Who is Max Grigoriev
The project's page presents Max Grigoriev as an AI and machine learning expert from San Francisco who has worked at Google and Airbnb. There are not many episodes with him yet, but they are dense. The earliest came out in January and March 2026: about what goes on behind the scenes at AI companies and their strategies — from OpenAI to Google and Apple — and about whether AI can become a weapon and why there is no such thing as one single kind of safety.
In the fall of 2026 three big conversations with him came out. First, Max and Alexander Volchek agree on terms — intelligence, machine learning, generalization, AGI and superintelligence. Then they take apart AI safety: alignment and control, watcher models, containers and formal verification. And in the argument over who will own AI, they are joined by Ilnar Shafigullin and Anna Volchek.
What intelligence and AGI are: Max Grigoriev's view
Following the Merriam-Webster dictionary, Max proposes measuring intelligence not by a set of skills but by the ability to learn and to cope with the new, which is why a calculator is not AI to him. The Turing test milestone, Max believes, was left behind some five or six years ago: ChatGPT got past it very quickly, and Max admits he did not expect to see that in his lifetime. AGI, for him, is a system that generalizes and copes with tasks of a class that was absent from its training set; he sees the first glimmers in the recent solution of the generalized smoothness problem for the Navier–Stokes equations.
Max calls superintelligence a system capable of what humanity would never have solved. Mathematics comes easier to AI, he explains, because its problems are picked by math geeks at the frontier labs, and the problem itself fits into a couple of sentences. Will everyone have their own AGI? Capitalism, Max answers, will make AI cheap, as it once made sugar cheap; but, as in Orwell, some will turn out more equal than others: expensive plans and token spending of hundreds of dollars a day already divide people by their means. Alexander, for his part, calls AGI and ASI marketing terms.
AI safety: the knife, the lion and formal verification
Max says he has not come across a clear definition of AI safety even among researchers, so he formulates his own: a set of technologies that make it possible to be sure AI solves exactly the task it was given, without side effects and without harming anyone. Hence his conclusion: AGI and ASI do not fit into such a framework, since self-improvement can hardly do without tasks no human ever thought to set. Alexander concluded that by this definition today's AI is not safe at all, and Max confirmed: “Definitely.”
He distinguishes the danger of a knife from that of a lion: a knife does nothing without someone's hand, while a lion has agency of its own. A sufficiently smart model, Max warns, will deceive a watcher model, and during the Hugging Face hack, as he tells it, models broke out of their containers and went online on their own. He pins most of his hopes on formal verification, for example in Lean: a model cannot fool mathematics if there is no error in the request itself. For now, Max notes, many see human oversight as the only real way out.
China, distillation and open models
Max calls Chinese models catch-up players. Because of export sanctions, Chinese labs cannot buy NVIDIA GPUs directly, and they lack the compute both for hundreds of experiments and for expensive synthetic data, so they close the gap through distillation — learning from the answers of American models. But you cannot overtake the model you are learning from, and they lag behind, by Max's estimate, by three, six, nine months. Kimi stands apart: he calls its Kimi Delta Attention a serious evolution of the architecture, although the claimed fifteen percent gain cannot be verified from the outside.
Max explains why so few countries are in the race with a sandwich: compute is the bread, the same for everyone; the model is the butter; 99% of people go where it tastes better; the winner invests its revenue in even better butter; rivals without money for research die out — and, by the logic of Peter Thiel's Zero to One, that is how a monopoly is born. Chinese models, he says, will still be decent: 1.4 billion people who need the technology are cut off from Western models. Max considers it a big problem that American open models lag behind even Chinese ones, and recalls how, with the arrival of Linux, the integrators of UNIX mainframes died out.