Slides posted here.

Drawing from her work on both Claude and ChatGPT, she'll share how she deepened her view of post-training as a blend of technical precision and creative intuition. This perspective becomes increasingly important as AI interactions grow increasingly multimodal, multi-agentic, and collaborative. At the heart of the discussion lies an interesting question: can we teach models to be truly creative and how do we design good evals for this?

To probe this question she'll briefly discuss Reinforcement Learning from AI Feedback (RLAIF): how synthetic data accelerates iteration, how the "asymmetric verification" paradigm—where checking is easier than generating—allow for new research methods, and what these advances reveal about fostering creative intelligence while keeping models aligned with human values.

But the path to AGI is riddled with deep, unsolved challenges: brittle reasoning, drifting goals, shallow memory, and poor calibration under uncertainty. Real-world deployment quickly reveals how fragile today's agents truly are. Solving this isn't just about model improvements—it requires a rethinking of how we design, evaluate, and deploy intelligent systems - from rigorous evaluation metrics to tight user feedback loops for building systems that can reason, remember, and recover.

In this talk, Div Garg explores a human-inspired approach to agent design—drawing from his work at frontier agent research and product design in the real-world. From new agent evaluation standards, online reinforcement learning training methodologies to enabling agent-agent communication, this talk offers a glimpse into the emerging frontier: agents that don't just complete tasks, but coordinate, adapt, and evolve with users in the loop.

Speaker Bio: Div Garg is the founder and CEO of AGI, Inc, a new applied AI lab redefining AI-human interaction with the mission to bring AGI into everyday life. Div prev. founded MultiOn, the first AI agent startup developing agents that can interact with computers and assist with everyday tasks, funded by top Silicon Valley VCs. Div has spent his career at the intersection of AI, research, and startups and was previously a CS PhD at Stanford focused on RL before dropping out. His work spans across various high-impact areas, ranging from self-driving cars, robotics, computer control & Minecraft AI agents.

Speaker Bio: Eric Zelikman has studied how we can improve AI reasoning, often with inspiration from human reasoning, and helped develop a now-popular algorithm for language models to learn from their own reasoning (self-taught reasoners/STaR). These days, he works on reasoning and agents at xAI, with a focus on how they can interact with us.

Speaker Bio: Denny Zhou founded the Reasoning Team at Google Brain, now part of Google DeepMind. His group is renowned for pioneering chain‑of‑thought prompting and self‑consistency, and for developing the mathematical foundations of in‑context learning and chain‑of‑thought reasoning. The team also created core techniques that power Gemini's reasoning capabilities. Denny co-founded the Conference on Language Modeling (COLM) and served as General Chair for COLM 2024.

Through visual walkthroughs and code examples, we will reveal how these capabilities are reshaping AI's ability to reason about the world, and glimpse the future of models that are truly general problem-solvers. We will also consider the implications for alignment: how to encourage helpful, honest, and harmless behavior when models are increasingly agentic.

Speaker Bio: Hongyu Ren is a Member of Technical Staff at OpenAI, working on reasoning algorithms for large language models. He led the development of the o-mini series of models, and previously received his PhD degree in CS at Stanford.

We will focus on the methods and findings of On the Biology of a Large Language Model, with some additional excursions and speculations. We hope to shed light on important behaviors like hallucination, planning, reasoning, (un)faithfulness, and emergent capabilities, and close with some suggestions for further research.

Speaker Bio: Joshua Batson leads the circuits effort of the Anthropic mechanistic interpretability team. Before Anthropic, he worked on viral genomics and computational microscopy at the Chan Zuckerberg Biohub. His academic training is in pure mathematics.

These advancements should go hand in hand, yet we still lack models that can tell us which biological targets to drug in which patient subpopulations. In this talk I'll describe one particularly promising approach to this problem: large multimodal world models of patient biology. The two core ingredients to this approach are quite general: 1) collecting a large dataset that spans many scales and modalities, and 2) training multimodal transformers that learn to fuse those data streams in a way that allows nuanced simulations with a "world model". I will give an accessible overview of these components, and share our progress in applying them to cancer immunotherapy.

Speaker Bio: Eshed is a neuroscientist and ML researcher working to understand biological systems with AI. He completed his PhD in neuroscience at Stanford, where he constructed self-supervised neural networks that incorporate biologically-inspired constraints to explain the structure, function, and development of primate visual cortex. Eshed is currently an ML scientist at Noetik, an AI-native biotech startup focused on curing cancer. In his work he develops novel transformer model architectures and tasks that learn from a large multi-modal dataset of patient tumor biology, and applies those models to drug discovery.

We set the context by briefly discussing some preliminaries around diffusion models and how they are trained. We then cover the UNet-based network architecture that used to be the de facto choice for diffusion models. This helps us to motivate the introduction and rise of transformer-based architectures for diffusion.

We cover the fundamental blocks and the degrees of freedom one can ablate in the base architecture in different conditional settings. We then shift our focus to the different flavors of attention and other connected components that the community has been using in some of the SoTA open models for various use cases. We conclude by shedding light on some promising future directions around efficiency.

Speaker Bio: Sayak works on diffusion models at Hugging Face. His day-to-day includes contributing to the diffusers library, training and babysitting diffusion models, and working on applied ideas. He's interested in subject-driven generation, preference alignment, and evaluation of diffusion models. When he is not working, he can be found playing the guitar and binge-watching ICML tutorials and Suits.

Speaker Bio: Andrew is a Research Scientist in Meta’s GenAI team, focusing on media generation. Over the past few years, his team has focussed on publishing research papers that push the frontiers of video generative models, including Emu-Video and Movie Gen. Prior to working at Meta, Andrew completed his PhD at Oxford’s Visual Geometry Group (VGG) under the supervision of Professor Andrew Zisserman.