There are so many nuggets of wisdom to share in this episode of Intentional AI with guest Cal Newport.

About how we’re giving up our ability to think and form our own opinions (very pro-fascism):

So it’s really worrisome how quickly people lose their ability to do any thought. And then it’s full cognitive surrender and [with] anything in their life, even basic decisions, even non-professional, non-academic decisions. “Oh, man, I gotta have the chat bot do it.” I’m very worried about cognitive surrender.

About how AI companies keep getting more specialized at verifiable tasks, not actually what they’re selling as “general broad knowledge”:

If you zoom in and say, what really happened in the last year, you’ll notice that almost everything specific that OpenAI or Anthropic have talked about is in one of 3 very related domains. Computer programming, cybersecurity or mathematics. That’s everything they’ve been talking about the last year. Why? Because those are the domains where they have the highly structured data that are very well suited to reinforcement learning train, post-train these LLMs. So they’ve been putting a lot of effort to the areas where they can actually make these models better. And as a side effect, by the way, the models are getting worse at the other things that they used to tout, they’re getting worse at writing. They’re getting worse at expression because when you super train on one thing, you get reduced capabilities in other things. So you could also see this last year as a cautionary tale, if you’re a potential investor in an IPO or something, is why are you talking so much about esoteric math problems? That’s the market you’re going for? Look, I’m a mathematician. We don’t have any money.

About how AI companies keep pivoting as they find the tech does not work in the ways they initially promised:

At the beginning of 2025, the big thing that the major AI companies were saying is this will be the year when the type of coding agents we see programmers use are going to go ubiquitous and in all jobs you’re going to use them. It will be the year of the agent, and everyone is going to have whatever knowledge work job you have is going to have this personalized agent that’s going to do so much of your work that it’s probably going to lead to a lot of job loss, but like ultimately you’ll be more productive. And it just nothing happened. And by the end of that year, they turned and said, no, no, no, actually, actually what we meant was computer programming agents are going to get better. And it’s because not for lack of trying, it’s just difficult. They’re like, actually, it’s hard to build agents that don’t do computer programming.

Cal’s tip was to try to understand what “AI” is. The industry intentionally want you to think it’s a single thing (it’s not):

I don’t think people understand what these tools are underneath it and how they actually work. And because of that, they give them sort of magical properties and then they begin using them in inappropriate ways or thinking they’re going to get uses out of them that they’re not.

About how AI companies just assume this is what’s happening, rather than making things people want:

Talk to me about use cases. I can automate that or I can make this, you know, lower-lift work. Great. [But instead AI companies say] “We’re inventing the future. maybe it’ll even kill us all or take all of our jobs, but what can we do? The future is the future.” It’s astonishing rhetoric… where [they’re] like, I’m not even trying to convince you it’s doing something useful.