AI Just Divided Humanity in Two. Do You Know Which Side You're On?

Last week Anthropic didn't release a model. That might sound like a non-story. But it's the first time in the history of this industry that a lab decided something was too good to release to the public. It's called Mythos. And we're not going to see it.

Not because it doesn't exist. But because someone decided some people deserve access and others don't. For the first time, access to the most advanced AI stopped being a subscription question. It became a question of who you are. The official justification is safety — but neither you nor I will have access to this model, at least not for now.

That's the signal. The moment technology stops being democratic. And if you didn't register it as what it is — the formal creation of two social classes — it's because you're still thinking this is a conversation about tools.

Meanwhile, the Drug Is Starting to Fail

But before talking about social classes, something is happening in real time that deserves attention.

Over the last few months, something strange has been appearing in forums, developer groups, and threads from serious people: models that in 2025 and until recently seemed like magic — compressing weeks of work into hours, generating apps from scratch in an afternoon — are declining in quality. It's not collective anxiety. Thousands of users are documenting that outputs are worse, that they have to wait until a certain hour when the system isn't overloaded, that the $20 subscription is no longer enough and you have to upgrade to the $200 tier.

The acceleration was so brutal that it created dependency before anyone could question it. A productive drug with magical results. And now the supply seems to be adjusting.

Coincidence? Technical degradation? Or the natural consequence of having created something powerful enough to become indispensable, and now calibrating who gets access to which version?

That's exactly the question I think matters more than any answer being given right now.

The AI That Builds the AI

And there's a meta-point that's even more vertiginous.

Anthropic announces it has a new model. How did they build it? Artisanally, engineers writing code line by line? Or was the previous model the one that helped build the next? Nobody is saying this clearly. But if the answer is the second — if AI is already building improved versions of itself with human supervision but not human authorship — then the board changed games a while ago.

Just a few days ago, Jensen Huang was saying on Lex Fridman's podcast that AGI had been reached. He dropped that bomb and people carried on like nothing happened. We've completely lost our capacity for awe. I'm sure aliens could show up tomorrow and we'd keep scrolling. But what Jensen says isn't marketing. He's the CEO of the company that manufactures the hardware everything runs on. When he says we've arrived, it at least deserves a pause to think.

Chess Is Over. Now It's Squash.

Mo Gawdat, former Chief Business Officer at Google X and author of Scary Smart, is one of the people who has seen most closely how all this gets built. And he uses a metaphor I find the most honest of all that circulate.

AI already won at chess. Everyone knows that. What nobody has finished processing is what it means that it's now playing squash.

Chess is deterministic. Fixed board, fixed rules, known positions. If you have enough computational power, you solve it. Deep Blue proved that almost 30 years ago.

Squash is different. The ball bounces at angles you can't fully predict. The opponent moves. The environment changes in real time. Calculating faster isn't enough. You have to adapt.

That's what's happening. AI is no longer demonstrating it can solve structured problems better than you. It's demonstrating it can navigate dynamic environments — your work, your industry, your market — far better than you.

Even those who built these tools are saying something that would have sounded ridiculous before: that coding as an act of writing code line by line is already secondary. Some industry references have gone months without opening an editor. "The most important programming language today is English" — the quote is a cliché now, but its weight hasn't landed yet in most enterprise IT departments.

It's not that programming doesn't matter. It's that the mechanical act of writing syntax stopped being the bottleneck. What matters now is understanding what you're building and why. Without that, the tool doesn't amplify anything.

FACE or RIP. No Neutrals.

Gawdat describes two response patterns to this transition. They're so clear they're uncomfortable.

The first is FACE: Fear, Anger, Conspiracy, Existential dread. Most people you see on social media discussing AI are here. Fear first, anger second, conspiracy theory as refuge, and nihilism as destination. The "couldn't they be making us dependent on purpose?" is a legitimate question, but it ends up paralyzing you instead of moving you.

The second is RIP: Reward, Innovation, Progress. It's the minority that understood the world before 2023 isn't coming back, that fighting against that is an energy loss, and that the only rational move is to learn to play the new game. Not with naivety — with eyes open to the risks — but moving anyway.

The two social classes AI is creating won't divide by internet access, formal education, or country of origin. They'll divide exactly here: between those who process disruption as data and those who process it as threat.

Mythos is the symbol. But the gap is already opening in your company, your team, at every moment you decide whether you use AI to think or to avoid thinking.

The Next 12 to 15 Years: Hell Before Heaven

Gawdat doesn't sugarcoat this. There's a window of 12 to 15 years that's going to be very difficult.

Not because AI is bad. But because jobs will disappear faster than new ones appear. The destruction curve always precedes the creation curve. In the Industrial Revolution it took society decades to reorganize. This time the cycle is shorter and more violent.

The end of that window, according to Gawdat, is something he calls the energy minimum: a world where AI and robotics produce enough that the cost of basic goods falls to nearly nothing, and human work is freed for what humans do well — creativity, connection, judgment, care.

But to get there, you have to go through the middle.

And in the middle, governments will have to do something nobody who believes in markets wants to admit: some form of universal basic income. Gawdat calls it the "communist pivot" — not as ideology but as mathematics. Sam Altman has been flirting with the idea in his experiments for a while. But when work as a commodity disappears at massive scale, you need a different distribution mechanism. Not because it's a good philosophical idea. Because there's no other viable option. And the cascading consequences for the entire economy are wild.

Nobody in enterprise is putting this in their digital transformation roadmaps. But the numbers won't wait until we're comfortable with the conversation.

IQ 700. But There's a Catch.

Gawdat says AI effectively gives any person access to reasoning equivalent to an IQ of 700. In terms of analytical capability, synthesis, pattern identification, and option generation, the ceiling rose for everyone.

The catch is this: a multiplier applied to zero gives zero.

If you don't have your own judgment, if you don't understand the domain you're operating in, if you can't evaluate whether what AI returns makes sense — then the IQ 700 is useless to you. It's like giving a precision saw to someone who's never held a tool. The accident isn't the saw's fault.

The term Gawdat uses is Augmented IQ. It's not that AI thinks for you. It's that you think with AI as an extension of your cognitive capacity. For that you need a cognitive capacity to extend.

That's where the deepest crack in enterprise sits. SAP consultants who built their career executing transactions, following runbooks, applying standard solutions have very little to amplify. AI doesn't enhance them. It exposes them.

Last week I asked SAP Joule to explain what a class was doing that was crashing a server. It invented everything. Gemini detected inconsistencies in the answer. But the final decision was mine. Not any AI's. That's the exact point: the person who has 15 years of understanding why SAP's data model is designed that way, what happens when you touch a certain configuration, how an integration behaves under load — has something real to amplify. IQ 700 applied to that accumulated judgment is devastatingly powerful.

Dependency without judgment is the FACE scenario. Augmented IQ is the RIP scenario.

The 4 Skills That Won't Be Automated

Gawdat identifies four capabilities that in the visible horizon won't be replaced. Not because they're mystical. But because they require exactly what AI doesn't have: situated judgment, lived context, embodied ethics.

AI mastery. Not knowing how to use ChatGPT. Understanding how these systems work, what their real limits are, when to trust and when not to, how to build workflows that amplify rather than atrophy. The difference between driving a car and knowing mechanics.

Agility. The ability to shift frames when the environment demands it. Learning a new domain fast enough to be relevant before the next cycle leaves you behind. This isn't trained by reading LinkedIn articles — and it's not the agility from Scrum courses.

Ethics. In a world where AI can generate anything, the ethical compass isn't philosophical — it's operational. The professional who can't distinguish what should be done from what can be done will produce damage at industrial scale.

Skepticism. Not paranoid distrust. Technical skepticism. The ability to receive an AI output and ask yourself: does this make sense? What did it assume that I didn't ask for? What is it not telling me? The person who accepts the first output without questioning it isn't using AI. They're being used by it.

These four aren't new skills. They're what always distinguished the professional who thinks from the one who executes. The difference is that before you could hide their absence behind routine technical work. Now you can't.

Your Diagnosis This Week

Back to the beginning.

Anthropic has a model they won't release because it's too capable to free without control. That's not a technical curiosity. It's the formal announcement that premium AI will be behind walls. And meanwhile, the mass version degrades enough that those who depend on it have to pay more or wait their turn.

The gap isn't future. It's happening now.

Today that gap is still crossable. Gawdat's 12-to-15-year window is still open. The game can still be learned.

But Mythos is the first sign that the window has a lock.

So, before closing, three concrete questions for this week:

Can you explain to someone what AI did for you today, without the AI? If you can't, you're not the orchestrator. You're the one who signs without reading.
What domain knowledge do you have that AI can't reproduce without your context? If it takes you more than 10 seconds to answer, that's the gap you need to close.
What did you learn in the last 4 weeks that wasn't generated by a model? If the answer is nothing, the process has already begun.

FACE or RIP. The chess board is over. Now it's squash. And the match has already started.

Shall we begin?