Interview with Scott Aaronson about philosophy and quantum computing [video] | Hacker News Reader

Interview with Scott Aaronson about philosophy and quantum computing [video]

youtube.com

134 pointsfurcyd30 comments

Lex Fridman's "AI Podcast" has quickly grown to have one of the best list of guests on any show ever [1] - extremely impressive!

[1]: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuK...

His quest list is awesome. Stephen Kotkin interview was really good.

His problem is the clear lack of preparation in several interviews. He gets thrown off and loses track just when the interview gets interesting. The momentum dies because he has no good followup questions and no idea where to go form there.

Instead of playing ignorant for the sake of the conversation like good host does, it seems that he tries genuinely to learn stuff during the interview. His backup strategy is to direct interview into the same set of overly generic questions where the quest has nothing interesting to say. It's kind of irritating.

If you compare Lex Friedman to Sean Carrol (they have had same quests in their program) its' clear that Carrol is better prepared and better host.

https://www.preposterousuniverse.com/podcast/

>His backup strategy is to direct interview into the same set of overly generic questions where the quest has nothing interesting to say. It's kind of irritating.

Couldn't agree more. I want such people to comment on their fields, not wrestle with some faux-philosophical fluff thrown at them at random. They're not some generic celebrities or CEOs. Their perspective is interesting precisely because they do real work in interesting domains.

Lex is on HN and from a prior thread [1], seems very thoughtful and receptive to constructive feedback. While I've only recently started watching his interviews, I think that reflects one of his main strengths as an interviewer--it seems he thinks both broadly and deeply about AI, physics, and a wide range of interesting topics, which is no small feat. Of course, I'm in awe of his guest list too--kudos to him for building this series. Though I've only seen a few so far, I do sort of share the feeling that the questions can tend toward too general, at least at the outset of the interviews. Like a slow-starting book, this sometimes makes it difficult for me to get hooked right away. But I think this reflects my personal preferences, and many others may prefer his style. In any case, I imagine it takes a tremendous amount of work to run something like this, and there are some great exchanges with world-class guests. It's amazing that everyone has access to interview series like this. On that note, thanks for the tip about Sean Carroll's podcast, I'll check it out.

[1] https://news.ycombinator.com/item?id=22047661

There is nothing wrong being broad and philosophical.

The problem is when he repeatedly pushes it to quests that have nothing to say about the matter or are clearly not interested discussing the subject. You get back only platitudes. He does well every time he adjusts to the guests frame of mind and stops doing his own thing.

Is it a competition? Sean is 20 years older than Lex, clearly loves talking and has way more experience across the table from other smart people. I've been subscribed to both for over a year now but end up listening to Lex's more for some reason. They are fairly different experiences and it's obvious to me that some will like one more than the other. That's ok.

Sean is more prepared, but I like Lex's open-ended style more overall, probably. He gets more philosophical and abstract than Sean does (though Sean certainly does, too). As others have said, it's not a competition; both podcasts are probably the best tech/science ones out there that I know of. Plus they have two different fields of expertise.

This describes my feeling about Lex's podcast as well. I was watching the interview with Jim Keller and was left dissapointed for the most part of the interview because of - as you said - lack of good follow up questions to presented statements and ideas. It's a shame really.

However if you skip trough the first about 30-40 minutes of that same interview, it get's a bit better.

Also, thanks for mentioning Sean Carroll's Mindscape Podcast, will check it out for sure.

I just thought about Jim Keller interview as an example of one of his failed interviews, at least partially failed. Keller had interesting stuff to say but Lex was constantly pushing him into too general or too philosophical direction he was not interested in and had noting to say about.

If you get one of the best microprocessor architects into your show, why not focus on that and dig deep.

I don't think many people have the capabilities to "dig deep" on microprocessor architecture with a sparring partner like Jim Keller. And the bulk of the audience probably doesn't have the capabilities to understand such a discussion anyway.

Don't get me wrong, I wish they'd talked about deeper details, but that's more like wishing for a Jim Keller book than a realistic interview for a fairly broad audience.

Leaving aside specific podcast episodes, podcasts, conference "fireside chats," and the like really don't lend themselves to deep dives especially podcasts that cover a fairly broad range of topics. Especially with a lot of listeners driving or multitasking, you're going to lose 75%+ of your audience on a single technical deep-dive episode, especially given the lack of any visual aids.

In general, people looking for deeply technical podcasts are mostly looking in the wrong place--although, of course, your mileage will vary about what "deeply technical" means.

I think that's antithesis to the goals of his project. He clearly wants to ask those broad philosophical questions of a variety of influential people, and mostly only talk shop when it relates to that.

I think it's one of the appeals of his podcast to perspective guests, although as you mentioned for some people it falls flat.

Coundn't agree more, my understanding is that this podcast if primarely about deeper philosophical questions. And that's exactly what makes this podcast interesting to me! Technical details, on one subject or the other, are easier to come by, but balanced and informed inquiry about deer nature of things across many areas of science is, in my experience, harder to find.

It may be a question of taste and personality. I absolutely love Lex's style. I thought the Jim Keller interview was fantastic.

They're both good. Sean Carroll is 'better' according to a few metrics I can think of, but for some reason I like Lex Fridman's podcast more.

My comment is too old to edit. I tend to listen to Lex's more but I thought this spectrogram comparing 8 seconds of Lex v 8 seconds of Sean is hilarious:

https://imgur.com/a/jo4eNWK

Pretty easy to figure out who's who. Sorry buddy, haha.

I've been bouncing around the smartie-pants podcasts over the last couple years, started with 'After On', moved to Sean Carroll's Mindscape in the last six months, I will have to add this one to the list.

In case you did not know yet: there are transcripts for some of videos on Youtube.

The transcript is time-stamped, the video will jump to the correct position when clicking a sentence and it loads completely at once, so you can use the in-browser search on it.

Open the three-dot-menu below the video content and open the transcript from there.

Nice video lecture series by scot

https://video.ethz.ch/speakers/bernays/2019/7b11b50e-f813-4d...

RSS link for the audio version: https://lexfridman.com/category/ai/feed/

slightly related to this, I read somewhere recently that deep reinforcement learning doesn't really work. Has there been any progress on that front since?

Well - all I can say is that it works for me!

More seriously, I think that it can be challenging to define the "game" or simulation that a DRL system (or any RL) system depends on. Clearly for things like go the game is preset, for the real world, perhaps less so. People have used it vs physical systems like robot hands though.

Also, one of the charms of DRL is the cheating that the agents discover, for example catapulting each other over barriers or blocking doors. But in real world scenarios scams that break the rules of the game (imposed by law or physics) are useless, so you have to rewrite your simulators to remove them, and start again.

What applications do you use DRL for ?

Training a chatbot; we thought of a wrinkle that would let us use DRL to improve performance over pure data trained ones. It was hard to get it going, but once we figured out the code it did add vs pure supervised.

I recently attended a panel discussion on machine learning at the ITA workshop. I asked the panel (several distinguished ML profs among them), which idea was most overrated in ML? The answer: deep RL.

There's been great success to play video games, initially Atari games and more recently Dota and Starcraft II.

Real world achievements are still lacking. Do check the recent advances from Covariant in factories though: https://www.wired.com/story/ai-helps-warehouse-bots-pick-new...

AlphaGo seemed to work.

So a digital perfect information game. It's pretty well-accepted that this is a great use case for RL. The question is how much broader the scope is.

It worked for Dota.

We changed the URL from https://www.scottaaronson.com/blog/?p=4616, which points to this.

Reply on news.ycombinator.com