Accelerating Frontier AI: The Compute Crisis is a Reinforcement Learning Problem

​DeepAI San Francisco

​Frontier AI has a compute problem. As models grow larger and training runs become increasingly expensive, the industry’s appetite for compute continues to outpace supply. Building ever-larger AI data centres is one answer, but what if we could make dramatically better use of the compute we already have?

​This event focuses on a very different approach to the problem: using reinforcement learning, the field built for sequential decision-making, to design more efficient infrastructure and to squeeze more from the compute we already have, at every layer of the stack.

​RL agents already design chip floorplans, generate and tune GPU kernels, control the cooling of gigawatt-scale AI factories, and are now starting to optimise the training process itself.

​But where does this go next?

​As frontier AI systems become increasingly autonomous, will they uncover entirely new opportunities for optimisation? How should recent thinking on continual learning influence the design of tomorrow’s training infrastructure? And if the compute crisis is a reinforcement learning problem, just how much of the stack can RL eventually optimise?

​These are the questions our panelists will debate.

​Join us at AWS Builder Loft for an evening of talks and debate with researchers and founders applying RL across the compute stack, from silicon to schedulers, followed by a panel debate, and drinks with the RL and frontier AI community. Following our London edition at Amazon's HQ, this is the second event in the DeepAI series.

​Agenda

• ​5:00 – 5:40 — Doors open, arrivals & refreshments

• ​5:40 – 6:00 — Guest talk I: Rick Gentry, InstaDeep

• ​6:00 – 6:15 — Guest talk II (speaker to be announced)

• ​6:15 – 6:30 — Recurvia: accelerating training with online hyper-parameter control (presentation & demo)

• ​6:30 – 7:00 — Panel discussion.

• ​7:00 – 8:00 — Social hour — refreshments provided by our hosts at AWS

​Who this is for

​Researchers and engineers in reinforcement learning, foundation model training, and AI compute infrastructure; and anyone whose GPU bill keeps them up at night.

​Venue

​The Loft (AWS Builder Loft), San Francisco. Refreshments kindly provided by Amazon Web Services.

​NOTE: All attendees MUST bring a valid government-issued ID (e.g. driver's license, passport card, passport) to present at check-in. This is mandatory with no exceptions.

​Pre-check-in on the AWS Builder Loft event space for smooth check-in: https://events.builder.aws.com/NAOVwB