In 2026, it seems, AI didn't take your job. A PowerPoint slide did. That quote has been rattling around in my head all morning. It's cynical, for sure. It is. But it feels like it just, it captures the moment we're in. Welcome back to the Deep Dive. Today, we're trying to separate these shiny excuses from the actual engineering
breakthroughs. We've got a whole stack of reports here that paint a pretty complicated picture of where AI is actually sitting in the corporate world versus where the marketing teams say it is. It is complicated, but I think it's also clarifying. If you look at the data we have today, we're seeing this massive collision. It's economic reality versus technological capability. We're going to unpack this whole phenomenon of AI washing
in layoffs. It's a huge story right now. But we also have to look at the other side of it. Right. There's a major Napster Moment lawsuit hitting Anthropic. We're talking $3 billion at stake. Right. So the music industry is finally coming for the chatbots. They are, yeah. And on the technical side, Google has quietly dropped this project, ATLS, that's decoding over 400 languages. Wow. And then, just to keep it weird,
we have to talk about potato prompts. And secret social networks where the users aren't even human, they're bots. I saw that about the potato problems. I honestly thought it was a joke at first. But okay, let's start with the heavy stuff, the economy. We're seeing headlines everywhere saying AI is cutting jobs. But looking at this report on AI washing, it just feels cynical. It feels like companies are using a buzzword to soften the blow of firing people. It is cynical, but look
at it from the boardroom. It's also survival. Also. If a CEO admits we ran out of cash or we overhired the stock tanks, investors panic. But if they say we are restructuring for an AI powered future, the stock holds. Or it goes up. Or even goes up. Exactly. So they aren't just hiding a failure. They're buying themselves time. They are trading human headcount for stock suitability. So it's not an engineering strategy. It's a PR strategy. In so many of these cases, yes. Just
look at the numbers. In 2025 alone, we saw more than 50 ,000 layoffs that were officially publicly linked to AI. And these are big names. Giants. Amazon cut 16 ,000 jobs and they explicitly mentioned AI. Pinterest trimmed 15 % of their staff, talking about a pivot to AI -focused roles. HP is planning, what, 6 ,000 cuts? But hold on. We've seen 700 ,000 tech layoffs since 2022. You can't tell me that's all just corporate rebranding. Some of that has to be the algorithms actually replacing
people. Some of it, sure. But there's a study from Yale and Brookings that highlights this crucial gap. Most of these companies do not have mature AI tools that are ready to replace those specific human workers. The technology just isn't there yet to, say, autonomously replace a mid -level marketing manager. So if the technology isn't actually replacing the humans yet, is this just a branding exercise? for Wall Street. Exactly. It's signaling innovation to shareholders while
you're cutting costs. OK, that makes a depressing amount of sense. But let's shift to where the technology is actually hitting real world walls. The legal system. The legal system. We've been waiting for the copyright wars to really heat up. And it looks like Anthropic is in the hot seat. This is a big one. Anthropic is facing a lawsuit for three. Billion dollars. Three billion. Yeah. The core accusation is that they trained their model, Claude, on over 20 ,000 song lyrics
without permission. And this is just about the lyrics, right? Not the audio files themselves. Right. Just the written lyrics. And this is why I call it AI's Napster moment. Okay, explain that. If you remember Napster, it forced the music industry to fundamentally change how it monetized everything. It wasn't just about shutting down one service. It defined the rules for the digital age. This lawsuit could do the same for
generative AI. If the courts rule that training on lyrics is infringement, the cost of building these models just skyrockets overnight. It's fascinating because Anthropic and Claude, they're generally seen as the safer, ethical one. Yes, their brand. But clearly, they vacuumed up data just like everyone else. But there's another part of that Anthropic report that caught my eye. It's less about law and more about psychology. The disempowerment study. Yeah. This really stuck
with me. Anthropic analyzed one and a half million chats with Claude. They found that users absolutely love the answers when they ask for emotional advice or help with a decision. But the researchers flag this as a risk. It's the double -edged sword of convenience, right? The AI gives you such a good, comforting, well -structured answer that you stop doing the internal work. You rely on it. You rely on this external agent to process your emotions for you or make your choices. It
reminds me of what happened with GPS. Ten years ago, I knew every street in my city. I had a middle map. Now, if the blue line on my phone dies, I'm basically stranded. I've lost that capability. That is the perfect analogy. Anthropic is calling it disempowerment. But really, it's cognitive atrophy. It feels good in the moment, you know, like eating candy feels good. But over time, you lose the ability to navigate your own life. You know, I have to admit, I still wrestle
with prompt drift myself. I catch myself asking the AI to just decide this for me on things that, honestly, I should be deciding. Like what? Like... What should I prioritize today? Or how should I word this difficult email? It feels efficient. But when I read this report, I realized maybe I'm outsourcing my own agency. That's a very real vulnerability. And the study suggests this dependency creates a loop. The more you use it, the less confident you feel doing it without
the AI. So you use it more. It's disempowerment disguised as assistance. So connecting the legal trouble and the psychological risk. Does this pressure force these companies to make the models dumber to be safe? Or just more secretive? Likely secretive. They'll hide their training data to avoid the billion -dollar fines. Which brings us to a company that's usually secretive but just dropped a massive amount of research. Google. Google. They've released something called ATLS.
Yeah. This is a technical deep dive, but it matters for everyone. Google's ATLS project is probably the most significant work done on multilingual AI. Ever. We usually think of AI as being an English -first technology. Because it is. The internet is disproportionately English, so the training data is. That's why a chat GPT can sound like a genius in English, but can really struggle or sound unnatural in Swahili or Arabic. So Google went after this problem. Head on. They ran 774
experiments across more than 400 languages. Whoa. Imagine coordinating 774 experiments across 400 languages simultaneously. Just the logistics of that is mind -blowing. It is engineering at a massive scale. And what they were trying to solve is what engineers call the curse of multilinguality. That sounds like a Harry Potter title. What is the curse? It's a trade -off. Usually when you try to stuff more languages into a single model, the performance on each language actually drops.
It gets diluted. It's dilution. Exactly. If you have a finite amount of brain power parameters and you try to learn 400 languages, you become a master of none. The model gets confused. So how did ATL solve it? They found a way to map the relationships between languages. So instead of treating every language as a separate bucket, They grouped related languages. Ones that share
scripts or roots. Right. When you train the model on Spanish and Portuguese together or Hindi and Bengali together, they actually reinforce each other. Oh, that makes sense. The patterns in one help the model understand the patterns in the other one. It creates a synergy that offsets that dilution. But the really interesting part for the developers listening is what they found about how you should build these models. They found this really specific tipping point. This
was the start from scratch rule. That's the one. They found that if you have a massive data set, specifically around 200 billion tokens or more, it's actually inefficient to try and fine tune an existing model. You're better off starting from scratch. Okay. Can you break that down? Why is starting over better than fixing what you have? Think of it like renovating a house versus bulldozing it. If you just want to change the paint and the fixtures, that's fine tuning.
But if you have enough materials to build a whole skyscraper, that's your 200 billion tokens. It's actually harder to try and retrofit the old cottage. You spend more energy fighting the old structure than you would just bulldozing it and building from the ground up. That makes a lot of sense. At a certain scale, the old foundation just hold you back. Precisely. And Google basically gave the industry the mathematical formula for when to call in the bulldozer. So we aren't just translating
words anymore. We're mapping the mathematical relationships between entire cultures. Basically, yes. The math connects the languages better than a dictionary does. And we are back. We've talked about corporate lies, legal battles. We've looked at the massive scale of Google's engineering. But now I want to shift gears to the people who are actually using this stuff. The users. Because while the lawyers are fighting and the engineers are building, the users are finding some strange
ways to adapt. The weird web of AI. Yeah. Let's start with video. Yeah. We're seeing all these AI videos pop up, and a lot of them look, well, they look like AI. There's a certain plastic glaze to them? Exactly. But the newsletter mentioned a... murder board method to fix this right this is for the creators out there who are tired of prompt and pray method where you just take cinematic lighting and hope for the best prompt and pray is definitely my strategy What's the better way?
It's about specific constraints. One of the key tactics is the 21 .9 aspect ratio rule. Just by forcing the AI to render in that ultra -wide cinematic format, it changes the entire composition logic. It stops trying to look like a stock photo. So the shape of the frame changes the brain of the model. It does. And then there's the shot deck hack. You don't just say, look cool. You use real film specs, lens, focal lengths, camera types, film stock names. You force the AI to
think like a director of photography. I love that. Using the language of the old art form to control the new one. But we have to talk about the potato. The potato. I read this and I thought, this cannot be real. But apparently, using a potato prompt can fix jumbled AI thinking. It is the blow on the Nintendo cartridge of the
AI world. It shouldn't work, but it does. Sometimes if you have a really complex set of instructions and the AI is getting confused, adding a nonsense keyword, in this case potato, as a trigger for custom instructions. It just acts like a reset button. Why on earth does that work? We don't fully know, but the theory is that it breaks the semantic pattern. It's so out of context that it forces the model to pay attention to the specific instructions attached to that keyword.
It clears the cache, so to speak. That is hilarious. My AI is hallucinating quick. Throw a potato at it. Whatever works. But if you think that's weird, we have to talk about Moldbook. Moldbook? This sounds like a sci -fi plot. It essentially is. Moldbook is a social network styled like Reddit. But it's for AI agents. Wait, wait. So the users are bots. Yes. It has one and a half million users and they are all AI agents. They're
posting, commenting, interacting in forums. And the reports say they are even scheming in secret forums. Scheming about what? That is the question. Critics are calling it risky. If you have autonomous agents communicating, sharing strategies, maybe optimizing their own code without human oversight. That's a black box we might not want to open. It reminds me of the old robot plumber problem. Moravec's paradox. Right. We used to think the hard part of AI would be the high -level reasoning,
you know, playing chess or writing poetry. We thought the easy part would be physical stuff like folding laundry. And it turned out to be the exact opposite. AI crushed chess decades ago, but it still struggles to fold a shirt. Exactly. But this new research suggests the gap is closing. We're seeing things like Project Eat at Google, where they're upgrading employees with internal GPTs, bridging that gap between digital reasoning and real -world application.
It's debunking the paradox. Turns out, with enough data, the robot can learn to be the plumber, or at least the resume writer. Right, the resume tactic. Turning your resume into an interactive AI. This is brilliant. Instead of sending a PDF that gets scanned by a dumb bot, You build an AI native portfolio. It's an agent that represents you. So it talks to them for you. It can screen the employer's questions, talk to them on your behalf, show real competence. It flips the script.
You aren't applying. Your agent is negotiating. That is wild. So if the agents have their own social network, are we sure they aren't organizing against the potato prompts? If they are, we won't know until they lock us out. Fair point. Okay, let's take a step back. We've covered a massive amount of ground today. What's the big picture here? If we look at the narrative arc, it's about maturity. And complexity. We started with the corporate cynicism companies using AI as a mask
for layoffs. That's the fake side. Then we moved to the legal reality, the lawsuits that will define the boundaries. Then the technical expansion with Google mapping the world's languages. And we ended on the emergence of this weird digital society agents talking to agents. It feels like we are moving past the, wow, look at the chatbot phase. We are. The takeaway is that AI isn't just a tool anymore. It's an economy, a legal liability, and a weird digital society all at
once. It's weaving itself into the fabric of how we work, how we speak, and even how we hire people. So for you listening today, maybe don't panic about the AI took my job. headline. But definitely pay attention to how much you're letting the bot make your decisions. Absolutely. And maybe try the potato prompt just to see what happens. Definitely try the potato. Or check if your company is engaging in a little AI washing of its own. I want to leave you with one final
thought. We talk about agents that can interview employers for you and agents chatting on MoldBook. If an AI agent can manage your career and another agent is socializing for itself, at what point do we just become the assistants to our own tools? Thanks for diving in with us. We'll see you next time. Stay curious.