#009 ML Сергей Марков. Русская GPT-3 и роботы-коллекторы от Сбера

Value of the Information & Strength of the Argument

The value lies in the insider perspective on how a major bank applies AI in real-world scenarios. The demonstration of the robot phone call is particularly informative, showing both the capabilities and limitations of current conversational systems. Markov argues convincingly that quality comes from meticulous engineering rather than just advanced algorithms, substantiating this with examples like handling mispronounced words and unnatural user interactions. His explanation of ruGPT-3 production is detailed and honest, including its size compared to the largest English models and the specific computational challenges. The argumentation is coherent, though largely anecdotal, relying on personal experience rather than systematic data or comparative studies.

Scientific Rigor, Source Quality, Title Accuracy

Markov does not cite external sources during the conversation, but the video description provides links to relevant materials, such as the Habr article about ruGPT-3 and the official SberDevices blog. His credibility as a researcher and manager strengthens the trustworthiness of his claims, although some statements, like the number of calls handled daily (850,000), are not independently verifiable. The title matches the content well, accurately reflecting the two main subjects. No critical inconsistencies were observed within the discussion.

The title accurately reflects the main topics discussed: the Russian GPT-3 model and Sber's debt-collection robots. Both are central to the conversation.

8/10

The information comes from a recognized expert in machine learning, head of experimental systems at SberDevices. He provides technical details that are plausible and consistent with known AI developments, but the podcast format and lack of peer-reviewed sources limit the verifiability. The claims about ruGPT-3 and the supercomputer are coherent and match external announcements.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 01:00Introduction of Sergey Markov and his educational background in game AI.
• 08:00Description of the robocaller system and live demo of a debt-collection call.
• 20:00Discussion on the technical challenges of phone call speech recognition and the importance of data diversity.
• 30:00Comparison of AI capabilities between Yandex and Sber, and public perception.
• 45:00Announcement of ruGPT-3 model, details on its size and training dataset.
• 55:00Technicalities of training large models on the Christari supercomputer and comparisons with OpenAI's architecture.
• 1:10:00Future plans for larger models and contributions to the AI community competition.

• Сбер выложил русскоязычную модель GPT-3 Large с 760 миллионами параметров в открытый доступ — Habr article linked in the description, announcing the release of the ruGPT-3 model.
• Сергей Марков - специалист по ИИ — The expert's personal blog, mentioned as his online presence.
• Блог подразделения SberDevices на Хабре — Blog of the SberDevices division on Habr, referenced for project details.
• Научно-популярный проект 22century.ru — A popular science project linked to Sergey Markov and colleagues.
• Выпуск подкаста на anchor.fm — Direct link to this podcast episode on Anchor.
• Подкаст на Google Подкастах — RSS feed for the podcast on Google Podcasts.
• Телеграм-канал «Стать специалистом по машинному обучению» — Telegram channel for learning ML, linked in the description.

• Habr article about ruGPT-3 (official Sberbank blog) — This official article confirms the model release and specifications as discussed in the podcast.
• SberDevices blog on Habr — The blog provides additional details on the technologies and products mentioned, aligning with the podcast content.

• None found — No contradicting sources were identified; the podcast's claims appear consistent with public information available at the time.

• anchor.fm
• music.yandex.ru