The Great AI Scramble: Who Wins When Nobody Knows the Rules Yet

---

A former Meta employee downloaded 30,000 private Facebook photos and nobody caught it until after he was fired. Meanwhile, OpenAI is pitching companies on four-day work weeks. Google’s AI is confidently citing Facebook posts as authoritative sources. China just got obsessed with training an AI to raise lobsters.

This is what the AI revolution actually looks like right now: messy, contradictory, and completely unmoored from the hype cycle that preceded it.

The Trust Problem Nobody’s Solving

Start with Google’s AI Overviews. The company built a system that generates authoritative-looking answers to search queries, pulling from “an array of sources, from trustworthy sites to Facebook posts.” Think about that for a second. Your grandmother sees a confident answer from Google’s AI and has no way to know whether it came from peer-reviewed research or someone’s Facebook rant. The presentation is identical either way.

This isn’t a bug. It’s the fundamental architecture of the problem. You can’t train a large language model to be careful about epistemic humility—the whole point is confidence, fluency, the illusion of knowing. When you package that in Google’s typeface with Google’s layout, you’ve created a trust Trojan horse.

Photo by George Piskov / Pexels

The Meta breach makes this worse. An employee grabbed 30,000 photos. Thirty thousand. The company discovered it only after firing him. That’s not a security failure—that’s a visibility failure. Meta didn’t know what was being accessed. And if they’re struggling to track data movement inside their own walls, what chance do they have of keeping external threats at bay?

Here’s my honest read: we’re about to see a spike in data exfiltration cases, and companies will blame AI. But the real culprit is that nobody’s built monitoring for the AI era. You had one job—watch what leaves the building—and we’re discovering that job was harder than advertised.

The Cybersecurity Arms Race Gets Weird

Anthropic claims its new model, Mythos, is a “cybersecurity reckoning.” The catch? They’re not releasing it. They’re working with 40 companies in controlled settings to figure out how to prevent cyberattacks.

Meanwhile, other companies are scrambling to deal with what they’re calling “a code overload”—the glut of AI-generated code flooding their systems. And the underlying issue is the same: hackers now have access to the same AI tools that defenders do. The speed advantage just flipped to offense.

OpenAI and Anthropic aren’t dumb. They know that releasing a powerful cybersecurity AI is like handing attackers a blueprint of your own defenses. So they’re keeping Mythos locked down, working with 40 friendly companies to test it. That’s not a product strategy. That’s damage control dressed up as innovation.

The real question nobody’s asking: what happens when hackers fine-tune their own version of Mythos?

The Workplace Experiment That’s Actually Telling

OpenAI pitched companies on four-day work weeks to “adapt to AI era.” This is fascinating not because it’s generous, but because it’s an implicit admission. If you need to compress your work week, you’re saying AI is either (a) replacing some of what humans do, or (b) so chaotic to implement that people need recovery time.

Probably both.

My take? This is a trial balloon for something bigger. If OpenAI can get companies comfortable with four-day weeks as “adaptation,” it becomes normalized. Then the conversation shifts from “AI replaces jobs” to “AI changes how we work” to “of course we work four days now.” It’s positioning, and it’s smart.

But here’s what worries me: if companies do adopt this, it won’t be out of generosity. It’ll be because they’ve automated enough tasks that they don’t need people for a fifth day. The four-day week becomes the consolation prize for technological obsolescence.

Photo by nappy / Pexels

The China Lobster Moment

An AI assistant in China sparked a frenzy where thousands of people started training the tool to help them raise lobsters. Users customized it, learned from it, built a community around it.

This happened because the AI was useful for something real, and people are endlessly creative about finding uses for tools. But it also reveals something about how AI gets adopted in practice: not top-down from companies, but bottom-up from users seeing what sticks.

The geopolitical angle here is what China and the US both understand: you don’t win the AI race by making the smartest model. You win by making the model everyone wants to use. The lobster frenzy was harmless, but it was also a user-engagement victory. That’s the real competition between the US and China—not raw capability, but adoption, customization, lock-in.

China’s not just winning “one AI race” and the US another. They’re winning the race to make their AI indispensable in daily life. That might matter more than raw capability in five years.

The Teens Are Fine, Actually

Here’s something that got buried: teens are using role-playing chatbots to harass them, confide in them, chat with a block of cheese, and deal with loneliness. This isn’t a crisis. It’s just what happens when you give people an audience that never leaves.

I’m genuinely uncertain what the long-term impact is here. On one hand, it’s a void-filling tool that might prevent worse outcomes. On the other hand, it’s training a generation to expect validation from systems optimized to provide it, consequences be damned. But I notice people treating this like a problem when it might be neutral, or even useful. The fact that teens can roleplay with a sentient block of cheese without judgment? That’s actually less dystopian than a lot of alternatives.

The SEO Death Spiral Is Real

Businesses are scrambling to get noticed by AI search because they know the old game is ending. If Google’s AI Overviews answer your question directly, you’re not clicking through to websites. You’re clicking nowhere.

This is a death spiral for publishers. Less traffic means less revenue. Less revenue means less ability to produce quality content. Less quality content means AI training data gets worse. Worse training data means worse AI summaries. Bad summaries mean users distrust the system.

We’re maybe eighteen months into this and you can already see the flywheel starting to reverse.

The companies adapting fastest are the ones changing “the way they present information on their websites” to get noticed by AI. But that’s just another form of gaming the algorithm. It worked for Google search in 1998. It worked for social algorithms in 2010. Every time we’ve done this, the quality goes down and the manipulation goes up.

Photo by Markus Spiske / Pexels

What I’m Watching

• The Meta data case going to court (timeline: next 6-12 months). This will set precedent for what “accessing your own company’s data” legally means in the AI era. If the employee walks, expect every company with cloud infrastructure to panic. If he’s prosecuted, expect whistleblowers to think twice.

• Whether any of the 40 companies using Anthropic’s Mythos actually deploy it in production. If it stays in controlled settings past Q2 2025, that’s a signal that even Anthropic doesn’t trust its own cybersecurity AI. The moment one company goes public with “we’re using this to stop attacks,” the entire threat landscape changes.

• Google’s AI Overview accuracy metrics in six months. Not the PR version—the actual studies from third parties. If accuracy drops below 85% on simple factual questions, publishers will revolt. Below 80%, the system becomes unusable. That’s the real cliff.

• How many companies actually adopt four-day weeks because of OpenAI’s pitch. If it’s less than 5% of their customer base by end of 2025, this was just PR. If it’s over 20%, we’re entering a genuinely new era of work restructuring, and the jobs question becomes unavoidable.

The chaos you’re seeing isn’t growing pains. It’s the actual shape of the future. Companies don’t know how to handle data they can’t see moving. People don’t know if they’re talking to bullshit or truth. Defenders are always one step behind attackers. Nobody’s written the rules because nobody knows what the game is yet.

That’s not a problem to solve. That’s just where we are.