Abstract: This talk will explore how scaling has driven progress in artificial intelligence over the past five years. In the first paradigm of scaling, the field achieved significant advancements by training large language models with increased compute and larger datasets. This approach led to the emergence of systems like ChatGPT and other AI chat engines, which demonstrated surprising versatility and general-purpose capabilities. With the introduction of OpenAI's o1, we are entering a new paradigm that extends scaling beyond training-time compute to test-time compute. These new models, trained through reinforcement learning on chain-of-thought reasoning, exhibit the ability to "think harder" for more complex tasks, enabling them to solve competition-level math and programming problems. The talk will conclude with reflections on shifts in AI research culture and speculations on future directions for the field.

Event is free and open to the public but registration is suggested.

Event co-hosted by Princeton AI Alignment (PAIA) and Princeton Language and Intelligence (PLI).