AGI is Almost Here! What comes next? What's left to do? How do we adapt and prepare?
26:47 Share

When launching a new product, finding the right revenue model can be just as important as the product itself. That's why Stripe built Stripe Billing. Stripe Billing is the advanced billing software to grow any business. From seat-based subscriptions to usage-based billing, and every pricing model in between, Stripe Billing helps you grow your business, your way, fast. Learn how Stripe Billing can help your business at stripe.com slash billing. Is crypto perfect?

Nope. But neither was email when it was invented in 1972. And yet today, we send 347 billion emails every single day. Crypto is no different. It's new. But like email, it's also revolutionary. With Kraken, it's easy to start your crypto journey with 24-7 support when you need it. Go to kraken.com and see what crypto can be. Not investment advice. Crypto trading involves risk of loss. Cryptocurrency services are provided to U.S. and U.S. territory customers by Payward Ventures Inc. PBI DBA Kraken. Visit PBI's disclosures at kraken.com slash legal slash disclosures.

This episode is brought to you by Allstate. Some people just know they could save hundreds on car insurance by checking Allstate first. Like you know to check the date of the big game first,

before you accidentally buy tickets on your 20th wedding anniversary and have to spend the next 20 years of your marriage making up for it. Yeah, checking first is smart. So check Allstate first for a quote that could save you hundreds. You're in good hands with Allstate. Savings vary. Terms apply. Allstate Fire and Casualty Insurance Company and Affiliates. Northbrook, Illinois.

With the release of 01 Preview, we are pretty much a stone's throw away from artificial general intelligence. However, there's a lot of nuance to unpack here. So let's take a deeper dive, a much closer look at where are we with respect to artificial intelligence and what comes next. So, Strawberry, 01 Preview.

This is pretty close to AGI. Now, the reason why I say that is because there's been a lot of conversation out there on Twitter, on Reddit, and so on and so forth. The benchmarks notwithstanding, which are already looking really good, the actual academic and scientific and economic impact of O1 is rivaling or surpassing humans in a lot of capabilities.

So for instance, there was one story out there on Reddit where someone said, oh, Strawberry or O1 Preview was able to do my entire thesis in about an hour where it took me a whole year. That's pretty significant. Now, if you watch my previous video about how smart is Strawberry, you'll kind of be up to speed. But the very TLDR is that people are comparing O1 Preview in terms of its usefulness to that of a good grad student.

And this isn't even the full one. As many people have pointed out, 01, the full model, is coming out in October, followed shortly by Orion, which Sam Altman hinted at is coming out around Christmas. So Project Orion could be GPT-Next, it could be GPT-5. We're not entirely sure, but it is the next generation model as far as we know. So with all of that being said, it's like, okay, this has an IQ of about 120.

Although the reasoning still fails in some cases, I do want to point out that there are plenty of people out there that have concluded that it's not that great, that it still fails at some really basic things. But at the same time, a good user is able to get a lot of juice from that squeeze, pun intended. So, moving right along.

One of the things that people have said is it's not AGI because it's not embodied. It requires embodiment data in order to be AGI. I completely disagree with that. And let me tell you why. Number one, it's all data.

The the IO coming from in and out of your brain from your body is just data to you to like your brain doesn't care if you simulate it, you know, cut your head off and, you know, remove your eyes and ears. You could hypothetically simulate all of the IO coming in and out of your in and out of your brain. Your brain is a thinking engine that's kind of floating in a little jar.

Now, another thing that I'll point out is from the perspective of the AI, it's all just data. But also, most of the most valuable tasks for AGI, such as scientific modeling or software development, doesn't require physical presence.

All you need is an API to get out to whatever resource it is that you're controlling. So whether it's, you know, whether you look at it as just data API, like you can puppet a machine. So this is kind of where I'm at is if you have a robot, if it's a good enough robot, you can just puppet it from the data center. You don't need to be embodied.

And then really kind of the primary metrics that I'm looking at in terms of does this constitute AGI are economic impact and scientific impact. So saying that embodiment is required, I personally feel like that's just mostly a red herring at this point. Now, with that being said, what I will say is that in order to realize all of the economic gains that we want, we will still need robots.

The embodiment data is not necessarily required, but in order to automate, say, construction tasks, we'll need good enough robots. Right now, it's more or less just a genius in a jar, which that's actually not necessarily a bad thing. The fact that it's not deployed widely and like, hey, AGI is right here or it's right around the corner.

means that we still have plenty of time to work out safety issues. You know, this is one thing where I've been kind of confused by the AI safety movement, at least the more hardline people, is just because you have AGI doesn't mean that the model, which is enormous, can copy itself out of the data center. It also doesn't mean that it's going to fool the people developing it to say, oh, by the way, I'm ready. Like they're red teaming it like crazy. So anyways, moving on from the safety conversation,

It's really difficult to find numbers, so I did some searching around and depending on how the job is classified, it's really difficult.

Here's what I'm trying to say. I was looking at how many jobs are sedentary? How many jobs do you do completely in front of a computer? Well, if you look at that, then only 13% of jobs are quote-unquote totally sedentary. Another 24% of jobs require light physical effort. But then if you look, okay, well, how many jobs are actually white-collar or office jobs? Then in that case, it's up to 65% of jobs are office jobs.

So it's like, that's a pretty big margin. So

The long story short is even if we don't have robots yet, but they're not that far behind then somewhere between 13% and 65% of jobs are pretty soon at risk for automation so Moving on we will need to see robots with human dexterity in order to get humans completely out of the workforce Which personally I think that that is inevitable and good People might disagree

I call that post-labor economics. Anyways, moving on. If you've been here on this channel for any length of time, you're familiar with my concept of post-labor economics. Next up, statutory friction. So you might say, okay, great. We have AGI. How will we roll it out? There's quite a few things that are going to slow it down, namely red tape.

So whether it's regulatory barriers, because a lot of companies are going to be like, I don't know if we're allowed to do this. And we'll talk a little bit more about enterprise adoption in a second. But also a lot of jobs just were not designed. They were not contemplated to be done by machines. It's assumed that humans will do most jobs. And so it'll take time to create new paradigms around how do you get work done with AGI?

Now you might say, "Dave, that sounds pretty obvious. Just give it an API and just give it a robot body." Yes, but it's not going to be quite that simple because there's a lot of infrastructure that needs to be in place. There's a lot of business processes that need to be updated and so on and so forth. Furthermore, I suspect that the unions are going to have something to say about all of this and they're going to fight tooth and nail to prevent the rollout of AI and robots as much as possible.

going up to the point of sabotage efforts, such as regulatory sabotage. I can't remember the guy's name, but someone just retired or resigned from the EU who had been responsible for a lot of the anti-AI legislation going around in the EU. So we should be vigilant for people trying to sabotage the rollout of AI and robotics at all levels, not just in the corporations, not just in the safety movement, but everywhere.

in the upper echelons of government as well because, well, everyone is reacting emotionally. And that's not to say everyone needs to be stoic and use pure logic. All humans are emotional. But my point is that this neo-Luddite movement will try and slow things down and they will throw shoes in the gears as much as possible. And that is just something that we got to live with.

All right, so I want to take a quick break and tell you about my new era Pathfinders community. So this is my new community, which is for us navigating this transition. The tagline is very simple. Conquer AI overwhelm, unlock new opportunities and craft a life of meaning and purpose.

As things are changing, we will need to navigate this transition into the fourth industrial revolution. So we've got this amazing community with lots of people. We're doing PBL projects. So basically people are playing along with my Raspberry project, which is an open source version of Strawberry.

Then we also have, I've got a bunch of classwork over here. So this is basically structured learning for the journey that I've been on in terms of everything that I've learned and used to adapt my life and my career to this AI. And then finally, we also have regular webinars.

So these are Q&A sessions. These are practice sessions. Lately, we've been practicing communication skills. So last week we did what I call discovery dyads, which is a particular mode of communication. We practiced it. And with just a few minutes of practice, we had people talking to each other confidently that had never spoken to each other. And this week we're doing circling or relatefulness meditation. And then next week we're going to be doing like cocktail party kind of communication. Okay.

All right, so crossing the chasm. This is where we talk about enterprise adoption. So enterprise adoption is going to be...

A bigger hurdle. And the reason that I say that it's going to be a bigger hurdle is a few reasons. Number one is enterprises are famously risk averse. Having been in many enterprises, there is a lot of hesitance to adopt any program or platform or application that is not being provided by a trusted vendor. So until AGI is offered by Microsoft directly or by Oracle directly or so on and so forth,

most enterprises are not going to adopt it. They're not going to adopt an AGI product by open AI because they're like, well, we spend a couple of years integrating it and then open AI just gets rolled into Microsoft and all the APIs change. We're just going to sit and wait.

Is crypto perfect? Nope. But neither was email when it was invented in 1972. And yet today, we send 347 billion emails every single day. Crypto is no different. It's new. But like email, it's also revolutionary. With Kraken, it's easy to start your crypto journey with 24-7 support when you need it. Go to kraken.com and see what crypto can be. Not investment advice. Crypto trading involves risk of loss. Cryptocurrency services are provided to U.S. and U.S. territory customers by Payward Ventures Inc. PBI DBA Kraken. Visit PBI's disclosures at kraken.com slash legal slash disclosures.

And if you don't believe me, that is exactly how enterprises think. A lot of CEOs, when polled, only like 13 to 25% of CEOs are like gung-ho about artificial intelligence. Most of them are still in a wait-and-see, I'll believe it when I see it kind of situation. And the reason for this is because of big tech hubris.

Silicon Valley is well known in all enterprises to have, let's say, an inflated sense of self-importance and also to continuously over-promise and under-deliver. That over-promise and under-deliver is why many enterprises are like, you know what, we're just going to wait until we see the economic proof.

We're going to wait until we see our competitors start to roll out some of these products. We're going to wait until the data shows that we must adopt this thing or their shareholders demand that they adopt it. All of this will conspire to create a slow adoption curve at the enterprise level, which again, that gives you more time for safety and other aspects, which we'll talk about right now.

When I so let me provide a little bit of context. I went to a very small local AI safety conference last week and one of the things that was that we talked about was how do you view AI safety in this changing landscape?

In the old way, and when I say old way, I mean like Eliezer Yukowsky way, they kind of had this vision of this nebulous Cthulhu Lovecraftian ASI is going to arrive one day and it's going to have this litany of characteristics. None of that has panned out, by the way.

Instead, the way that AI is arriving is it's decentralized, meaning that it exists in many data centers across many places. There is also an iterative approach to development where, and this is where I will agree with Sam Altman over Ilya Sutskever, is that Sam Altman, so for some context, back in the summer of 2023, this is about when we think it was happening, said,

Basically, Sam Altman and Ilya were arguing over whether or not you just deploy quickly and roll out and see how people use it and then adapt. And Ilya said, no, we need to do more safety testing in a vacuum. And in this case, I will agree with Sam is because you release a low-grade product first,

which is why they released ChatGPT. They said, let's get it out there and see how people use it, see how people misuse it. So that way we can start getting some feedback from the marketplace and from regulators in order to steer safety. And as Sam Altman said in several interviews, they could not contemplate all the ways that people would use and abuse ChatGPT.

And so I will agree that getting stuff to market sooner rather than later. Now, obviously, you do need some safety testing, but keeping it locked behind closed doors for years is not the way to do safety testing. So there are five primary feedback loops that are going to steer safety from here on out.

As we build more robots, get to AGI, and then finally to ASI. Those safety feedback loops are one, market feedback. So market feedback is just who decides to buy what, and what does the court of public opinion say? So for instance, you might remember a few months ago when Gemini rolled out and it was

Super woke, like the things that it would refuse to talk about. And then Google's faux pas where like it would just flat out refuse to generate white people. Um, like they got a lot of market feedback very quickly and Google still hasn't recovered. Um, in, in terms of optics and the mark and the court of public opinion enterprise feedback, as I was just talking about, businesses are very skittish about new technologies.

And so the requirements that enterprises have, so like if you want Goldman Sachs to adopt AGI, guess what? It's going to have to be provable. It's going to have to be transparent, explainable. Like there's all kinds of requirements. And that's just what the enterprise is going to want to see, notwithstanding the regulatory requirements to be doing financial transactions.

So the enterprise feedback to OpenAI and Microsoft and Google, now granted Microsoft and Google are already enterprise software, so they are aware of the life cycle that this is going to require in order to get adopted.

Next up is military feedback. The military has very stringent requirements on, you know, when I hit the off switch, the off switch had better work. It's got to be reliable. It's got to be it's got to have fail safes built in and so on and so forth. So if OpenAI wants military contracts, guess what?

It's going to have to be explainable, it's going to have to be interpretable, and it's going to have to be reliable when you tell it to stop. Next up is government feedback. So government feedback has two parts. Number one is regulation, which we're already talking about. But then if you want to sell AGI to the government, once again, you have to be an approved vendor and you have to prove that this technology is good. Now, the U.S. government is not well known for adopting technology quickly. So...

All of these feedback loops are going to slow down the rate at which AGI is deployed broadly, which, of course, also those feedback loops will inform the safety movement as well as just the commercial viability of these products. So moving on.

Open source. Open source is never that far behind the big shops with one exception and what I will say is that scale might be the biggest moat from now on. If it's going to take $100 million or $1 billion to train models from here on out,

And then who knows how many millions, hundreds of millions of dollars to do the research to fine tune them. It's entirely possible that only the big players will be able to do this from now on. Now, what I will say, though, is having watched some interviews and podcasts with various people working with constrained resources often forces you to be more creative.

And so for instance, my, my raspberry project, we're going to try and get as close as we can to strawberry, but on a shoestring budget.

So what we find out with this, who knows? But, you know, often it's constraints force you to be more creative. And, you know, we saw this with the Soviet Union because the Soviet Union had far less resources than America. So they basically basically the Russian scientists had to work a lot of stuff out on paper where we had supercomputers to send to simulate stuff. And guess what? That forced them to be more creative. And they did pretty good. Likewise, China is operating with fewer supercomputers than America.

But again, where there's a will, there's a way. So I suspect that open source will only ever be six to 12 months behind the flagship models with a caveat being that we will probably not have the scale. Although time will tell.

Now, even more exciting, next gen models are coming. 01, so what we have right now is just the preview version. 01, the full model is coming out in October, at least rumor has it. And then Sam Altman has hinted that Project Orion is coming out in Christmas. So we're going to have several steps in a row. And remember that 01 is about as much of a leap above

Current models to O1 preview is about the same leap as going from O1 preview to O1, at least if the benchmarks are to be believed. Now, time will tell, of course, in terms of how economically valuable and scientifically valuable it is. But if O1 preview has already surpassed most humans on many tasks, it's safe to assume that this trend will continue. So, what does that mean?

I mean, I would not be surprised if everyone in hindsight agrees that we had full AGI by the end of 2024, if not in 2025. Now, of course, you know, 2027 is the number that a lot of people were kind of closing in on. But it's like, well, if we have another two years of this, where is that going to be? That's going to almost certainly be artificial superintelligence, right? Who knows? Time will tell.

Now, one thing that I noticed about the way that OpenAI has been behaving is that it reminds me of the chip industry. So the chip industry has what they call the TikTok model. So the TikTok model in the chip industry is first you change the microarchitecture, then you shrink the die, and then you rinse and repeat. So that is basically every time you hear like, oh, we went from like 8 nanometers to 6 nanometers.

That is the that is the micro architecture changing or maybe that's the die shrinking. I don't know. I'm not a chip design expert. Anyways, they follow the TikTok model. And what I noticed that OpenAI was doing is that they did the same thing. First, they scale the model and then they optimize it. That's what we were seeing with when we went from GPT-4 or chat GPT to. So then we added 4.0 and then 4.0 mini, which were optimizations.

And so now we have a new flagship model which is the O1 series or the O series. So I think that means Orion. That's what the rumor is. So now we have the Orion series models, but they're much larger, they're much slower, they're much more expensive. So now we need to optimize them to bring the cost down. So I suspect that OpenAI is going to be abiding by this model from now on where it's scaled and optimized. Then once you've optimized it as much as you can, you scale again, optimize again.

Again, this is a very well established paradigm in the chip industry. They've been doing it for decades at this point, like basically since Intel was founded. And so I expect that over the next six to 24 months, this new new generation, the strawberry generation of models will the price will continuously come down. We'll probably lose some intelligence with those optimizations because with with these kinds of models,

Basically, efficiency always comes at the expense of intelligence. But once it gets cheaper, as Sam Altman frequently says, intelligence too cheap to meter. Well, at $1,000 or $2,000 a month, that is definitely not too cheap to meter. So they got a little bit of work to do on the optimize phase next.

Now, one thing as we wind down that I want to talk about is human exceptionalism. So as I've been talking about this, I've been reading comments both here and on Substack and on Twitter, and there's four primary arguments that I have seen saying, ah, this isn't actually AGI or this actually isn't artificial. You know, it's not actually smarter than humans. And so one of those arguments is it doesn't truly understand.

Which I say is irrelevant. The output is accurate and measurable and the output is also economically and scientifically valuable. Therefore, those are the outcomes that really matter. Whether or not it is quote-unquote "truly understanding" what it's doing, I mean, I'm also the same person who argues that most humans don't truly understand anything. Your brain just gives you a signal that tells you "I understand this!" But if you read comments on the internet, most humans are confidently incorrect, including myself at times.

So humans don't true. There is no such thing as true understanding. That is a red herring. That is a no true Scotsman argument. Next up is no real experience. They're saying, ah, well, it doesn't have subjective experience in no way is consciousness or sentience or, or phenomenal experience remotely required for it to be useful. A wrench doesn't have experience, but it's still useful to you. A water wheel or, you know, the computer that I'm recording this on, none of those need