Скачать с ютуб Continuous Thought Machines, Absolute Zero, BLIP3-o, Gemini Diffusion & more | EP. 41 в хорошем качестве

Continuous Thought Machines, Absolute Zero, BLIP3-o, Gemini Diffusion & more | EP. 41

In this episode of Hidden Layers: Decoded, Ron Green, Dr. ZZ Si, and Michael Wharton explore the latest AI breakthroughs, including Sakana AI’s biologically-inspired “Continuous Thought Machines,” the self-taught coding model Absolute Zero, and Salesforce’s unified vision-language system BLIP3-o. They discuss the growing importance of reinforcement learning in a data-constrained world, Google’s diffusion-based language and video models, and Anthropic’s industry-leading interpretability efforts. The team also covers Apple’s AI missteps and a new study revealing why single, well-structured prompts outperform long chat sessions. Throughout, they reflect on alignment risks, emergent reasoning, and the changing shape of model development and training strategy. 00:00:04 - Introduction 00:00:42 - Sakana AI’s “Continuous Thought Machines” 00:08:23 - Absolute Zero Reasoner: Self-Taught AI Coders 00:16:53 – Why Reinforcement Learning Will Dominate the Future 00:20:14 - Salesforce BLIP3-o: Unified Vision-Language Model 00:25:21 - Anthropic Safety Research and Interpretability 00:29:00 - Claude 4 Release: Opus and Sonnet Upgrades 00:34:04 - Google’s Gemini Diffusion Language Model 00:35:36 - Single-Prompt vs Multi-Turn: Microsoft–Salesforce Study 00:38:32 - Apple’s AI Delays and Competitive Challenges 00:43:03 - Veo 3: Google’s Text-to-Video Breakthrough