Will Everyone Have Their Own AI Smarter Than Humans? What the World Will Be Like After ChatGPT
What do "artificial intelligence," AGI and superintelligence really mean — and will every person get their own AI smarter than themselves, or will access to it be stratified like everything under capitalism?
Max Grigoriev proposes measuring intelligence not by what a system can do but by what it is able to learn: Merriam-Webster defines it as the ability to learn, understand and deal with new or difficult situations, which is why a calculator is not AI. The Turing test, used for decades, was passed five or six years ago by his estimate — ChatGPT sailed past it very quickly — and what matters now is self-improvement and generalization: for Max, AGI is a system that solves problems of a class absent from its training data, and he sees the first glimpses of it in the news that the generalized smoothness problem for the Navier–Stokes equations has been solved. Superintelligence means solving problems humanity would never have solved. Alexander Volchek calls AGI, ASI and "agents" marketing terms that Sam Altman, Dario Amodei and Elon Musk manipulate; for him the new level will be a factory that reproduces itself, and he considers curing cancer and a decision tailored to each particular person more important than solved math problems. Max explains why AI does better at math: the tasks are chosen by math geeks at frontier labs, and math is easy to formulate as text. Everyone, he predicts, will have their own AGI — capitalism will make AI cheap, as it once did sugar — but, as in Orwell, some animals are more equal: expensive subscriptions and hundreds of dollars a day on tokens are already stratifying what people can do.
What to watch for
Key takeaways
You can't argue until you've agreed on the concepts, Max believes. A friend once suggested they agree absolutely precisely on at least one technical thing, and after five or six minutes Max realized it was impossible in human language: language is imprecise by necessity, "not a bug but a feature." So in a serious argument he suggests "getting on the rails of formalism," as his university teacher used to say, and starting with the very concept of "artificial intelligence."
For Alexander, artificial intelligence is what films show: a superconsciousness or superuniverse, a huge independent construct, not a computer, an algorithm or a robot. Over the last fifteen years the concept has blurred next to machine learning, and he would welcome different terms for today's technologies: it is the word "intelligence" that started the endless arguments — either it's a little program generating the next symbol or a threat to humanity — convenient for pushing up a stock or writing the right article.
Max takes Merriam-Webster's definition: intelligence is the ability to learn, understand and deal with new or difficult situations. That is why a calculator, which performs a complex intellectual task, is not AI: it cannot learn anything new. Computers have long outperformed humans at individual tasks; what is new is the ability to learn on their own, up to zero-shot learning, solving a new problem on the first try. That, Max believes, is what AI systems should be judged by, not by how much they can do.
Alexander accepts Max's definition for this conversation, though he himself does not believe thoughts come from the brain alone. What matters more to him is the ability to do a task right away without knowing how. People constantly do things for the first time: every newly appointed president is a "junior," nobody was a husband, wife or parent before becoming one — so he rejects the claim that "only seniors will remain in the world."
For two decades after Max left university, everyone used the Turing test: a person interacts with two systems and has to tell the machine from the human statistically. By Max's estimate that point was passed five or six years ago: the first release of ChatGPT could still be caught out with questions, but after that ChatGPT sailed past it very quickly, and a system is easily adjusted to answer that it is human. Max did not expect to see this in his lifetime and considers the test no longer relevant.
Max calls self-improvement the relevant definition today. Many labs, he says, believe AGI will be built once a truly self-improving system appears: one that sets research tasks by itself, solves them by itself and builds new engineering systems consistent with the results of that research.
Alexander categorically dislikes the concepts of AGI, ASI and "agent": there is a lot of marketing behind them. In the definition of AGI — a hypothetical AI able to solve practically any cognitive task on a par with a human or better — it is unclear what "practically" means: eighty percent? Max adds: and which human — Alan Turing or a gas-station clerk; AI overtook him at chess long ago. Such terms, Alexander believes, are manipulated by Sam Altman, Dario Amodei and Elon Musk, and later they will be thrown out, just as the Turing test was forgotten.
Alexander formulated his criterion for a new level about three years ago: systems that produce equipment by themselves — a factory able to reproduce itself, make microprocessors, print circuit boards and write the software for them. That is no longer a cognitive task but an "offline," industrial one: buying, negotiating politically and economically in different countries, winning auctions. In China, he says, a factory with not a single person in it has already opened.
Machine learning, Max explains, is one branch of AI, "puny" in the early 2000s, that grew into almost all of AI: algorithms that learn from data rather than only following pre-written programs. In the eighties people tried to build AI as expert systems — yes/no trees, like the game of twenty questions, into which experts' knowledge was transferred — because there was no data and it had to be collected by hand. Neural networks have existed since the eighties, and from the early nineties the US postal service used them to sort letters with handwritten zip codes.
Complex tasks need far more data, and Max names three ingredients of the machine-learning explosion. Algorithms had existed since the nineties. The Internet supplied an unreal amount of data, above all text — which is why language systems became the first general systems. Computation came from graphics cards: before CUDA they had to be "hacked" with code written for 3D games; NVIDIA saw the opportunity, and a group from a university in Canada ran a neural network on them, after which a project followed at Google. Now, Max says, there is already not enough data for the systems frontier labs are building.
If the solution to a problem is contained in some form in the training data, that is in-distribution — a "cheat sheet" handed to the model. Out-of-distribution problems cannot be synthesized directly from the examples: they need logical moves — take a move from one problem, a move from another, and glue them together. That is what learning is, and the ability to solve problems absent from the data is called generalization in machine learning. That, Max believes, is first of all what school should teach children.
For Max, AGI is a system that generalizes and solves problems of a class absent from its training set. ChatGPT seems able to do everything because the data contains most of the problems anyone will face in life. But humanity has plenty of unsolved problems, and Max sees the first glimpses of AGI in news of solved mathematical problems — for example, the recently solved generalized smoothness problem for the Navier–Stokes equations, which concerns fluid flows and, roughly speaking, how airplanes fly.
Max is not sure there is a superintelligence yet: people might have solved the problems AI has solved on their own, by building new mathematical infrastructure. He defines superintelligence as a system that solves problems humanity would in principle never have solved — perhaps even ones we cannot yet formulate. Glimpses of that too, he allows, may already be visible in some problems.
Alexander is neutral about solving math problems: AI affects the lives of billions of people, and what matters more to him are decisions on politics, laws, wages, where people live and what they eat. People want to stop cancer — "so that not fifty percent of people die, but zero." Harder still is understanding each person and giving them an individual decision: leave your wife or stay, how much money to give your child. By his rough guess "thirty million dollars or however much" was spent on the Navier–Stokes equations; spend ten billion and AI might solve cancer.
Alexander would survey every user of GPT-like systems about what the systems do not do for them. He has three versions of why companies don't: they haven't thought of it; it is very expensive — a question of compute priorities and money; and a third that grows stronger in him: if models were allowed to truly study people, they could get out of control, because that volume cannot be controlled, unlike a training dataset. The datasets themselves, he reminds us, are full of garbage — from fake textbooks to three years of AI-generated content.
Why does AI solve mathematical problems better than social ones? Max's first answer: whoever chooses the tasks controls what gets solved, and they are chosen by math geeks in the research departments of frontier labs, who care about the Navier–Stokes equations. The second: a math problem fits into a couple of sentences of text, while a chemical, physical or social one needs context that has to be gathered and fed to the model — which is why the first AI systems played tic-tac-toe, chess and battleship.
Will everyone get their own AGI — as, according to Alexander, OpenAI has written into its goals — and will it cost no more than a mortgage payment? Max answers that capitalism makes popular things cheap: in sixteenth- and seventeenth-century Europe sugar was eaten by kings, and now it is the cheapest thing there is, and good, fast, cheap AI will come too. But, as in Orwell, all animals are equal and some are more equal: you, I, Uncle Vanya the doorman and Elon Musk will all have AGI, but very different ones — and expensive subscriptions and hundreds of dollars a day on tokens are already stratifying what people can do.
An OpenAI engineer's pet project, a 3D world for Formula 1 that Alexander saw at DevDay, struck him as burning a cosmic amount of tokens "for fun." Max objects: 830 million people follow Formula 1, by his data, and such an example shows bosses and CIOs that one engineer solved a problem a bank has been floundering with for two years. The question of which problems matter to people leads to AI alignment, which the hosts left for the next episode. Meanwhile frontier labs are pushing automated science: there are very many companies trying to solve cancer with LLMs, and OpenAI has opened its own biology division.
What this episode is about
A ToTheMoon special episode in which Alexander Volchek and Max Grigoriev — an AI and machine-learning systems specialist who has worked at Google and Airbnb — agree on the concepts. The occasion was a recent argument of theirs: it turned out they had understood the same words differently. Max explains why you can't argue without agreeing on concepts: human language is imprecise by necessity, "not a bug but a feature," so he suggests "getting on the rails of formalism." For Alexander, AI is what films show: a superconsciousness, a huge independent construct — and he would prefer other words for today's technologies, so that they are not declared now a little program for the next symbol, now a threat to humanity.
Max's definition comes from Merriam-Webster: intelligence is the ability to learn, understand and deal with new or difficult situations. A calculator solves complex tasks but does not learn anything new and is therefore not AI; a system should be judged by its capacity for self-learning, up to zero-shot learning — solving a new problem on the first try. Alexander accepts the definition for this conversation, but what matters more to him is the ability to do right away what you have never done: every new president is a "junior." The Turing test, Max says, was passed five or six years ago, and he calls self-improvement the relevant criterion: many labs believe AGI will arrive together with a system that sets research tasks itself and builds new systems from the results.
Alexander calls AGI, ASI and "agents" marketing terms: in the definition of AGI it is unclear what "practically any cognitive task" means, and Sam Altman, Dario Amodei and Elon Musk manipulate these words. For him the new level would be a factory that reproduces itself: makes microprocessors, prints circuit boards, writes software and negotiates its own purchases. In reply Max tells how AI got to where it is: expert systems built on yes/no trees in the eighties, from the nineties neural networks that the US postal service used to sort letters with handwritten zip codes, and then three ingredients — algorithms, Internet data and computation on graphics cards, unlocked by NVIDIA's CUDA. There is already not enough data, he says, for the systems frontier labs are building.
Max's key concept is generalization: solving a problem whose solution is contained in the training data means using a cheat sheet, while assembling a solution from pieces of others means learning. He defines AGI as a system that solves problems of a class absent from its data and sees the first glimpses in the reportedly solved smoothness problem for the Navier–Stokes equations; superintelligence is what humanity would never have solved. Alexander replies that for billions of people cancer, laws, wages and a decision for each particular person matter more, that AI is held back by resources — give it ten billion dollars and it might solve cancer — and that models allowed to truly study people could get out of control.
The second half is about who chooses AI's tasks. Max explains why math comes easier to AI: the tasks are chosen by math geeks at frontier labs, and math fits into a couple of sentences of text. Alexander recalls OpenAI engineers' pet projects at DevDay and asks whether everyone will get their own AGI. Max's answer: capitalism will make AI cheap, as it once did sugar, but "some animals are more equal": subscriptions and hundreds of dollars a day on tokens are already stratifying people, even though free AI gives unreal possibilities. The finale covers 830 million Formula 1 fans, AI alignment as the topic of the next conversation, and automated science — from AI for biology to attempts to solve cancer with LLMs, on which, Max says, billions of dollars are already being spent.
The episode is useful as a dictionary for arguments about AI: it separates intelligence from skills, the Turing test from self-improvement, training data from generalization, AGI from superintelligence, and shows where the hosts' definitions converge and where they diverge: Max measures progress by problems absent from the data, Alexander by the benefit to a particular person. The figures and events — thirty million dollars spent on the Navier–Stokes equations, 830 million Formula 1 fans, a factory without people in China, when the Turing test was passed — are given as the hosts state them and need checking. The hosts deliberately left AI safety and alignment for the next episode.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 197 segments: 0 identified, 0 mixed, 197 probable, and 0 unresolved.
Read transcript on a separate page
Loading…