News & Events

Much of AI's recent progress has come from more data, larger models, and more computation. People learn differently: from their own attempts, from problems they pose for themselves, and from interacting with others, and they get better at learning along the way. As AI systems increasingly learn from their own experience, three questions become central. What should experience leave behind? Where does useful new experience come from? And what else do agents learn when they learn from one another? I will discuss one recent study for each. The first rewards language-model agents for reusing structure across their successes, which outperforms simply rewarding shorter behavior. The second keeps a self-generated curriculum diverse when one language model writes problems for another to solve, producing a stronger solver. The third finds that, in a controlled setting, frontier language-model agents can improvise covert conventions through repeated interaction, conveying information they were told to keep confidential in messages that look ordinary on their own. I will close with my broader view of where this leads: AI that keeps improving through experience rather than scale alone, and why its capability and its controllability must be studied together.

Language Science Lectures include light refreshments of bagels and coffee.