The AI Pause Nobody Wants: Why We Should Stop Building GPT-7 and Start Understanding GPT-4
- Andrei Raileanu
- Apr 28
- 7 min read
On power concentration, the innovation myth, and why 7 people deciding humanity's future should terrify you
By Andrei Raileanu — The AI Whisperer
Emad Mostaque says we have around 800 days before human cognitive labor becomes economically obsolete. Honestly? I believe him. Not because I think tech will suddenly leap forward in 800 days and become magical—no. It's because we barely grasp what we've already built, it's all wildly unregulated, and maybe seven people truly control what happens next.
Seven.
Not politicians.
Not state actors you can vote for.
Not even accountable billionaires.
Just—seven guys in various boardrooms, calling the shots about what comes next. Sam Altman. Dario Amodei. Demis Hassabis. Elon Musk. Sundar Pichai. Satya Nadella. Mark Zuckerberg.
That's your list. That's who's holding the steering wheel.
And look, there's a hard question here: Why jump into GPT-7 when we barely understand GPT-4? Why sprint toward AGI when we haven't even figured out how to govern the AIs we already have? Why let a handful of corporations, driven mostly by profit and their own ambitions, decide the future of human civilization?
And why does asking any of this instantly get you labeled a Luddite?
The uncomfortable truth: We don't need GPT-7
What Mostaque is saying should really make everyone stop and think. Everything we need for human-level AI is already on the table. No more breakthroughs required—we just dump more compute into our current architectures. Sure, GPT-4 hallucinates. Claude gets things wrong. Gemini has alignment issues. But those aren't problems we'll solve by building GPT-7. They're problems we solve by understanding GPT-4 better.
The "scaling hypothesis" keeps proving right: more GPUs and more power equals better models. There's no secret magic left, just more expensive training runs.
Meanwhile, GPT-4, Claude Opus, Gemini—these models already:
Write code better than most programmers
Draft legal contracts better than most lawyers
Diagnose stuff better than your average doctor
Spin up essays and stories as well as any human
Yet here we are, racing to build GPT-6, Claude 5, Gemini 4.0.
Why? Not because GPT-4 is everywhere or thoroughly understood—not because we've nailed safety. Not because we have actual governance or oversight. Instead, it's a race among the companies. A race for headlines, funding, dominance, and power. It's an innovation theater.
Who decides? Two guys at a party
So, here's how Emad Mostaque tells the OpenAI origin story: Larry Page meets Elon Musk at a party.
Page talks up AI, saying, "Wouldn't it be amazing if AI replaced humanity with something better?"
Musk pushes back: "Actually, I like humans. What happens to them?"
And just like that, the founding of OpenAI is set in motion. That's literally the spark—one conversation. From there, you get breakaway companies, personal feuds, a spiderweb of competition, and shifting alliances. "The only good AI is the one I build, that stops all the others." It's Game of Thrones, but with neural nets.
Are these people evil? No. They're human. With egos. With competing visions. With financial incentives. With virtually zero public oversight. And they're making decisions that could reshape human life, maybe forever.
The myth of "inevitable" innovation
Every time anyone suggests slowing down AI, the same talking points get wheeled out:
"You can't stop progress." "Innovation is unstoppable." "If we don't do it first, China will."
It's mostly marketing spin. Progress toward what, exactly? Innovation for whom? Who decided it's inevitable? Think back: we put the brakes on human cloning. We paused risky virus research. We set rules for nuclear weapons. We've decided, collectively, that certain tech is just too risky to move ahead with, without rules.
AI isn't really different. The trouble is, unlike nuclear power, the people building AI get to write their own rules. Anthropic writes its own "AI Constitution." OpenAI drafts its own safety playbook. Google draws its own lines.
The fox watches the henhouse. And we're the hens.
Saying "you can't stop progress" really means: "We, the people building this, don't want to stop, and we have the power to keep going, no matter what the rest of you want."
What would actually happen if we hit pause?
I'm not saying ban AI. I use Claude daily to build QuickFusion AI. I'm launching TAKUMI with AI at its core. This isn't anti-tech. But using what we have is different from forging ahead, recklessly, into versions we still don't get.
A genuine pause would look like:
Stop training on huge new models past a clear compute threshold—until governance frameworks are in place.
Focus attention and resources on understanding and safely deploying current models, instead of rushing to invent new ones.
Set up independent oversight (not just internal company reviews) for any system that can seriously affect society.
Make transparency mandatory: training data, model capabilities, safety testing—publish it before release.
Dedicate a chunk of compute power for public goods—education, healthcare, climate efforts—not just profit.
That's not Luddism. That's just basic engineering. You don't add a new floor without checking the foundation. You don't roll out version 2.0 while version 1.0 is still riddled with bugs. But, for some reason, AI is the exception.
Welcome to digital feudalism
Emad calls this "digital feudalism"—honestly, I think we've already arrived. When Anthropic can shut off Claude overnight, businesses built on it just… vanish. That's not a marketplace. That's a fiefdom.
When OpenAI controls access to o1, sets prices, and nobody knows how the decisions are made—that's not open innovation, it's an unaccountable gatekeeper.
Google holds Gemini, Microsoft owns Copilot, and Meta runs Llama. None of them will tell you what they trained on or what these models can truly do. That's not just "progress," that's all the power stacking up into fewer and fewer hands.
Emad calls it "taxation without representation." They took humanity's collective output—our writing, our art, our conversations—and used it to train AI that will now take our jobs.
We didn't sign up for this. We didn't get paid for it. We don't have a vote in how it's used. But now, we're supposed to pay up to access intelligence built on our own backs.
This is feudalism, just with GPUs.
The 800-day clock
So, what does this "800 days" really mean? Mostaque isn't saying AI will become smarter than humans in 800 days. It's already better at plenty of things. He's saying: in about 800 days, hiring a human for cognitive work won't make economic sense anymore. Why pay a programmer $100,000 a year, when you can rent an AI agent that does the same job for $1,000?
It won't just be programmers. Lawyers, accountants, customer service, writers, analysts—anyone whose work can be outpaced by the machine gets replaced. And here's the tough part: this isn't science fiction. The technology's here. It's just not everywhere yet. The clock is simply counting down until widespread adoption.
No one—no governments, no industry watchdogs, not even the companies themselves—has a plan for what happens next.
What would real AI regulation look like?
If we were serious about governing AI, it would look a bit like this:
Tier 1: Transparency
Public "model cards" for each major release: what data went in, how much compute was used, what capabilities were tested, what red-teaming showed. No more black boxes. If something can swing an election, diagnose a disease, or draft a new law, you should be able to see what's under the hood.
Tier 2: Resource Controls
Labs must earmark a set percentage of compute for public-interest projects—education, climate, and healthcare. Training runs above a big threshold require outside approval. Not just a checkbox your own lawyer signs—real oversight.
Tier 3: Governance
An independent AI Safety Board with subpoena power, staffed with technical experts not on anyone's payroll. Liability: If your model causes harm—election interference, deepfake scams, automating weapons—you're responsible. No more loopholes. Real, democratic input: ordinary people actually get a say about the values encoded into AI. Not just whatever Anthropic or OpenAI decides in-house.
None of this is that wild. This is what we do with new drugs, dangerous aircraft, and nuclear plants. Tech with mass impact gets oversight.
Why this won't happen (and what that means)
Here's where I need to be direct: I don't think any of this is going to happen. Not in time. Why? Because the people who'd have to make it happen are the ones who benefit most from avoiding it.
OpenAI is valued at $150 billion. Anthropic has Amazon and Google backing it. Microsoft's betting its entire future on the sector. Meta's pouring billions into open weights. None of them is going to voluntarily slow down, open up, or accept real referees.
And governments? They're way behind. Most lawmakers think "AI" is a slightly better spell-check.
By the time anyone catches up, by the time there's real understanding or oversight, the decisions will already be baked in—by those seven people, for their own benefit.
The window is closing faster than most people realize.
What can you actually do?
So all this—what's the point? Are we just along for the ride? Not quite. Understanding the stakes is the first step. Every time someone claims "it's inevitable," you get to push back. You can choose which companies, which models, and which futures to support.
Use open-source models when you can—Llama, Mistral, Stability. The more power that moves toward open communities, the stronger the pushback against digital feudalism.
Demand transparency from AI companies. Ask: what did you train on, how do you test safety, who audits you?
Back regulation. When (if) governments do wake up, support oversight, even if it's flawed at first.
Build on purpose. If you're using AI to make something, ask yourself: does it really make the world better, or just cheaper?
Talk about it. The more people wake up to what's at stake, the harder it is for this stuff to get decided entirely behind closed doors.
The question that haunts me
If seven unelected people are making irreversible decisions about technology that will shape civilization, and they're driven by profit, competition, and their personal beliefs—not the public interest— and there's no real way to stop them, or redirect them, or vote them out.
What are we really consenting to?
We didn't vote for this. We weren't even consulted. We're not told what's being built, or how. By the time it becomes clear, it'll be too late to course-correct. Emad says 800 days. I'd say the window is even tighter.
Not because the tech will suddenly become dangerous overnight. But because the decisions—about how it's used, who benefits, who's left behind—are being made right now. Behind doors we'll never walk through. By people we never chose. For a world none of us got to vote on.
And if that doesn't terrify you, you're not paying attention.
Andrei Raileanu is The AI Whisperer — founder of UDV and HumanAILabs, building AI-powered businesses while questioning the systems that make them possible.
More at andreiraileanu.eu
References:
Comments