Интерпретируемость важнее улучшения метрик | Татьяна Гайнцева

Tatyana Gaintseva is an AI researcher and PhD candidate at Queen Mary University of London, researching the interpretability and behavioral control of generative models. She is a recipient of a DeepMind scholarship, a lecturer at Deep Learning School and Nebius Academy, a co-founder of Deep Learning School, and the author of the DLStories Telegram channel and the Deep Learning Stories podcast. Previously, she conducted research at Huawei and Philips, working on computer vision and medical AI. This episode focused less on a career in AI and more on the research mindset. Why do some specialists spend years improving model metrics, while others try to understand what's going on inside them? Why even bother researching the interpretability of neural networks? Is it possible to control the behavior of LLMs through the activations of individual layers? And why is it that sometimes the most interesting question in machine learning is not "how to make it better," but "why does it work at all?" We discussed Tatyana's journey from competitive mathematics and MIPT to her PhD in London, the creation of Deep Learning School, her work at Huawei and Philips, the transition from applied AI to fundamental research, and the current state of AI research. Viktor Kantor's Telegram channel: https://t.me/kantor_ai Tatyana Gaintseva's Telegram channel: https://t.me/dl_stories __________________________________________________________________________________________________ In less than a month, our ML System Design course with Valery Babushkin will start. Sign up now while there are still spots available! Website with the program and course pricing: https://mlinside.ru/system-design/?ut... Pre-registration form: https://forms.yandex.ru/u/6a0436bd902... __________________________________________________________________________________________________ 0:00 - Guest introduction 0:20 - Tatyana talks about herself 1:16 - How Tatyana got into Phystech 3:37 - About Tatyana's story and introduction to AI 9:30 - About the Google Collaboration "prototype" 11:08 - Continuation of the story about introduction to AI 13:06 - About Deep Learning School 14:31 - About working in AI 18:57 - About working at Huawei 22:47 - About the moral side of technology development 26:06 - About working at Phillips 32:22 - About getting a PhD 36:34 - About Tatiana's research #MLinsidePodcast #TatianaGaintseva #MachineLearning #DeepLearning #AIResearch #LLM #Interpretability #ActivationSteering #DataScience #DeepLearningSchool