Samuel Colvin, CEO of Pydantic, predicts that OpenAI and Anthropic will soon create closed “intent databases.” They will store every interaction a developer has with AI coding tools—it’s useful, free, and unexportable. The main advantage: by clicking on any line of code, you can see the entire chain of AI reasoning behind it.
Samuel Colvin, CEO of Pydantic, predicts that OpenAI and Anthropic will soon create closed “intent databases.” They will store every interaction a developer has with AI coding tools—it’s useful, free, and unexportable. The main advantage: by clicking on any line of code, you can see the entire chain of AI reasoning behind it.
Pydantic, the company behind one of the most popular AI frameworks, works closely with leading model development labs and AI developers, including Anthropic and OpenAI. This gives CEO Samuel Colvin a first-hand look at the rapid evolution of models, agents, and coding tools. In an interview with Business Insider , he shared his vision for how tech giants' strategies might change to keep users firmly tied to their AI ecosystems.
According to Colvin, a year ago, Anthropic and OpenAI’s top priority was revenue, so they welcomed any use of their inference. But now, as both companies prepare for IPOs, profit margins have become paramount.
He notes that competing purely on the quality of models is the worst way to maintain high margins, as it requires enormous costs for training AI and minimizing the cost of generating answers for users. That is why companies are now actively looking for alternative ways to engage customers that are not related to the quality of the models themselves. Colvin is convinced that this is what led to the emergence of tools such as Claude Code, Codex, and other similar projects.
Samuel Colvin notes that the reasons for the deep discounts on tools like Codex and Claude Code are obvious. While users pay only about $200 per month for a subscription, the actual costs to companies for inference (generating model responses) for these queries can run into thousands of dollars. In this way, developers try to maximize their market share and attract as many users as possible.
However, according to Pydantic’s CEO, there is a much deeper strategy behind this move. When customers create huge databases, virtually entirely written by artificial intelligence, they find themselves at a point where a human is no longer physically capable of maintaining such code on their own. Colvin gives an example: if 20,000 lines of code were generated overnight using AI, a human would never be able to understand them for further maintenance — they would have to turn to the model again to fix errors.
Ultimately, with such huge amounts of generated code, corporate clients will be forced to continue using AI services from Anthropic and OpenAI. Colvin is convinced that once companies finally tie users to their ecosystem, they will most likely significantly increase the prices for their services.
According to Samuel Colvin, companies will soon expand the capabilities of their corporate subscriptions. They will offer customers not only code generation and the operation of AI agents, but also the preservation of the full history and logic of interaction between the user and the model during the process of writing programs.
Pydantic's CEO explains that as a result, companies will receive a unique database that will allow them to instantly determine the developer's original intention for any line of code. AI creators will argue that this is an extremely useful feature that will make the work of programming AI agents even more efficient.
However, Colvin predicts another side to this innovation: most likely, such an option will be provided for free, but the possibility of its export will be closed. Thus, the business will find itself ultimately dependent on the ecosystem of the AI provider it has chosen for its work.
Colvin explained in more detail how this might work. He suggests imagining a situation where a software bug or some unusual code behavior occurs. Typically, a developer—whether human or AI—leaves a comment explaining the logic behind that line. However, too many comments can make the text unreadable. Instead, the Pydantic CEO describes a much more efficient approach: the ability to simply click on a line of code and see the full history of a colleague’s interaction with the AI model as it was created. This includes the entire chain of reasoning of the model and the human input, providing a comprehensive explanation.
According to Colvin, this approach provides a much deeper understanding of the original intent behind the code base. This makes changes much safer and less risky, as the programmer knows exactly what the code authors were trying to achieve and can easily determine whether it is a bug or the intended logic of the program.
"I think this idea of essentially 'we store your trajectories and we give you some kind of database of your trajectories' is attractive and valuable. These two things are not necessarily the same thing, but in this case it would be really valuable, as well as attractive."
