The Role of Multi-Agent Learning in the Path to Powerful AI

When I started my PhD, I had that exact same intuition that indeed the interaction of intelligent agents is what has driven intelligence. It now turns out looking back that simply training supervised models, self-sized models, large-scale language models on large amounts of human data is a faster way of bringing agents to current or proximate levels of human abilities in terms of simple reasoning tasks. Now that doesn't mean that these methods will also allow us to radically surpass human abilities. So I think there's a bit nuanced about getting up to human levels, which I think current systems can do and it's an open question how much further we can push it without going to multi-agent learning. And also