The report hammers home that we remain in a period of rapid improvement in AI and, just as importantly, ever expanding scope of what AI can do. The apparent triviality of some past breakthroughs in AI - identifying cats in videos or beating 1970s video games
- is long gone. Ian and Nathan highlight important new results in protein folding, molecule synthesis and warehouse automation.
Another striking conclusion is the centralisation of cutting edge AI in a small number of organisations - and the growing structural barriers to reversing that. DeepMind’s Starcraft model, AlphaStar
, took $26m in computational resource to train, while remuneration for top AI talent is now close to $1m annually. As Ian and Nathan note, this makes the question of AI governance increasingly important - but there are not yet any successful models
(and, as the public attitudes section makes clear, no consensus on what good looks like). This is troubling, and an important space to keep an eye on.