Сейчас не использовать ИИ-инструменты – это проигрыш автоматически | Подкаст Андрей Жогов

Summary

The podcast follows ML engineer and educator Andrey Zhogov’s path from mathematics competitions to work on code models at Sber. It also discusses AI-assisted programming, model evaluation, teaching, and career development.

Career and learning path

• Zhogov became interested in mathematics through olympiad preparation in school, then chose a practical software-engineering track at MIPT.
• He initially expected machine learning to involve more neural networks and elegant mathematics. The early focus on simpler models was less exciting, but SVMs appealed to him for their mathematical ideas. An internship at MTS helped convince him that he wanted to pursue ML.
• After working on analytics and HR-related modeling at MTS, he moved to Sber’s research team. He describes not getting through his first interview, then applying again about six months later and receiving an offer.
• He says university mathematics—especially calculus and linear algebra—remains useful as a foundation for understanding models and preparing for interviews, even when it is not used directly every day. He would have preferred more clearly separated study tracks rather than requiring students to take unrelated areas such as mobile or frontend development.

Evaluating code models

At Sber, Zhogov has worked on evaluating code models and generating synthetic data. He describes evaluation as both a measurement problem and an infrastructure challenge:

• Code-model benchmarks need an environment where an agent can inspect a repository, edit files, and run tests.
• A benchmark typically consists of tasks, an execution environment, and a metric. One simple metric is the proportion of tasks an agent completes successfully, determined by tests it cannot see.
• Running many experiments quickly requires scalable infrastructure. Zhogov describes work involving Kubernetes and coordination between GPU and CPU systems.
• Benchmark results can be misleading if benchmark tasks leak into training data. He says detecting this requires a “decontamination pipeline” that checks for similarities between training examples and benchmark tasks.
• He also gives an example of a training dataset that allegedly included many copied benchmark tasks, warning that high reported scores are not necessarily evidence of genuine general capability.
• Evaluation is difficult to keep current: new models, benchmarks, and methods appear rapidly. A model’s mistakes or risky tool use do not always mean it is incapable; sometimes experimentation helps it find a better solution.

Using AI tools at work

Zhogov’s central point is that avoiding AI tools altogether is now a disadvantage because they can save substantial time—but their value depends on using and checking them well.

• AI is useful for implementation work, such as completing a feature with a precise specification or modifying a well-defined function.
• It is less reliable for architecture and design, especially in complex distributed systems where requirements and dependencies may be unclear.
• Agents can alter or delete files, make incorrect assumptions, or produce code that looks plausible but misses the intended goal. Reviewing every action matters; rushing through approvals can let mistakes slip by.
• AI can be particularly helpful after a person has already worked out the system’s requirements and architecture. A clear prompt can then speed up implementation.

Practical guidance for using AI without becoming dependent on it:

• Understand the input data, the intended output, and what the system is supposed to do before delegating implementation.
• Use AI for coding where the task is bounded and clearly described, rather than treating it as a substitute for design judgment.
• Read and check the generated code, even if only at a high level; remove unnecessary checks, assumptions, or clutter.
• Be able to explain the result and adapt it when a problem changes slightly. If someone can answer only the exact question they prompted the model with, that may indicate they have not understood the work.
• Watch for signs of unchecked generation, such as boilerplate code, excessive try/except blocks, unexplained comments, or code that the user cannot justify.

Teaching, NLP, and studying

Zhogov says he was drawn to teaching by the combination of sharing substantive ideas and helping learners feel motivated. In his view, teacher engagement can turn a clear lecture into an excellent one, but it should be supported by good explanations, examples, and practice.

• He recommends combining theory with applied work: for example, implementing a model or training loop, then using libraries such as PyTorch or Hugging Face, and connecting the material to practical applications.
• Analogies can help build intuition, but should complement the underlying concepts and mathematics. Working through a derivation with students—and investigating questions together when the answer is not immediately obvious—can deepen understanding.
• He sees NLP knowledge as valuable for both research and engineering. Developers building systems such as chatbots need enough understanding of how models and memory work to anticipate issues such as growing context and declining response quality.
• He sees a useful NLP course as one that combines theory with current practice, rather than separating foundational material from topics such as model inference and optimization. He expects courses to need updates as the field changes.
• In discussing likely directions for LLM development, he points to agent environments and tools, as well as reinforcement learning. He also notes growing research interest in smaller models that can perform well on particular tasks.

Advice for learning ML or NLP:

• Use a structured course or learning plan, particularly when rebuilding forgotten mathematical foundations; an AI chatbot alone may confidently give incorrect explanations.
• Avoid delaying all applied ML until after years of studying mathematics. Refresh the necessary fundamentals while beginning practical work, adjusting the depth to your goals and motivation.
• Seek interaction with instructors, peers, or a learning community. Live sessions and checkpoints make it easier to ask questions and verify that the material has been understood.
• Test understanding by building something yourself. For example, to check whether you understand BERT, implement a basic version and examine its inputs, attention, masks, and outputs.
• Use both conceptual study and practical exercises, so that learning does not remain at the level of memorized explanations.

Career advice and work habits

• For an entry-level job search, Zhogov recommends practical projects that reach at least a working MVP, including personal projects, hackathons, and competitions. Be prepared to explain your own code: an interviewer may inspect a repository and ask why it was written in a particular way.
• Articles or other published work can strengthen a profile, though he says they are not always required.
• He sees opportunities in both AI research and applied AI engineering. In his view, developers who adapt and use AI tools competently are more likely to remain valuable; he does not expect AI to replace everyone.
• He advises choosing work according to personal interests while remembering that career choices can change. Employment can provide experience and income while leaving room to explore a personal project.
• To maintain energy, he recommends changing activities and making room for physical movement, relationships, and rest. Working on interesting tasks for long hours without a broader life can still lead to burnout.
• Finally, he advises following substantive technical developments rather than getting absorbed in sensational headlines and speculation about AI.

Speakers and sources featured

• Viktor Kantor — host of the MLinside podcast.
• Andrey Zhogov — guest; ML engineer at Sber and educator at MIPT and MLinside.
• Sources and organizations mentioned in the discussion: Sber and its GigaCode work; MTS; MIPT; MLinside; machine-learning benchmarks and training datasets; research-paper reviews and arXiv articles; ChatGPT and other AI coding tools.

No additional interview speakers are featured in the subtitles.