AI's impact depends on task complexity and real-world bottlenecks

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 32:48

   find i don't i question the premise of the two extremes i'll give you my perspective so first part of the reason i think you have the juxtaposition of the models aren't changing and oh my god everything has changed is how much intelligence you need for the task so if you're using chat gpt to plan your vacation and you use gpt40 and use gpt53 my guess is the experience won't be that different because planning a vacation on chat gpt doesn't require a huge amount of intelligence to do it was already had reached sort of sufficient quality you know a year or two years ago that actually like everything is sort of feels like incremental on top of it in contrast if you're using the model to write a rust module that is going to do something low-level systems software with like high sensitivity around both correctness and latency

2. 33:44

   the models were like woefully and like they just could not do it two years ago and all of a sudden over the past three months they could i think what's ending up happening is if you end up i don't think it's really possible to do but if you're just like take every task you do in your life and and sort of order them by how much intelligence is required to do them we're sort of moving down that line and if you're testing something that we'd already like the horizon had already passed you're like nothing's changed and that's actually people's lived experience with chat gpt a lot of it is these sort of simple and they're useful right but we already we already solve that. So now it's a commodity. And then if you're trying to do drug discovery or software engineering, you're seeing it change on a daily basis. That's why I think there's a juxtaposition.

3. 34:28

   And put it out of the way, one interesting sort of thought exercise is we've reached sufficient intelligence for a lot of tasks. Put it out of the way, if we paused innovation and just absorbed the intelligence of all the existing models, my guess is there's still trillions of dollars of economic value we haven't realized yet, which is interesting unto itself. Uh, I think it's really interesting. Then you go to like the, what's going to happen. I think a couple of things are true. I'm an optimist in all of this. The reason why I don't, I'm not in camp Mai Tai is I just don't believe that humans will stop doing things. I think we've, uh, through automation, like eliminated a lot of jobs from the jobs of a bank teller to a lot of agricultural jobs, which have, you know, I think it's like what, 5% of our jobs are in agriculture today compared to

4. 35:19

   like 95%, you know, a few hundred years ago. And so just because we automate jobs, I think we often lack the imagination to imagine what comes on the other side of it. And I firmly believe that, you know, I just think that we are, I believe this technology is fundamentally a tool, even though it's a really impressive and super intelligent tool, but it's fundamentally a tool and we will create an economy and identity around it.

5. 35:46

   And I don't believe that because it took something we're doing now, it's taking away our identity and we want to go sit on a beach. I just don't believe that as a human being. The other thing, though, to temper the excitement is suffering. engineering is both getting the most attention right now because the research labs know to create self-improvement for AGI, a automated AI researcher is a prerequisite. So as a consequence, all the smartest research labs are specifically working on this. So it can write its own, write the next series of models. That's right. And if you saw the OpenAI post on 5.3, there was a lot of openness about sort of using the model to help build the model, which I thought was really interesting and what's happening in all the labs right now. So first, just because software

6. 36:34

   engineering is getting great, the idea that that's completely general is not, it may be true, but it's not obviously true. There's a lot, how much this generalizes the G and AGI is the hard part. And it certainly generalizes, but how broadly is an interesting question. The second thing is like which parts of the economy can basically absorb intelligence completely fluidly. Software engineering is definitely one of them. It's a purely digital profession, right? You write code, you compile code, you produce, you know, binaries. And so the entire process is digital and you can test it digitally, contrast that to, you know, a drug discovery. It might work until you need a wet lab and then you need a wet lab and all of a sudden you need robots. Okay. Then unless you did robots in the wet lab, then you want to bring it to market. Well, you need a

7. 37:22

   clinical trial. And that clinical trial, even if you have great ideas on how to do it better, it's still a process that is regulated by the government. Like there's no amount of like intelligence you can just like pour onto that process and make it run faster. And so the interesting thing I see is like clearly finance and software engineering are professions where you can take intelligence and sort of absorb it into the profession quite efficiently. You know, because a lot of the job is essentially operating in the world of information.

8. 37:52

   And then you say, okay, like, what are all the other parts of the economy? And not only are we absurdly focused on coding agents just because of its importance to the AGI labs, but also it's not obvious to me that every other profession is as easy to take an information-oriented digital agent and transform the profession. So I think it's going to have a really big impact, but I think the takeoff in the AI labs might be faster than the takeoff in society to some degree, just because different professions will be sort of differently impacted by it.