Picture a chess program that always finds the perfect move. It just needs a few billion years of computing to do it. Does that program play chess better than a person?
Sergey Markov says no, obviously. He leads the experimental machine learning systems directorate at SberDevices, part of Sberbank, and he uses this absurd program to show why one of the most quoted terms in technology has stopped meaning anything. The term is artificial general intelligence, or AGI.
The stakes are practical. Sooner or later a company will stand up and announce that it has built AGI. Markov is nearly sure of it. What you will be told, and what the announcement will actually mean, may be two different things.
The phrase comes from Mark Gubrud, who used it in the late 1990s, in a paper about military uses of nanotechnology. Markov finds that detail funny. Gubrud's idea was a distant, pleasant dream: one day we would build universal systems that, like a human brain, could solve a wide range of intellectual problems.
As a dream, it works. As a measuring stick, it falls apart the moment you touch it. Markov's complaint is that the definition is left unfinished at both ends.
Start with "any task a human can solve". Which human? Different people solve different tasks. Should the machine match an average person, or an expert? Maybe there is a problem that only one person on Earth can solve. Do we count that person too?
Then "solve". How long may it take, and at what cost in energy? That is where the chess program comes in. It technically finds the best move, and no one would call it a chess master. Try to turn Gubrud's sentence into a test, Markov says, and you land in a puddle. Everyone in the field knows it, Sam Altman included.
A researcher's paper from Meredith Ringel Morris, published in Proceedings of Machine Learning Research in July 2024, points at the same two holes: "what tasks?" and "which people?". Her answer is a framework that scores performance and generality separately.
Altman's answer was different. Markov's version of it is a company saying: thanks, Gubrud, it's a nice marketing name, but we'll write our own definition. Then they wrote it.
They've reached the point, in their talks with Microsoft, where AGI is, well, a system that can, like, earn a billion dollars.
И они там в своих переговорах с Майкрософтом дошли уже до того, что AGI у них, это, значит, система, которая сможет, там, миллиард долларов заработать.
— Sergey Markov, Gleb Solomin Podcast, 08:59
Markov says the host's "seriously?" got a serious answer. He points to a leak from OpenAI that had come out a week before the conversation. In it, the company discusses a system good enough that, once turned into a working service, it could make a billion dollars.
Press reports put even bigger numbers on it. Business Insider reported in December 2024 that OpenAI and Microsoft had signed an agreement defining AGI as a system that can generate $100 billion in profits. Markov's figure is a billion; Business Insider's is a hundred. Either way, the test is financial earnings.
Notice what happened to the question. "Can it think like a person?" is hard to answer, and a bank balance is easy. A contract can check a bank balance.
Markov doesn't endorse this approach, but he understands the motive. Large public companies want to be the first to announce AGI. His bet is that someone will do it in 2026, or by 2027 or 2028 at the latest. The question is what the announcement will contain.
He calls it speculation, a market-and-marketing move that feeds on an under-defined term. Reality is less tidy. Today's large language models still handle some classes of tasks very badly, and they have a number of fairly basic limits. None of that, he adds, takes away from the progress, which he calls huge.
So when the headline arrives, ask which AGI it means: Gubrud's brain-like generalist, or the one that clears a billion dollars?
