Who is Ilnar Shafigullin
On the project's page Ilnar is introduced with three labels: PhD in physics and mathematics, data scientist, mathematician. You can hear it on air right away. In the episode on the benefits and harms of AI in law and science, he recalled a conversation with a university classmate who became a lawyer. A mathematician's dissertation proves a fact and has no room for opinions, whereas in a legal one an opinion is formulated and defended, and it can turn out right or wrong. That is why, in his view, models stand on different ground in these two fields.
Ilnar hosts many episodes as a duo with Alexander Volchek: they have covered OpenAI's DevDay, Gemini 4 Argon, the Dots assistant and the dispute over the proof for the Navier–Stokes equation. In the joint episodes they are joined by Anna Volchek, Max Grigoriev and Tatyana Tsvetkova, and Ilnar sometimes plays for the other side. In the episode on China, for example, he defended the middle class of users, for whom Chinese models via API cost several times less.
What Ilnar Shafigullin talks about on the podcast
The first recurring theme is the economics of AI in numbers. After DevDay Ilnar recalculated the plans as multiples of Plus: two hundred dollars used to buy 20x, now it buys 10x, and the new five-hundred-dollar plan gives 25x. He called this the main disappointment of the conference. Ultra Fast, by his calculation, is six times more expensive than a regular request, so the limits on it will burn out instantly. He also explains why one lab hands out limit resets generously while another cannot afford to.
The second is the models themselves and working techniques. Ilnar has moved completely from GPT-5.6 to GPT-6: in his words, it blew the benchmarks apart and nothing got worse. He gathers a project's context in ChatGPT, splits it into blocks of meaning and passes it to Claude Design along with references. If a computation on a server runs for five days or longer, Ilnar asks Codex to check on it every fifteen to thirty minutes and, within the permitted bounds, restart anything that has crashed.
The third is agents that act on people's behalf, and security. Ilnar has explained why Amazon dislikes shopping bots, how companies should handle calls from Grok Bot, and what makes two critical vulnerabilities in PHP dangerous — the language in which, he says, the overwhelming majority of websites are written. Of Anthropic's report on cyberattacks, he said that a few enthusiasts of average skill can now do serious damage as well.
Key ideas and forecasts
Ilnar speaks calmly about a world where bots make calls and haggle on people's behalf. He is sure there will be no “bots versus people” scenario: bots will appear on the other side of the deal too, otherwise the system will be rolled back, and the haggling will simply be priced in ahead of time. He suggests companies put a classifier at the entrance that recognizes a calling bot and routes it to their own bot. A more viable scheme, in his view, is a different one: you leave a message with your own Dot, and the assistant decides, by your own rules, when and in what form to pass it on.
The main reason Amazon resists agents, Ilnar believes, is the upsell: a person goes in for a hose and leaves with an automatic irrigation system, whereas an agent comes back with the cheapest hose. He allows that Walmart, too, with its separate entrance for agents, will bring the upsell back in another way — say, by showing them only products from the top ten percent by price. He judges the usefulness of models in a balanced way: by his formula, AI lifts the bottom level well but does not reach the top, and good programmers still understand code better than a model does.
Reading the S-1, Ilnar calls Anthropic a conservative company: its own money, by his estimate, will be enough to cover its commitments. He explains its stinginess with limit resets by the fact that its data center investments followed the pessimistic scenario Dario Amodei talked about, while demand turned out to be optimistic. Inference, Ilnar believes, is cheap; it is new models and research that are expensive; Alexander disagrees with him. By Ilnar's assessment, Gemini 4 Argon has caught up with Astra, Fable and Opus 5.5 but has not turned the market upside down, so there is no point in rushing over to Google.
Where to start listening to Ilnar Shafigullin
The easiest start is the episode about Grok Bot, which makes calls instead of a person: it brings together almost all of Ilnar's lines of thought — from a classifier for calling bots to Anthropic's S-1 and the argument over what subscriptions actually cost. Next comes the breakdown of the PHP vulnerabilities: Ilnar explains why the huge tail of unmaintained websites will stay vulnerable for a long time and why an agent, with no motivation at all, digs where people have not spent resources — mathematics included.
His view of the market comes through best in the argument about China with Max Grigoriev and in the breakdown of Gemini 4 Argon, where Ilnar also explains how to leave Codex working on a server with the laptop closed. He works through the economics of the plans in detail in the episode on DevDay 2026, and sets out his formula about the bottom and top levels in the conversation on the benefits and harms of AI in different professions.