Posts

Showing posts from August, 2023

Dead rats and dopamine - a new publication

My new paper, coauthored with Yohan John , Dakota McCoy and Oliver Braganza , is out in Behavioral and Brain Sciences. " Dead rats, dopamine, performance metrics and peacock tails " is about the universal emergence of proxy failure. When you measure and incentivise performance by a single metric (a proxy), the proxy will always become a worse measure of performance than it was before you added the incentive. This effect is also known as Goodhart's Law  in economics, and by other names in other fields - but the same underlying process drives the effect across multiple domains. Our paper studies it in management, economics, biology, neuroscience and other areas. Recent concerns about AI alignment are closely related to this phenomenon. The paperclip problem is a good example - if a sufficiently clever AI is given a single goal, to produce as many paperclips as possible, it may eventually destroy all of humanity and take over the whole universe in its efforts to maximise o