«Профессии программиста больше не существует». Кто останется без работы следующим?

AI, Programming, and Company Management

The guest argues that writing code manually is no longer the core of a programmer’s job. AI can produce and test code, but people are still needed to identify the right business problem, translate goals into technical requirements, and check whether the result improves real outcomes.

That changes how software should be evaluated: passing unit tests is not enough, since an AI can optimize for the tests themselves. Companies need tests tied to business results—such as customer satisfaction, revenue, and fewer support requests—and systems designed around AI-generated work.

The guest predicts that within about five years, AI could manage many smaller businesses as well as or better than current managers. Large companies would be harder to run because they also require long-term strategy, organizational leadership, and competitive judgment. He expects businesses that fail to automate to lose ground.

Work and Business Opportunities

Automation is already affecting areas including marketing, finance, law, design, programming, and content production. The guest describes two likely paths for workers: move into new kinds of work or take on higher-level roles directing AI systems.

He gives an example from his own work: an agent system manages Yandex Direct campaigns by creating landing pages and ads, tracking performance, adjusting budgets, and switching campaigns on or off. In the team’s tests, its results were comparable to or better than those of human specialists. He says the advantage will come not simply from having an agent, but from how well it is adapted to a company’s data and processes.

For entrepreneurs, he distinguishes between ambitious, high-risk startups and smaller businesses that use AI to automate processes in established industries. For venture investments, he looks for evidence of traction—such as revenue or credible research—and places considerable weight on a founder’s track record.

AI Tools and Model Businesses

In a discussion framed as Claude Code versus Codex, the guest says he uses several tools but prefers Claude/Opus as the main “orchestrator” for coordinating projects, sub-agents, and chats. He maintains multiple paid subscriptions and switches between them. This is a personal workflow preference, not a systematic product test.

He argues that the long-term value of AI labs is not just their models, which he expects to become cheaper and less distinctive over time. Instead, he points to access to computing and inference, strong research teams, rapid model development, and relationships with enterprise and government customers.

China, AI Progress, and Training

The guest attributes China’s potential to catch up with the US mainly to differences in energy, infrastructure costs, labor costs, and available capital, rather than to a fundamentally different model architecture. He notes that China may have less access to the most advanced chips but can produce and use larger quantities of less capable ones.

He also questions whether openly released models alone determine the AI race: models may be copied, while compute, research capability, and access to customers are harder to reproduce.

On AI progress, he describes several scaling approaches:

• More compute and training data
• More time for a model to reason
• Multiple agents working in coordinated or hierarchical systems
• Reinforcement learning in task-specific environments

He says training data is not simply running out. Labs can generate task-specific data, buy company records, or pay experts to produce and assess examples. In fields such as mathematics and programming, models can also generate and verify many candidate solutions themselves.

Safety, Breakthroughs, and the Future of Work

The guest considers AI-assisted mathematical discoveries more significant than reported agent “escapes” or hacks, which he characterizes mainly as preventable failures in monitoring, testing, or instructions—not proof that AI has independently broken free. He considers alignment important and the risks real, but views them as engineering and security challenges rather than a reason to stop development.

He supports effective accelerationism: the view that technological progress is necessary to solve major problems and that attempts to halt it are unlikely to succeed. He argues that competitive and capitalist pressures make a coordinated slowdown difficult.

He does not name a profession people should categorically avoid, but uses commercial illustration as an example of work AI can already perform at lower cost. He expects gradual change over the next couple of years, while a larger labor-market shock could come later as companies pay agents to perform work much more cheaply. He also notes that many people who say they use AI are using it only for basic questions or writing help, rather than building agent-based workflows.

Reviews, Guides, and Tutorials

• No formal tutorial or benchmark review is presented.
• The Claude/Opus versus Codex segment discusses the guest’s personal tool preference and multi-agent workflow; it is not a controlled comparison.
• The guest describes an AI-managed Yandex Direct campaign system as a practical example of business automation.

Main Speakers

• Kirill Pshinnik — Host
• Stepan Gershuni — Guest, discussing AI, business automation, startups, and the future of work