AGI Is a Number in a Contract
The industry sells AGI as a scientific finish line. In the fine print, it's a profit target.
Ask ten people in this industry what AGI is and you will get eleven answers. It's a machine that can do any cognitive task a human can. It's a system that can improve itself. It's the moment the computer wakes up. The definitions are gauzy on purpose, because a finish line you can see clearly is a finish line someone can hold you to. And the one place AGI has ever been defined with actual precision is the last place anyone romantic about the technology wants to look. It's in a contract.
Reporting last year surfaced the definition buried in the OpenAI and Microsoft agreement. AGI, for the purposes of the deal that determines who gets the money, is the point at which OpenAI's systems generate a hundred billion dollars in profit. That's it. Not a Turing test. Not a benchmark. Not a philosophical threshold. A dollar figure that triggers a clause. When the most important document in the field needs to know whether the machine has arrived, it doesn't check the machine. It checks the balance sheet.
Sit with that, because it explains almost everything about how these companies talk. AGI is not a scientific milestone they are racing toward. It's a marketing horizon they are selling toward. The vagueness is not a bug in the discourse. It's the whole point. A precise definition can be met, and once it's met the story is over and the fundraising ends. A vague one can be perpetually two years away, which happens to be the exact distance that keeps a valuation levitating.
A finish line you can see clearly is a finish line someone can hold you to.
Watch what happens every time a model gets genuinely good at something. For years, passing the bar exam was shorthand for real intelligence. A model passed it, and within a week the goalpost moved. "Well, it's just pattern matching, it doesn't understand." The same thing happened with the Turing test, which was the holy grail of machine intelligence for seventy years until a chatbot cleared it, at which point everyone decided the Turing test was actually a bad measure and always had been. Every capability that gets achieved gets instantly demoted from "sign of AGI" to "narrow trick." The target is defined as whatever the machines cannot do yet. By construction, you can never arrive.
This is a beautiful machine for printing money and never being wrong. If the models keep improving, you point at the improvements and say the future is arriving, invest now. If someone notes that the models still can't be trusted to book a flight without hallucinating an airline, you say that's exactly why we need another hundred billion in compute, the real thing is almost here. Progress is evidence. Failure is also evidence. Heads the story wins, tails you're not thinking big enough.
Now the fair objection. The work is real. The models are astonishing compared to five years ago, the capability curve is steep, and pretending nothing is happening is its own kind of denial. All true. I use these tools every day and they've changed how I work. But "this technology is genuinely powerful and improving fast" is a completely different claim than "we are on a defined path to a defined thing called AGI." The first is observable. The second is a narrative device, and the people repeating it most confidently are the people whose comp depends on you believing it.
Here's the tell. If AGI were a real engineering target, the labs would compete to define it tightly, publish the criteria, and race to hit them. Precise goals are how engineers work. Instead the definitions get softer as the funding gets bigger, and the word gets stretched to cover whatever the latest release happens to do well. That's not how you describe a destination. That's how you describe a receding one, on purpose, because the journey is the product and the journey has to never end.
So when a CEO tells you AGI is coming in two years, don't argue about the timeline. Ask him to define it in a sentence he'll let you screenshot. Ask what specific test, passed on a specific date, would make him stand up and say we did it, it's done, stop sending money. He won't give you one. The number's already written down. It's a hundred billion dollars, and it was never about the machine.