Distinguished lecture: "Instructing Machines Effectively"

Abstract: "Turing considered instructing machines by programming, but also envisaged 'child' machines that could be educated by learning. Today, we have very sophisticated programming languages and very powerful machine learning algorithms, but can we really instruct machines in an effective way? In this talk I claim that we need better prior alignment between machines and humans for machines to do what humans really want them to do with as little human effort as possible.

First, I'll illustrate examples and results from machine teaching. In particular, I'll present a new teaching framework based on minimising the teaching size (the bits of the teaching message) rather than the classical teaching dimension (the number of examples). I'll show the somewhat surprising result that, in Turing-complete languages, when using strongly aligned priors between teacher and learner, the size of the examples is usually smaller than the size of the concept to teach.

This tells us much about the way humans should teach machines. But this also tells us much about the way machines should teach humans, what is commonly referred to as explainable AI.

Second, I'll argue that the shift from teaching dimension to teaching size reconnects the notions of compression and communication, and the primitive view of language models, as originally introduced by Shannon. Nowadays, large language models have distilled so much about human priors that they can be easily queried with natural language prompts combining a mixture of textual hints and partial examples. The expected teaching size for a distribution of concepts presents itself as a powerful instrument to understand how language models can be effectively instructed to do a diversity of tasks. With this understanding, 'prompting' can properly become a distinctively new paradigm for instructing machines effectively, yet deeply intertwined with programming, learning and teaching."