**Dwarkesh Patel** (0:01)
Like, how long can an AI do a task? Is it like seven seconds and it loses its train of thought? Is it 14 seconds? And it seems to be doubling every seven months. So before, they're going to be right to a little function. Now, they can write you a whole package of code, a whole file of code. Now, they're going to be right to many different lines of code and keep engaging with you over a session as you say, no, I want to add this feature or something. So if that keeps doubling, then you have this really exponential process where you have this discontinuous impact. Going from like 30 minutes to an hour is a big deal, going from an hour to two hours is a big deal, and going from like a day to two days is like an incredibly big deal, right? Like the amount of increased scope you can have as a result will increase exponentially. So I expect these kinds of discontinuities. I am still confused about why these models have no traction whatsoever on things which are like short-horizon tasks. It doesn't take you that long to, let's say, edit a video or book a flight or start playing a new video game, which you've never seen in your trading data. And there they have very little traction whatsoever, or organize the logistics for like a happy hour or something, right? So the big question, it seems like free training is exhausted. The reasoning paradigm, the big question is how far that gets you. And so we had 01, 02, I guess we skipped out with 02, and now we went to 3
The next question is like, okay, look, they've conclusively shown that this gets you a ton of traction on math and code. What are the other kinds of oral environments you can spin up where you get this thing where you're like, now you're doing tasks that are like humans' day-long worth of input? I think in three months, we will know if, do we have models which can reliably do computer use? Can you set up an oral environment for that? If not, then I think the next like we're not going to have some huge economic impact from AI that we haven't seen so far in the next 18 months. Because fundamentally, there's not that much in the economy that's just like chat based. But if you do have that, then I think you can start seeing tens of billions of dollars of revenue per AI company and other kinds of big impacts.
**Ashlee Vance** (2:18)
Hello, hello, it is Ashlee Vance back here with y'all for another episode of the Core Memory Podcast. In this episode, the headline is that the podcaster becomes the podcastee. I sat down for about an hour with Dwarkesh Patel, has a massive YouTube show and podcast where he gets into deep, goes deep on science, history, you name it, tech, everything and he's a great, great chat. We went to Dwarkesh's studio in San Francisco, so I really turned the tables on him properly. And there's more headline news, which is that he has a book coming out. It's an oral history of artificial intelligence technology. Obviously, we get into that at length in the show, we get into his background, everything you could ever want from a podcast on Dwarkesh. And per usual, too, as a favor, give us a review, give us a like wherever you get your podcast. I promise I'll stop harassing you about this once we have asserted our position as one of the giant podcasts of the universe. But until then, we need your help. Anyway, come check this out. Come hang out with us into Dwarkesh's studio. Here we go.
Dwarkesh, how are you?
**Dwarkesh Patel** (3:46)
I'm good. Ashlee, good to see you.
**Ashlee Vance** (3:47)
The tables have turned. You're the interviewer, but I'm in your lair. I'm on the home turf.
**Dwarkesh Patel** (3:55)
We should have actually turned the tables, because this is usually the side I sit on. And I actually should have gone on the other side.
**Ashlee Vance** (4:01)
I was wondering. I was wondering where to see if I was in. But it's better if you can see what's going on. I'm so excited to talk to you. There's a whole bunch of things I wanted to do. One, we'll start with in a sec, is you have a book coming out, which is fantastic. I feel like you and I have known each other a bit, but there's not a ton about you on the internet biographically. I guess partly you're so young, but you've been on this rocket ship. And so I thought this was a good chance for people to get to know you a little bit. And then obviously I've got some AI and other questions.
65 more minutes of transcript below
Try it now — copy, paste, done:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090
Works with Claude, ChatGPT, Cursor, and any agent that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000700778111