Sander Dieleman, Research Scientist (Director) at Google DeepMind, explains why every word on his blog is handwritten. Asking people to read something you didn't write yourself feels disrespectful to him, so he keeps AI out of the writing. He still uses models, though, to test a turn of phrase, find the right metaphor, or work through dense research papers while he learns. In this clip, Sander covers: • Why publishing AI-written text feels disrespectful to readers • How he checks phrasing and metaphors with a model • How he used Gemini to question papers while writing his post on flow maps • Where he draws the line between AI as a writing tool and AI as the writer

Sander Dieleman, Research Scientist (Director) at Google DeepMind, explains how much modern diffusion models depend on guidance. Turn it off and the output falls apart, with problems in both the global structure and the fine detail. Sander argues that same dependence helps explain why diffusion models can run with far fewer parameters than large language models. In this clip, Sander covers: • What happens when you sample from a diffusion model without guidance • Why both global structure and fine-grained detail break down • How guidance lets diffusion models get away with being smaller • How diffusion model size compares to large language models

Sander Dieleman, Research Scientist (Director) at Google DeepMind, walks through a short history of depth in deep learning. Each breakthrough let networks go deeper, from two layers to five, then to 20, then to 100. Sander sees diffusion and autoregression as the next step in that same line of progress. In this clip, Sander covers: • How layer-wise pre-training took networks from two layers to five • Why ReLU and other non-saturating nonlinearities pushed depth to 10 to 20 layers • How residual connections made 100-layer networks possible • Why he frames diffusion and autoregression as the next step in that evolution

While much of the recent AI boom has focused on text, images, and other unstructured data, Léo Dreyfus-Schmidt says the enterprise demand he saw during nearly a decade at Dataiku was clear: structured data. Léo discusses: • Why structured data remains central to enterprise machine learning • The disconnect between where AI receives attention and where businesses have demand • What his experience at Dataiku revealed about real-world enterprise use cases • Why tabular data remains critical to business decision-making

What does it take to trust the output of an AI system? According to Léo Dreyfus-Schmidt, the better starting point may be not trusting it at all. Léo explains why AI systems need to be tested and verified rather than blindly trusted, including: • Why AI outputs should always be checked • The risks of relying on systems without sufficient verification • Why human oversight remains critical • How testing can help ensure AI behaves the way we expect

If you loved a movie, song, or book and then discovered it was created by AI, would you feel differently about it? Léo Dreyfus-Schmidt explores the question through a debate with his partner, revealing two very different perspectives on AI-generated creativity: • Whether the value of art comes from the experience or its creator • Why knowing something was AI-generated can change how we perceive it • The role of human intention in creativity • How generative AI is challenging our assumptions about what it means to create