Invited Talks

AI’s nextphase will be shaped by learning from real-world experience. In industry, however, such experience is embedded in proprietary workflows, physical experiments, and decisions with real consequences. Turning such experience into a reliable source of learning requires more than deploying capable models. In this talk, I will describe our efforts to build the industrial learning loop through four connected foundations. 1) Access to training. Frontier model development remains concentrated in a handful of AI companies. To enable others to train their own models, we released an open, end-to-end, frontier-grade pipeline—from cluster setup to inference, including code, weights, corpus, environments, and logs. 2) The hands to build. Industrial data often cannot leave the building, and domain experts should not have to become AI engineers. Our agentic R&D system is designed for in-house development: experts define tasks, constraints, and acceptance criteria in natural language, while agents coordinate experimentation, model post-training, and workflow development. In one deployment, 300 AI agents worked together to post-train a roughly 400-billion-parameter model in two months. 3) Readiness on both sides. General-purpose AI often falls short of industrial requirements, while industrial infrastructure is rarely designed to support iterative learning. To address these challenges, we developed specialized models—from quantum chemistry at the scale of 100,000 atoms to a DNA model 200 times faster than before—alongside autonomous laboratories for high-throughput experimentation with AI models in the loop. 4) Training grounds built on trust. Industrial environments cannot simply be scraped. We therefore designed institutional arrangements in which enterprises define problems and partners share risks, creating incentives to open the loop. Across six enterprise groups and nine joint programs, individual engagements began to form shared infrastructure. Together, these pieces set the industrial learning loop in motion on a materials problem in the semiconductor sector that had resisted progress for decades, delivering a result that met industrial standards and was accepted by domain experts.

Satinder Singh Google DeepMind & University of Michigan There is no such thing as reward-hacking, there is only bad reward design and good reward design. Indeed, Reinforcement Learning is *the* solution to alignment, not the problem. I will make the case that what we need are powerful AIs that have their own (human-designed) reward functions that care about long-term consequences of behaviors including instruction-following. I will also briefly lay out a possible research program for building these better AIngels of our future from our frontier models. This program will advance (1) designing good rewards that reflect the better angels of our own nature and (2) planning with local-world models—models that include us. *From joint position paper with Richard Lewis, University of Michigan

Scoring the performance of a machine learning model is a key choice we all make, whether it is the loss for training a model, a gap for understanding alignment, or a metric such as accuracy for predictive performance. This talk considers the score we use in adaptive experimental design for scientific discovery, where we use bandit type approaches to choose experiments. I show some recent theoretical and practical work on adaptive design and discuss the needs of cross disciplinary collaboration for AI in science, for example to empirically validate bandit designs in a genomics lab. On the other side, machine learning researchers may be tempted to generalise the idea of optimising a score to other aspects of work and life. I argue that it is problematic to convert the diversity of human experience into a single scalar score, and then optimising that score. Finally I discuss some preliminary results on what could be in between scoring and not scoring.

This talk reflects on the future of learning as generative artificial intelligence becomes both a powerful engine of epistemic production and of epistemic dispossession. The latter refers to the wide-scale removal of knowledge as a good created, enacted and held by human beings, challenging the mainstream view of AI as a promise of human knowledge abundance. After outlining five dimensions of dispossession realized through the extraction, severance, control, distortion and enclosure of knowledge, I explore the alternatives to epistemic dispossession, and the conditions under which it might be able to be resisted.

AI evaluation is finally shifting from aggregate performance metrics on specialised benchmarks to rich profiles of capabilities and propensities. Determining task demands and incentives, and comparing them against these profiles makes it possible to predict and explain performance and harm; this is the true objective of AI evaluation. Just as we are beginning to extend this to evaluations that are more human-centred and society-relevant, new challenges arise. In performance evaluations, we struggle to build verifiable, superhuman tasks of arbitrary difficulty. In safety evaluations, we can no longer build testing sandboxes that are realistic, contained and secure. In both cases, we face continual learning, with ever-evolving AI systems, and evaluation awareness, with deliberate sandbagging or judge manipulation. In this new world for AI evaluation, I propose a blend of pragmatism and foundational work. We must accept that AI systems will be unreliable, insecure, and misaligned, requiring evidence-based red lines and monitoring. We must also overhaul scaling laws and supervision, and explore new formalisms, such as relative 'supraintelligence', back in computational terms. And we must choose the right level and extent of AI evaluation automation to make AI oversight both scalable and controllable. AI development may have entered open-ended recursive self-improvement, but we cannot afford to let that happen with AI evaluation.