Incredible podcast with @demishassabis the CEO of Google DeepMind with @lexfridman Some raw human notes on my favorite parts - 50% change of AGI by 2030. Jagged intelligence does not work. Consistency of intelligence across the board is needed (creative, inventive capabilities). - Gemini’s VEO3 video model is incredibly good at physics and realism (lighting, materials, liquids) despite having no form in the real world. It’s understood physics through observation and this hints at a deeper understanding of our world. - AlphaEvolve by DeepMind evolves algorithms. AI/LLMs collaborate with evolutionary algorithms (biology inspired optimization: iteratively evolving populations toward higher fitness via operations mimicking natural selection) - Taste (What's the right experiment or question) and judgement will be extremely hard to nail with AI. Systems can not do this today. - Noam Shazeer (Transformers), David Silver (AlphaGO), Demis is pretty clear he thinks Google has the best AI talent. “It’s harder to invent things vs fast follow”. Most original AI breakthroughs have been via Google. - He does not think we are running out of data to train. - Fusion and Solar are the key energy sources for AI. Batteries and transmission are the core issues. - Seems very unlikely text will be the user interface with multiple modality AI. Expecting AI generated interfaces tuned to how each person’s brain works. - Benchmark results are hard since as you optimize for one area another falters, it's a very difficult balancing act. End users are the ultimate arbiter (Lex said this last part but I liked it) - Demis says Meta is not at the frontier but is behind so overpaying for talent to catch up. - AGI is the most important technology that will ever be invented. - People who embrace AI and become one with them will become super humanely productive. - Number one thing is to create more abundance of resources so this isn’t a zero sum situation.