English Español Français Deutsch
← Back to the list
Systems and scaling Foundational figures Methodologically adjacent

Chelsea Finn

How language and learning systems adapt to a new task with very few examples — and what the theoretical structure of that adaptation tells you about what they have or haven't actually learned.

Finn's MAML (Model-Agnostic Meta-Learning, 2017, lead author) formalized the question of meta-learning at architectural level: train a model not to perform a specific task, but to be quickly fine-tunable to any task in a distribution it's seen related examples from. The framework predates the in-context-learning era by several years, but its theoretical vocabulary — task distributions, adaptation steps, the loss landscape of fine-tuning — is what modern LLM in-context-learning research either explicitly builds on or independently rediscovers. For ai100, the meta-learning lens applies directly: when an AI engine answers a probe query it's never seen before, the question of whether it "knows" the answer or is adapting on the fly is methodologically equivalent to the few-shot adaptation questions Finn's work formalized.

Worth following when
you want theoretical grounding for how LLMs handle queries they weren't specifically trained for, and want the meta-learning literature that worked out the relevant frameworks before LLMs arrived.
Topics
model-agnostic meta-learning (MAML); the theoretical structure of few-shot adaptation; the bridge between meta-learning and in-context learning in LLMs.
Key works
MAML: Model-Agnostic Meta-Learning (2017, lead author); long body of work on meta-learning and robotic learning; Stanford and Physical Intelligence publications on RL/robotics.