English Español Français Deutsch
← Back to the list
Retrieval-augmented generation Foundational figures Directly upstream of ai100's methodology

Yoav Shoham

What can be measured about the state of AI from outside any single company — and what retrieval-augmentation looks like when you don't need to retrain anything.

Shoham's career spans multiple eras of AI — multi-agent systems and game theory in the textbook he co-authored, then a long stretch chairing the Stanford AI Index, an annual measurement of the field that has become the closest thing the industry has to an official accounting. The "In-Context Retrieval-Augmented Language Models" paper (2023, senior author) sits inside the AI21 stack but reads as a method paper: retrieval-augmentation works on frozen LLMs with off-the-shelf retrievers, no joint training needed, which is exactly the regime that applies when you can't open up the model you're trying to improve. The two lines — measurement of the field and methods that work in the closed-model regime — meet at a position ai100 shares: you don't get to claim industry-level insight without admitting what you can and can't measure.

Worth following when
you want a senior AI figure who treats "measuring the state of the field" and "making methods that work without model access" as one continuous project.
Topics
annual industry measurement (Stanford AI Index); in-context retrieval-augmentation without joint training; multi-agent systems as a foundational subfield.
Key works
Multiagent Systems textbook (2009, with Leyton-Brown); In-Context Retrieval-Augmented Language Models (2023, senior author); Stanford AI Index annual report (chair).