The discussion explores the current state and rapid evolution of technology and artificial intelligence, particularly focusing on Remote jobs that can be done through laptops. Historically, technology replaced blue-collar jobs, but now it is the bureaucratic tasks of white-collar positions that are more at risk of automation. The advent of tools like GPT models, which simplify complex processes like programming, shows how human intelligence is indispensable for guiding such technologies to achieve impressive outcomes, even as it replaces routine tasks.

The video dives into the potential for Artificial general intelligence (AGI) and the challenges involved in unlocking creativity within structured environments. While progress is inevitable, predicting how quickly AI can fulfill specific jobs remains uncertain due to the variety of tasks each job entails. Discussion around current advancements like GPT-5, 6, and beyond, points to a future where AI might need to evolve and teach itself, mirroring gaming environments where experimentation leads to creative solutions. Human intervention is still needed, however, to infuse AI with more relatable and emotional understanding.

Main takeaways from the video:

Please remember to turn on the CC button to view the subtitles.

Key Vocabularies and Common Phrases:

1. Remote [rɪˈmoʊt] - (adj.) - Situated far from the main centers of population; distant.

Where we are right now is that if you have a job that is Remote, doable remotely...

2. Bureaucrats [ˈbjʊrəˌkræts] - (n.) - Officials in a government department.

...but now technology is replacing the Bureaucrats, and the blue collar, like, plumbers.

3. Artificial general intelligence [ˌɑː.tɪˈfɪʃ.əl dʒɛn.ər.əl ɪnˈtel.ɪ.dʒəns] - (n.) - A still-hypothetical AI that possesses the ability to understand or learn any intellectual task that a human can.

...about AGI, Artificial general intelligence, which is people call AGI...

4. Robotics [roʊˈbɑtɪks] - (n.) - Technology dealing with the design, construction, operation, and application of robots.

...Robotics are probably going to need AI to write that, to figure out how to build robots...

5. Stochastic [stoʊˈkæs.tɪk] - (adj.) - Randomly determined; having a random probability distribution or pattern.

...the models are Stochastic, so, you know, it's definitely gonna be...

6. Neural networks [ˈnʊrəl ˈnɛtˌwɜrks] - (n.) - Computational approaches inspired by the networks of neurons in the brain.

This is the thing that genuinely happens. Each one is some new way of...

7. Reinforcement learning [ˌriːɪnˈfɔːrsmənt ˈlɜrnɪŋ] - (n.) - An area of machine learning about how agents should take actions to maximize some notion of cumulative reward.

...you start by training these models, like gp four on the Internet...

8. Magnitude [ˈmæɡnɪtuːd] - (n.) - Great size or extent.

...there's probably an order of mag. It takes, you know, ten or 100 neurons...

9. Paradigms [ˈpærədaɪmz] - (n.) - A typical example or pattern of something; a model.

I always assume there's this and then another s and then another s.

10. Iterating [ˈɪtəreɪtɪŋ] - (v.) - Performing or uttering repeatedly.

Interesting. So that's actually a more dynamic place for them to live is like learning how to write code and Iterating on what happens with code it writes.

Unlocking Creativity Beyond Human Capabilities

Where we are right now is that if you have a job that is Remote, doable remotely, which at this point is most laptop jobs. Right. Computers, white collar jobs, maybe get better at that. Yeah, it feels really achievable that we're gonna be able to get there. This is very unintuitive because forever technology replaced blue collar work. But now technology is replacing the Bureaucrats, and the blue collar, like, plumbers are like, we're fine, but, like, the annoying lawyer people are like, they're in trouble. I mean, the thing right now is what we're seeing is that technology is basically eliminating the rote parts of the jobs, right?

So when I used to program, I remember I would have to write all these XML configs. And nowadays you just ask GPT four and it'll write the XML configs for you. But you want to do something really interesting, then you need to bring your human intelligence to the problem, even if it's just telling the model how to go about it. So you're closer to this than 99.999% of people in the world. So I want to ask your intuition, because I've talked to Sam, obviously, who's involved in this. I think you may even be more technical than him, and you're really close to the research.

So, you know, the next 1020 years, like, what else are you gonna. I mean, are you just gonna be able to do everything people can do in five or ten years? Is this like, a total change in all of reality by the 2030s? Are there actually a lot more steps potentially still, like, what's gonna happen here? Yeah, I mean, the thing about AGI, Artificial general intelligence, which is people call AGI, is that it's very clear to me, and I think it's clear to people who are working in the field, that progress is going to continue to happen. What's hard to predict is when you actually are able to do a particular job, when the. There's so many different things that are wrapped up in any particular job that it may turn out that something we didn't even expect was the last thing that needed to be automated.

So I think we're going to have a long time. We may not have perfectly realistic Westworld robots walking around in seven or eight years. Robotics. Robotics are probably going to need AI to write that, to figure out how to build robots for us. But, I mean, you're gonna get. So, I mean, that's the other question is, you know, I don't know what public or not, so push back on me, but, like, you know, the thing I've heard from various people is you could probably do GPT five and six and maybe seven with people, and then at some point you're gonna need to get to eight or nine or whatever.

You're gonna need GPT itself to do that. If I was running a research group, I'd already be trying to probably figure out how to get this thing to teach itself in different ways. That's probably like, there's lots of concepts there. I'd imagine that you're trying to figure out like how far are we away from it teaching itself? And is that a thing you guys are working on that you could say?

Yeah, I mean, I think that's what everybody's, I think that's sort of the open problem that everybody sees right now, is that, you know, the current, as you say, the current GPTs are all just mimicking what humans do. So how do you, how do you unlock creativity? Right? And if you think about it, it's that same problem we had with Dota or with games, where because you have this structured environment of a game, it was able to be created within the structured environment by trying new things. So how do you give it, how do you give it the ability to be creative within the structured environment of reality? And that we don't know.

And I think the interesting thing even now is just how do you build a structure of reality for the agent to play in? That's interesting. I imagine you already have lots of versions of 3d structures of reality, obviously for Robotics to play in, but I guess it's more complicated to have a structure of human reality. I think the most interesting structure we have now is actually the compiler, the interpreter, and being able to write code, run code, see what happens.

Interesting. So that's actually a more dynamic place for them to live is like learning how to write code and Iterating on what happens with code it writes. Yeah, I actually spent a lot of time at OpenAI thinking about 3d worlds, and ultimately the problem with the 3d world is it's very limited because you have to create everything interesting that's in there.

That's fair, I guess you could create a bunch of human and emotional actors or something like this for it to interact with, but you're going to create some really dorky robots that are all just going to be really good coders. Hey, you know what, you can, you could be all the non dorkiness, humans can bring the non dorkiness, and, you know, I will not be able to bring that, but normal people will be able to bring that for you, it's almost like you kind of need more human data. Like, you need more data about how people just do normal human things and how they're feeling and how they're thinking. Right.

Because that's what we're missing. Well, it's unclear. The funny thing is, there's actually a ton of just. If you just think about the Internet, it is so huge. There's so much, and so there is a lot of data out there about how, if you think about a math problem, like how to approach the math problem, how to iterate on it, how to work with it, and then the other thing that we've done and that other companies have done in the last couple of years is hire humans to give us data.

So you start by training these models, like gp four on the Internet, and there's this finishing step that you do, which is called Reinforcement learning from human feedback, or RLHF. Reinforcement learning. It's like padlocks, dogs, you know, you give it a treat if it does the right thing. And human feedback means that you have humans who are doing it, and so they'll have it, you know, try to solve a problem, and if it has a good approach to solving a problem, then they'll say, great, you had a great approach, whether it got the right answer or not, that turns out to be really helpful, and just do this over and over again, and you don't have a ton of data compared to what you pre trained on.

One thing that's been unintuitive to a lot of us, we haven't caught up on any of this, is just that you think it would get better and better over time. But a lot of people seem to experience it getting worse over time. At least their latest interactions. Are they correct that somehow some parts of it seem to be getting worse over time, or do you have any thoughts on that? We think it's actually getting better over time, that some of the reports are confounding various different things that people are confused about.

I think people are just confused about this because even really smart friends of mine who are close to me feel like it's not answering them as well as it was. But maybe they just got to expectations too high or got confused on that somehow, huh? Well, I mean, the models are Stochastic, so, you know, it's definitely gonna be the case that the answer you remember is the one when it got it right, and it doesn't display that flash of brilliance all the time. And what's your intuition that you could say, for.

I mean, obviously, GPT-3 was just fundamentally different than two, and four is much better than three. Like, are we getting to an asymptote here with your work? Is five just like, way better than four? How are you feeling about this? The way I think about it is that we are still pretty far away from the level of scale that is the human brain, and so there's no reason for there to be an asymptote. I think it just keeps working.

When you say level of scale, what's the intuition for that? Like, what's the scale of three or four versus five or six? Yeah. So you can count the number of neurons, and, you know, this again, like I said, this is very rough, because a computer neuron and a brain neuron are not at all the same thing. You know, there's probably an order of mag. It takes, you know, ten or 100 neurons to simulate a neuron in the brain.

But, you know, if you look at something like GBD three, you can say, oh, well, that's maybe a lizard. Number of neurons I and GP four is like maybe a cat, and you still have many orders of Magnitude to go before you get to something that is actually the same size as a human. Is that really relevant, though? Because, I mean, you'd say this is human. You say, this is a blue whale, and it's like, a lot more. And blue whales aren't. I assume they're not much smarter than us.

Yeah, we don't know, actually. They could be doing really cool philosophy in the ocean that we don't know about. But I think the other thing is that if you think about the data blue whale is trained on, it's trained on, where are the fish? When does it open its mouth? You know, when does it swallow? And, you know, it's not.

We're training. We're training on the human data. So you want something a size human brain train on the human data. You still think it gets a lot better the next three, four, five years? Like, we're on a curve right now. Like the world. I mean, it makes sense for forever to work on AI, because it's going to be so much more important in three years. So that's your. I don't see any reason for it to stop.

There's no number of years in which you would start to asymptote intuitively to you. Well, look, I mean, I don't think I can put a number of years on it, but I think when you pass human level intelligence, that's where it gets kind of crazy in lots of different ways. Maybe learning by mimicking humans no longer works. Maybe there's an asymptote there. Maybe we have to switch to one of these other techniques. Maybe you'll finally feel understood by a creature that exists.

The way I sleep at night on this is I assume that there must be multiple different s's. Everyone always assumes there's an exponential. It just goes to the moon and the world's singularity, and I always assume there's this and then another s and then another s. Is it possible there's different Paradigms you have to figure out? That could take a while, or you think you guys have it all figured out enough to go all the way? I think there really are. I think these things are fractal.

And so, in some sense, the paradigm is Neural networks. But in another sense, well, you needed to figure out transformers. And so each of these s curves, this is the thing that genuinely happens. Each one is some new way of making sure that you can fit more compute and more data into a larger network. And every time you sort of see a peaking, you need to have some sort of breakthrough. But it's not as fundamental a breakthrough as it was before. The difference between Neural networks and linear regression was really big. And the difference between Neural networks now and Neural networks two years ago is just a series of tricks.

Artificial Intelligence, Technology, Innovation, OpenAI, Future of Work, Neural Networks