Member-only story
You Can’t Fetch the Coffee if You’re Dead
AI-Human Alignment — is it hopeless?
The title comes from the Chapter ‘Overly Intelligent AI’ from Stuart Russell’s book ‘Human Compatible’, and was used to illustrate the seeming impossibility of prescribing safe goals to an AI.
I used it from 2020 on to motivate the fake newspaper article below. The article was presented in the last lecture of a second year BSc course I taught on The Mathematics of Deep Learning. My aim was to get my students to think beyond the maths we’d done, and to grapple with the enormity of what AI was becoming.
There is a lot of AI Doomerism around at the moment and while I don’t wholly subscribe, I thought I’d share my take on Russell’s observation. By the way — his book is a strong buy from me: It was relevant when it was published in 2019, and it’s even more relevant now.
In 2032 Ben E Volent as head of the UK’s leading tech firm NY Eve Future Technologies authorised the release of an AI robot product called Bean. Bean worked in large open plan office spaces, with the job of visiting each desk, taking a coffee order, and then making and delivering the coffee to the customer. Bean was trained using the usual techniques in AI — the performance index, or cost function, minimized in Bean’s neural network to give its weights and biases, involved such things as: the error…
