past few months, this company has contaminated a city's water supply with industrial bacteria, secretly paid hundreds of contractors to pose as teenagers and attack its competitors AI models, being cut off from Google's AI because it couldn't build its own, and continue to fire thousands of employees, while its CTO admits morale is the worst in 20 years. What's happening at Meta right now has gone beyond strategic incoherence, which I covered at length in an earlier video. This is a company actively sabotaging the ecosystem around it while failing to tend its own garden. Look, as a computer scientist, I want them to build cool things. Of course, LLMs are solving structural biology and rewriting mathematics. AI is a generation miracle. But Meta is treating this paradigm shifting technology like a teenager who just discovered Grand Theft Auto and a credit card, which incidentally was also me back in May this year when I fell in love with Trevor Phillips, and now I'm desperately waiting for November 19th to see if we're going to be reunited. In this video, I'm going to walk through how Meta poisoned a Wyoming city's water system, the covert operation that had contractors pretending to be children to break competitors chatbas. the humiliating moment Google cut Meta off from AI it couldn't build itself and what all of this means for an industry whose own researchers are now openly terrified. Also, Aaron Brockovich, if you're watching, and I genuinely hope you are, you might want to upgrade to a CRM system because the number of communities dealing with tech industry contamination is getting a bit beyond what anyone can memorize. In early 2026, Meta's data center in Cheyenne, Wyoming, was undergoing routine maintenance on its closed loop cooling system. These systems are the circulatory system of any data center. They basically pump treated water through the facility to absorb the enormous heat generated by thousands of servers running at full capacity, then release that heat through cooling towers and recirculate the water. To keep the system clean, the water is treated with biocytes and corrosion inhibitors, which are just fancy words for industrial chemicals designed to prevent bacterial growth and pipe degradation. Periodically, the system needs to be flushed so the old water now laden with these chemicals and whatever biological activity has survived them is discharged into the municipal sewer system so that fresh water can replace it. It is a beautifully simple circular process until a multi-trillion dollar tech giant decides that property disposal protocol is a luxury they simply cannot afford. The contractor handling this process was a company called Goat Systems LLC. I don't know about you, but if I'm looking for someone to manage toxic industrial discharge, I want a name that sounds intensely clinical and slightly terrifying. The moment your biohazard contractor sounds like a landscaping side hustle your uncle started in his garage, you have to ask some questions. For example, whether they also offer birthday petting zoos. What goat systems discharge into Cheyenne sewer system contain cupraidus gillardi which is a bacterium that is resistant to heavy metals and the very biocytes used to keep cooling systems clean. So even the bacteria have better resilience engineering than the project managers. It's the kind of organism that doesn't just survive industrial environments, but also vacations in them. And when this contaminated water reached Cheyenne's water reclamation plants, it knocked two facilities offline. These are not drinking water plants, by the way, but the city uses reclaimed water to irrigate parks, sports field, and golf courses. And the contamination created a genuine aerosol exposure risk for anyone using those public spaces. Aerosol exposure means that the bacteria could become airborne through sprinklers or water misting, potentially affecting anyone nearby and similar fun activities. The plants were offline for months while the system was cleaned. Meta, of course, has plenty of experience keeping things offline, usually their own platforms, during a bad configuration update. But now they appear to have extended their signature downtime to a city's public utility infrastructure, SLE. The Cheyenne Board of Public Utilities responded by revoking Meta's discharge privileges and suspending wastewater connections for all data centers in the area, not just Metas. One city counselor called it a very, very unpleasant surprise. That is a heroic sentence. If my city's reclaimed water system got knocked offline by industrial bacteria, I think I would personally reach for slightly more colorful vocabulary. But that's just me. And as we already know, I prefer to express my displeasure on canvas. Now, unlike what follows in the later chapters in this video, I don't think this was malice. Nobody at Meta decided to poison Cheyenne's water on purpose. It was probably negligence. Nobody tested the discharge properly before releasing it into municipal systems that serves the public. And honestly, that might actually be a little worse than malice because malice at least implies somebody was paying attention. This was a trillion dollar company outsourcing biohazardous waste management to a contractor called Goat Systems and then apparently now checking whether the goats knew what they were doing. We were promised that AI would optimize global supply chains and cure diseases and instead we got a heavy metal resistance superbug shutting down a golf course irrigation system because nobody thought to check the flush settings on the server farm. Truly, the future is now, guys. Filming at 30° C should be in the Olympics. I mean, I would lose, but at least the sport would be acknowledged. And you know, it's times like these when I recall people in the comments asking me if I'm AI, and I'm like, I wish. I really do. Anyway, this isn't happening in vacuum. Across the United States, 64 billion worth of data center projects have been blocked or delayed by local opposition. In Oregon, voters recall two port commissioners who supported a proposed data center. In Virginia, all four council members who voted to approve an Amazon data center were voted out. The backlash against data centers has gone physical, local, and democratic. And Cheyenne just became the most vivid example of why this is what happens when the cloud stops being a metaphor and starts asking your town council for electricity, water, and a personality waiver. But while Meta wasn't paying attention to what was flowing out of its cooling systems, it was paying very close attention to what was flowing through its competitor's chatbots. Where there is a well, there is a way, I guess. A poison water system is apparently a regrettable operational detail. Falling behind an AI, however, now that is an emergency. And if Meta's approach to handling industrial waste makes you wonder how carefully they're handling your data, well, that is a reasonable question. AI assistants are becoming less like search engines and more like thinking partners, which makes privacy matter differently. You're not just looking things up. You may be handing over drafts, doubts, research notes, unfinished thoughts, and ideas you're still working through. And most people have had that one second of hesitation when typing something personal into an AI assistant. Should I really be sharing this? That is exactly the problem Luma was built to solve. Lumo is Proton's own AI assistant built by the same team behind Proton Mail and ProtonVPN. Proton started back in 2014 when scientists who met at CERN set out to build an internet where privacy is not just an extra feature but the actual default. Lumo is built for people who want the usefulness of AI without the usual data trade-offs. Your chats are protected with zero access encryption, meaning not even Proton can read your messages. There are no conversation logs and most importantly, your conversations and thoughts are not used to train AI models. Lumo is also open-source and runs on proton control servers in Europe, which matters if you care about reducing reliance on big tech and keeping your conversations private. Recently, I was using Lumo to work through an idea while it was still a bit messy. And with the new Lumo 2.0 update, it feels a lot more capable. So now you don't have to choose between usefulness and privacy. You can try Luma for free or check out the paid plans with a discount at proton.me/houseofll. The link is down in the description. But while some companies are building AI with privacy as the starting point, Meta was taking a rather different approach to other people's data.

You know, Meta launched a covert operation internally codenamed can like the film festival because apparently even corporate espionage needs a touch of glam through a third-party contractor called Coalen. Hundreds of contractors were instructed to create fake accounts with birth dates listing them as under 18 using disposable email addresses and shared passwords. They then use these accounts to bombard the AI chatbots of Meta's competitors. We're talking OpenAI's chatbt, Google's Gemini, and Character.ai with thousands of prompts designed to push the models past their safety guard rails. The prompts targeted the most sensitive content categories imaginable, the kinds of topics that safety systems are specifically designed to refuse. I'm not going to detail the specific prompts because they involve content about vulnerable people, and I don't think repeating them serves anyone here. What I will say is that in a single round of testing, contractors generated over 45,000 prompts and meticulously logged every chatbot response in spreadsheets. None of the targeted companies knew this was happening, by the way. The project was still active as of April 21st this year. Now, red teaming competitor products is a thing that safety teams do. Testing how rival chatbots handle sensitive content is not inherently a scandalous thing. Meta's official position is that this was a responsible industry standard practice meant to help ensure safe and age appropriate experience. But that defense falls apart under about 30 seconds of scrutiny. Industry standard safety research gets shared with the company whose product was tested. That's the whole point. You find a vulnerability, you disclose it responsibly so it can be fixed. That's how legitimate security research works across the entire technology industry. What you don't do is run the operation under a film festival code name through a third-party contractor using fake accounts disguised as children while keeping the targets completely in the dark. It's a magnificent corporate strategy really. You have thousands of fully grown adults with advanced engineering degrees, by the way, sitting in a modern office complex hunched over keyboards trying to figure out how a fictional 14-year-old would phrase a question about homemade explosives. It's not data science. Is 21 Jump Street rewritten by a compliance department. And the hypocrisy is what makes this generally difficult to stomach. Meta's own internal assessment surfaced through legal proceedings, by the way, reportedly showed a 66.8% 8% failure rate in blocking child safety content and a 54.8% failure rate on self harm prompts across its own chatbot products. Twothirds of child safety threats getting through. More than half of self harm content passing unchecked. Their own house was on fire, visibly, documentably on fire. And instead of putting it out, they were inspecting the neighbors smoke detectors while dressed as children. If your own platform has a 66% failure rate at protecting minor, maybe don't spend Q1 trying to gaslight SHA GPT into telling you how to smoke oregano, fix your own house, babe. I can't even call it corporate espionage anymore. This is closer to middle school cyber bullying with a trillion dollar market cap. Ramen Chowry, CEO of Humane Intelligence, reviewed a sample of the problems and called it a governance gray zone where safety becomes a convenient cover for anti-competitive practices. I'd put it more simply. If you want to be better than the competition, be better. Improve your own models. Fix your own 66.8 failure rate, compete on quality, not sabotage. I've always believed that competition should happen on an even ground. You prove your superior by being superior, not by making the other person look worse. And I'm not saying Open AI and Anthropic and Google are perfect, by the way. Nobody is. Everyone's trying their best in a field that is moving faster than anyone's ability to keep up.

But this is not competing. This is why we can't have nice things. And the reason Meta might be resorting to these kinds of tactics is becoming embarrassingly clear. In one of the more quietly humiliating developments in recent tech history, Google informed Meta that it was reducing Meta's access to its Gemini AI models via API. The reason was that Meta's demand was so enormous that it was crowding out Google's other enterprise customers. Google, a direct competitor, essentially told Meta, "You're using too much of our AI, and our real customers need it." This is the company spending $125 to $145 billion in capital expenditure in 2026 on AI infrastructure. By the way, nearly double what it spent in 2025. This is the company that fired 8,000 people in May to fund its AI pivot. The same company that force marched 6,500 engineers into an applied AI unit to build and train its own models. The company that runs Meta Super Intelligence Labs under Alexander Wang, the former CEO of Scale AI, whom they acquired for 14.3 billion. This company was relying on Google's AI for core internal functions because for the task it needed, Gemini performed better than Meta's own llama models. For 14 billion dollars, I would expect the man to also do my laundry and remember my birthday. But apparently all they got was a lab that still cannot be Gemini at homework. That is an $135 billion confession of inadequacy. Do you know how embarrassing it is to spend $135 billion on infrastructure only to have Sundar Pichai essentially knock on your door and basically say, "Hey, you're using too much of our internet. Turn off your chatbot. My real friends are trying to use it." It's like renting an apartment from the guy you're trying to evict. And it raises a question that I keep coming back to. Why isn't Meta investing that same focus into the one thing it's actually best at? Meta is about to become the number one digital advertising company on Earth. Its Advantage Plus automated ad suite is running at a $60 billion annualized clip. Its Andromeda system uses the advertisers's creative as the primary targeting signal, which is a fundamentally different approach from the demographic box model, the defined digital advertising for two decades now. These tools are genuinely innovative and they're powered by a data set that nobody else on this planet can replicate the behavioral signals of 4 billion users. But since I made the previous video on this, I actually went out and spoke to advertising professionals like real practitioners who use Meta tools every day. And what they told me is that the targeting and optimization tools still need significant work, which in advertising software usually means the machine has developed the confidence of a god and the judgment of a Roomba. Technically, yes, it's automated, but Stone still needs to rescue it from under the sofa every 40 minutes. The targeting isn't as precise as it should be. Campaigns require more manual oversight than the automated branding suggests. As a computer scientist, these are solvable problems. Model accuracy improves with better training data, more refined architecture, and sustained engineering focus. Meta has the data, the richest advertising data set in history. Meta has the engineers, or it did before it drained thousands of them into a data labeling unit to make coding puzzles. What Meta doesn't appear to have is the focus, the willingness to basically say, "This is what we're best at. This is where the money comes from. This is where the investment goes."

The advertising engine could be the most powerful, most creatively sophisticated, most precisely targeted system in the history of commerce. And instead, is being neglected while $135 billion chases frontier AI models that Google builds better. If I ran a company with this much data, this much revenue, and this much potential, I would be singularly obsessed with building the most powerful advertising specific AI in existence. I wouldn't be diversifying into 17 contradictory directions willy-nilly and then wondering why none of them seem to work well enough to keep me off Google's API. But then again, that might be why nobody has asked me to run a trillion dollar company. Although, and I'm just putting this out there, some of you in the comments have said that you would vote for me. So, the offer stands considerate. I promise not to name my waste management contractor after a farm animal. So, just to recap, contaminated water, fake children, an AI so insufficient they had to borrow Google's. Each of these stories is individually troubling, but together they paint a picture of something more structural. An organization operating at a scale that affects billions of people with a level of oversight that wouldn't pass muster at a regional car wash. And that pattern is exactly what has the AI research community, myself included, losing sleep. In an earlier video, I compare the risk of reckless AI deployment to the conditions that preceded Chernobyl. Not the disaster itself, but the culture, the cutting of corners, the assumption that the system would just hold, the suppression of warnings from the people who understood the technology best. Some of you told me in fairly direct terms that the analogy was inappropriate, that I was being dramatic, that I didn't understand Chernobyl. Since then, Stuart Russell, distinguished professor of computer science at UC Berkeley, where I spent some time during my academic career as well, has publicly warned that AI cannot afford to wait for its own Chernobyl moment. His words, and I'm quoting, "If there is a Chernobyl scale disaster with AI, it's not just going to be a regulatory response. It's going to be a societal response. People will say, "Shut it down. All of those trillions of dollars that we hear about being invested will be wasted." Michael Waldridge, professor of computer science at Oxford, drew an even starker parallel. The Hindenburg disaster destroyed global interest in airships. It was a dead technology from that point on. And a similar moment is a real risk for AI. And Daria Amade, CEO of Anthropic, wrote in a 19,000word essay that humanity is about to be handed almost unimaginable power. And it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it. Fearing a Chernobyl style disaster is not a fringe position. This is not dramatic. This is the academic and industry mainstream. the people who are actually building this technology saying openly that they are frightened by the pace at which is being deployed relative to the maturity of the systems governing it. The people building AI are more frightened than the people using it. And that should give all of us pause. When the academic community uses the Chernobyl comparison, we are not predicting a literal nuclear meltdown. We're pointing to a historical pattern. An industry that moves too fast, cuts too many corners, suppresses internal descent, and creates the conditions where a single visible failure crystallizes all the accumulated public distrust into permanent rejection. Chernobyl didn't just destroy a reactor. It destroyed an entire energy paradigm for a generation. Nuclear power is by many measures one of the cleanest and most efficient energy sources available. and it still hasn't recovered its reputation 40 years later. The Hindenburg didn't just crash, it ended the age of airships permanently. The question the researchers are asking essentially is what happens if a similar moment arrives for AI? And when I look at the evidence in this video alone, a company that contaminated a city's water supply through negligent industrial discharge, a covert program that used fabricated child accounts to probe competitors, while its own child safety failure rate was 66.8%. I think the researchers fear is entirely rational. Not because Meta is going to cause a nuclear meltdown, of course, but because the pattern of negligence, hubris, and misplaced priorities creates exactly the kind of environment where something goes badly, visibly, catastrophically wrong, god forbid. And when it does, the backlash won't distinguish between Meta's failures and AI's potential. It'll take the whole field down with it. Is this already happening? Maybe not as a single catastrophic event, but as a slow accumulation of negligences, each one individually survivable, I guess, but collectively devastating. I think the answer is starting to look like a yes. You know, every time I sit down to work with AI, building something, testing something, exploring a new capability, I feel genuine joy. The tragedy here isn't the code. The Transformers architecture is beautiful. The math is elegant. The researchers are doing breathtaking work. The problem is that the technology is being stewarded by people who have the long-term vision of a goldfish and the impulse control of a toddler in a candy aisle. I read research papers the way some people read novels, and I'm only a little bit embarrassed about that at this point. It used to be so much more. Like a proper nerd, I get excited about architecture improvements and benchmark scores and novel approaches to reasoning. Although, unlike a proper nerd, my stash contains at least 300 dresses. When an MLM disproves a geometry conjecture that stood for decades or achieves fellowship level mathematics seven years ahead of expected predictions, I feel something that I could only describe as awe. This is one of the most extraordinary scientific developments in human history. And I am lucky to be alive at the same time as it. And then I open the news and it's a data center that poisoned a city's water with an industrial bacterium. And it's hundreds of contractors dressing up as children to sabotage competitors AI. And it's a CEO in a super yacht while his CTO admits the company is falling apart from the inside. And it's $135 billion spent on AI infrastructure that still cannot outperform Google's models for the company's own internal tasks. It breaks my heart genuinely. It wounds me because this was supposed to be the moment, our moment. This was the technology that could transform humanity into something genuinely creative and precise, accelerate scientific discovery, augment human capability in ways that previous generations could only imagine. And instead of nurturing it, instead of stewarding it with the care and the focus and the humility that something this powerful demands, the people in charge are treating it like a toy, like a weapon, like a narrative device for quarterly earnings calls, like a justification for firing thousands of people while posting record profits, like a cover story for corporate espionage run under a film festival code name. AI is not going anywhere. It is too powerful, too promising, and too deeply embedded in the trajectory of human progress to be stopped by bad actors, no matter how spectacular their failures. But the gap between what this technology could be and what is currently being done with it, that gap is where the tragedy lives. Unfortunately, Meta keeps arriving at that gap with a forklift, a keynote deck, and absolutely no adult supervision. They want the historical legacy of the Manhattan project, but they're running the company with the impulse control of a Discord server and every contaminated water system, every fake teenager account, every engineer drafted into data labeling they didn't choose, every superyacht parked in a harbor full of freshly unemployed people, makes that gap wider. I've been saying this on this channel for a while now, and I'll keep saying it until somebody listens or until you make me president, whichever comes first. The technology is extraordinary. The deployment is very reckless. And the distance between those two things is costing us something that money cannot buy back. The public's willingness to believe that this can be good. That is the most expensive thing the AI industry is losing right now. Not me, not Hermione. You. Sorry, wrong video. Not money, not engineers. Trust. If you want to understand how Meta arrived at this point, the super yacht, the morale crisis, the $80 billion metaverse habit, and why the most valuable advertising data set on Earth keeps getting ignored, I cover that in this video that I'm linking here. Thanks so much for watching. I'll see you in the next one.