Yes, and I found Derek Thompson and Ezra Klein to be very credible on Lex Fridman:

Ezra Klein and Derek Thompson: Politics, Trump, AOC, Elon & DOGE - https://www.youtube.com/watch?v=DTPSeeKokdo

I like when the right and left can actually talk to each other -- solutions are more likely to emerge that way.

They had a meta-discussion of the fact that Fridman has been "coded" by the left (Thompson and Klein firmly representing the left).

I get that, because Fridman can be so uncritical that it can rise to the level of shilling.

But I also find it curious that many on the left won't sit for 3 hours with him. In contrast, Thompson and Klein sat for 3 hours, which shows me that they have something to say which stands up to scrutiny.

They have something to say that doesn't have to be carefully boxed into 30 or 60 minutes of talking points.

---

Related: even though Fridman can be annoyingly uncritical, I think this also serves the purpose of journalism. Because he gets the primary sources to talk freely.

For example, IMO this part an interview with Demis Hassabis is revealing. He asks if they're worried they will run out of high quality training data:

https://youtu.be/-HzgcbRXUK8?t=3931

From my perspective, Hassabis gives a mealy-mouthed answer about generating synthetic data of the right distribution, and then they change the subject. I would bet there's a lot more to it than that. If they had a good angle of attack, I feel like he'd be more excited to talk about it, and say something more substantive.

I guess you can argue that he's being cagey to not reveal anything to competition, but it seems like a real point of concern to me.

The rest of the interview is talking about AGI time frames and similar sales talk. Whereas my takeaway is that there's significant worry that LLMs are limited by training data, because they interpolate from it (rather than extrapolating), and are inefficient at using it.