Invited Talks

Recent advances in AI agents have been driven by imitation learning with reinforcement learning in the digital world, based on large scale generative models, yielding strong performance in many online tasks but limited capability in physical world settings. I argue for a shift toward AI agents grounded in world modeling, allowing them to understand the physical environment, to understand user intentions and social contexts, thereby enhancing their ability to perform complex tasks autonomously in the real world. World modeling encompasses the integration of multimodal perception, planning through reasoning for action and control, and memory to create a comprehensive understanding of the physical world. I argue that achieving advanced machine intelligence requires modeling both the physical world and the mental world, including latent variables such as intent, attention, and context. I outline key challenges toward building context-aware, interactive agents in the real world. This essential trajectory demands continued efforts to develop robust world models and embodied agents that can truly assist humans with real tasks in the real world.

How do we answer causal questions about sequence data such as text, career job sequences, or customer journeys? This talk will consider methods for estimating average treatment effects, conditional average treatment effects, and decompositions of differences across groups in average outcomes. It will consider both experimental and observational data.

Modern large language model pretraining is governed by complex heuristics — from cosine learning-rate decay to batch-size scheduling. Yet, a growing body of work suggests that an analytically simple quadratic model can accurately predict much of this large-scale optimization behavior. In this talk, I will argue that the quadratic model is not merely a convenient theoretical toy, but a useful lens for pretraining practice — both for compute efficiency and for serial runtime. We will begin with exact computations of critical batch size and time-dependent learning rates in linear systems, establishing a principled foundation. From there, we will see how the same analysis yields batch-size scaling laws (where we estimate batch-size exponents in LLMs) and motivates two pretraining improvements: SeeSaw, a scheduler that trades learning-rate decay for batch-size growth and matches loss at lower serial runtime; and Horizon-Free Pretraining, which shows how anytime schedules with weight averaging can match carefully tuned cosine decay without committing to a horizon in advance. We will close with lower bounds on the interaction between momentum and batch size, which suggest the quadratic model captures fundamental limits about what any first-order method can achieve. Taken together, these results make the case that quadratics deserve a more central place in how we think about pretraining.

Making effective medicines is challenging: more than 90 percent of drug candidates fail in pre-clinical research or clinical trials. A major contributor to this low success rate is the enormous space of biological and therapeutic possibilities. In the underlying biology of disease, there are thousands of different cell types and states, about 20,000 genes in our genome, more than 105 disease associated loci, and perhaps 1013 or more ways in which they could meaningfully combine. To make medicines targeting this biology, one could consider at least 1060 possible small molecules with medicine-like properties, approximately 2032 relevant antibodies to consider, billions of people, and about 10,000 different diseases. Now, however, we are at a major inflection point: we can collect large-scale data, at high-resolution, from human biology, and crucially, combine these large datasets with AI to be able to represent, reason and generate over these enormous spaces to yield testable predictions of missing or nonexistent information and iteratively improve our models. Although it is not possible to test every possibility in a lab, clinical trial, or even an entire population, with the scale of data it is currently possible to generate, we can use AI to bridge different layers of biology, determine the impact of combinations of genetic mutations or drug perturbations, predict disease progression, and generate therapeutic molecules de novo or through optimization. Key to the success of this approach is an integrated interplay between data and AI, or a “Lab in the Loop,” where experimental or clinical data are used to train models, the models are used to help predict and design the next set of experiments, and the process is iterated, at scale, both to yield key predictions in any specific project and improve the model for all projects. In this talk, I will describe how we built such a Lab in the Loop of experiments and AI in Genentech across our target discovery, drug discovery and drug development efforts to serve patients across therapeutic areas.

How can agents make safe autonomous decisions in complex, dynamic environments? While significant progress has been made in establishing post-training guardrails to enforce conversational compliance in generative models, these rigid constraints often prove brittle in open-world environments. I argue that achieving generalizable agentic safety requires Normative Alignment: a new paradigm that moves beyond passive harm avoidance to equip autonomous systems with Agentic Integrity. This approach provides agents with the structural capability to interpret, reason through, and dynamically apply abstract principles when literal instructions fail. Realizing this paradigm presents a triple challenge of capability, measurement, and governance. First, it requires a shift in model capability toward normative competence beyond generic reward maximization, moving toward the contextual reasoning needed to adjudicate complex trade-offs in non-verifiable domains. Second, it demands new metrics that move optimization targets beyond immediate preference satisfaction toward long-term human well-being. Third, it requires deliberative governance to ensure these systems avoid top-down paternalism by grounding alignment targets in pluralistic, representative societal input.

Given rapid advances in AI, how should researchers and developers shift how we allocate our time? What new skills should we build so that we’re not obsolete in the future? I argue that there will be plenty for us to work on, grounded in the “AI as normal technology” thesis, which holds that there are many bottlenecks between AI capability improvements and automation of tasks or jobs. The evidence suggests that AI is better seen as an augmentation than an automation technology. The balance of human effort will shift towards tasks that are less verifiable — from developing models to scaffolds, and from building towards evaluation and monitoring. Over the long term, as purely technical skills are devalued, both researchers and developers will have to adapt. In research, human effort will migrate from problem solving to question asking and conceptual progress; in industry, relational skills, domain knowledge, aesthetic and normative judgment will gain in importance.