Nice job by Cal debunking misconceptions about AI model capabilities. The segment highlights a few points I cover in my unpublished NoLM - Not Only Language Models blog post. Specifically the fact that Language Models on their own can't do much and need to be connected to data sources and other systems. Complex AI systems will be built with more specialized roles and leverage various components for their planning and execution. In the end though, models will require integration into existing systems. Those integrations need to be done by people, meaning humans are still in control of the AI-assisted system capabilities.
Later in the podcast, Cal takes a question about distributed webs of trust. I agree with Cal's point of using existing open standards like RSS for content consumption. It's the reason you often hear the phrase, "or wherever you get your podcasts". Assuming you have a program that can read an RSS feed, you can follow all types of content. On the topic of discovery, Cal makes the suggestion of using distributed webs of trust. Using domain names and linking as ways of discovering content. While blogrolls were not directly called out, it's one of the benefits a curated set of links provides.