#033 ML Анатолий Старостин. Танцующие скелетики, генеративное искусство и философия ИИ

Value of the Information & Strength of the Argument

The value of the information lies in the insider perspective of someone working in AI at a major tech company, offering insights into practical limitations of generative models. The argumentation is mainly based on personal experience and reasoning rather than empirical evidence. The host and guest engage in a thoughtful dialogue about creativity and technology, but the claims are not systematically justified. The discussion of music generation methods is informative, with clear distinctions between raw audio, MIDI, and sample-based approaches. However, the philosophical parts remain subjective, and the lack of concrete data or hardware benchmarks slightly weakens the overall rigor.

Scientific Rigor, Source Quality, Title Accuracy

The episode mentions several specific sources: the OpenAI Jukebox project, the arXiv paper ‘Zero-Shot Text-to-Image Generation’ (DALL-E), and other podcasts. The description provides relevant links, which supports credibility. However, the speakers do not critically evaluate these sources in detail; they reference them as examples. The title aligns well with the content. No comments are provided for analysis, so no audience feedback is assessed.

The title accurately reflects the content: the episode covers generative art, AI philosophy, and explicitly mentions 'dancing skeletons' as an example of generative imagery.

7/10

The discussion is conducted by knowledgeable practitioners in the AI/ML field (host Mikhail and guest Anatoly Starostin, who leads technology development at Yandex Music and Kinopoisk). They reference specific technologies and papers, but the content is largely subjective perspective and anecdotal rather than systematically sourced. The quality is decent for a podcast but not rigorous scientific analysis.

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

• 02:00Introduction and context about the podcast and guest's background.
• 05:00Discussion on whether AI-generated works can be considered art, referencing social consensus.
• 10:00Explanation of GANs for image generation and their limitations in originality.
• 15:00Comparison of visual and auditory perception, leading to music generation challenges.
• 20:00Detailed discussion of raw audio generation with Jukebox and quality issues.
• 25:00MIDI-based music generation and the lack of global structure awareness.
• 30:00Sample-based assembly approach and its applicability to electronic music.
• 35:00Reflection on creativity, natural selection, and the democratization of artistic production.
• 40:00Future prospects for video generation and the need for better world models.
• 45:00Discussion on multimodal datasets and pathways to stronger AI.

• OpenAI Jukebox Blog — Referenced as an example of raw audio generation with low-quality output.
• Zero-Shot Text-to-Image Generation (arXiv) — Mentioned as a paper on DALL-E, likely discussed in relation to image generation.
• Podcast NII (Unidentified Artificial Intelligence) — Anatoly's own podcast, linked as a related resource.
• Podcast NII with Boris Yangel — Earlier episode mentioned as a favorite discussion on DALL-E.
• Anatoly Starostin's music — Linked as his musical work.
• Robert Downey Jr. AI series — Referenced as a documentary about AI.

• anchor.fm
• anchor.fm
• music.yandex.ru
• podcasts.apple.com
• podcasts.google.com
• t.me
• t.me