Vasily Ryazanov, candidate of physical and mathematical sciences and head of the Data Science team at Twinby, sat down with journalist Alexander Shulepov to separate hype from practical value in neural networks. The conversation covers the technology’s long history, how contemporary language models operate, real business deployments, and the persistent need for human expertise.
From family history to Data Science: why neural networks entered his life long before ChatGPT
Ryazanov explains that his path began in the 1970s when his father worked on artificial intelligence and pattern recognition under academician Zhuravlev at the Academy of Sciences after finishing graduate studies around 1977. Early networks processed tables of only hundreds or thousands of rows on computers that ran for hours. By the 2000s the workflow had become clearer: load tabular data, train an algorithm, and obtain results, often written in C++.
He stresses that neural networks are not 10–20 years old but approximately 80 years old. Interest has risen and fallen in waves, with renewed attention around 2014–2015 driven by advances in image recognition, face detection, and projects such as Google DeepDream.
How TikTok became a professional platform
Ryazanov began posting on TikTok in 2019. Initial experiments with monowheel videos and song translations did not stick; a friend’s suggestion to discuss remote work and Data Science produced his first successful clip. The platform’s recommendation engine, which analyzes watch time, likes, on-screen content, speech, and music, quickly matched his material with an interested audience.
Skills and education required
For engineers and data scientists, Ryazanov insists on a solid foundation in probability theory, linear algebra, statistics, programming, and data handling. Entrepreneurs and marketers need only a clear understanding of capabilities, limitations, and data-handling rules. His own education at MIPT remains essential because much of his current work involves code and model training rather than simple prompting.
Distinguishing artificial intelligence, neural networks, and ChatGPT
Artificial intelligence is the broad goal of automating tasks associated with human thinking. Neural networks are one mathematical approach inspired by biological neurons. ChatGPT is a conversational system built on the Generative Pre-trained Transformer architecture; it predicts the next token while respecting dialogue history, safety constraints, and other controls.
Training begins with enormous datasets and GPU clusters to create a base model, followed by domain-specific fine-tuning or connection to corporate data. The cost of training frontier-scale models is dominated by GPUs, electricity, cooling, and infrastructure.
Prompts, Claude, and paid tiers
A prompt is simply a textual instruction that may assign a role, state goals, specify output format, and supply source material. Ryazanov prefers natural conversation with models, often dictating thoughts via keyboard while walking. He uses Claude most frequently because it maintains long context well and produces precise phrasing; ChatGPT remains useful for its multimodal features. Paid plans mainly remove rate limits and provide priority access during peak load.
Hallucinations, confidentiality, and the “AI psychologist” trap
Common beginner mistakes include uploading sensitive documents and placing excessive trust in fluent answers. Models can invent films, misattribute works, or cite nonexistent papers. Ryazanov warns against sending passports, medical records, trade secrets, or proprietary code to external services. When the cost of error is high—in health, finance, law, or safety—every output must be verified against primary sources and domain experts.
Although models can help structure reflections or prepare questions for a therapist or physician, they are not substitutes for professional care. The convenience of receiving a personalized summary in minutes makes the interface psychologically compelling, increasing the risk of over-reliance.
