This afternoon I watched a great conversation with Dwarkesh Patel talking about key shortcomings in current large language models.
He points out that while the current crop of AI models are very powerful they still lack a fundamental capability that’s required to make them of real value in many contexts: the ability to learn via on the job training. That’s something that’s fundamental to human teams but it’s completely lacking in the current AI models.
Until there’s some breakthrough that solves this weakness, Patel feels the application of AI to many real world scenarios will be necessarily limited.
Here’s a Substack post on this topic that Patel wrote a few weeks ago: https://www.dwarkesh.com/p/timelines-june-2025
Many people are very optimistic about the potential for humanoid robots in the home. But this current lack of a learning capability is such a fundamental gap that it’s hard to imagine anyone wanting a robot that can’t take instruction and learn!
Maybe this capability just isn’t that important for self-driving cars — but for many other applications learning is totally fundamental and the lack of learning is a show stopper.