Summary
Notes
Transcript
So what the US government just did both with Amtropik, they also did with Planet Labs in terms of restricting access to the space-based imaging data that we have. It's kind of an interesting time. I'm sure it'll probably engender a lot of discussion because we can talk about this a little bit. So I'll give you my background. So I've been at this, about to start my 20th year at Stanford, so I was a Stanford alum. A little bit different background than most of the people here. I came to Stanford as a technologist. So I had spent an early part of my career, I founded two large public software companies. I was a chief technology officer.
Came to Stanford. My company got bought while I was at Stanford. I ended up founding another company when I left that became the highest funded private company in the US. I led that for about eight years to its IPO. And then retired from the company at the IPO to take a faculty While I was here, actually, I did a fourth company that was one of the first big machine learning companies spin-outs out of the CS department.
That was back in 2009. That company grew. It's now called Nielsen IQ. We didn't use Nielsen data, but that was the data science team we had spun out. About 16 years ago, the person that was teaching the adjacent class to mine turned out to be Eric Schmidt. Eric was CEO of Google at the time. Eric and I ended up partnering together. We founded a venture capital firm almost 17 years ago that manages about $3 billion in capital.
It does very, very deep tech investments. I helped Eric write the check that created Ed Funds when he was from Dario out of Oakland, Indiana. So my work here, being really technical at the business school, wasn't really cool until about three or four years ago. Suddenly, all the stuff I did prior to my career turned out to be And then in addition, so I teach a couple different things here at the GSB. So I teach our capstone entrepreneurial class, taught that for about 15 years. Our core research class, we teach our MBA ones as part of the work that I did here.
And then most interestingly, recently, I'm the faculty lead for our applied AI initiative at the university. So I spend a lot of time thinking about how How do we teach our graduate students to be facile with the technology? So we have a lot we can cover. I brought my four times more than a class So we'll pick and choose. So maybe just to triangulate a little bit, I know many of you come from very different backgrounds.
If that's helpful, that's probably a good place to dive in. We've got a pretty heterogeneous group, but that's probably a good place to start. So, interrupt. This is way more fun for me if you guys want to engage, and I can give you specific examples of things that are going on. And I'm not going to belabor this, so I have one slide from the 50th reunion class I gave, but what was interesting is I stepped back and looked at the world through their eyes and just looked at what major perturbations did this 80-year-old cohort experience in their life and just looked at it across a spectrum. And this chart is basically how long does it take something to reach effectively a market level tipping point, which is typically 25 to 30% penetration.
Normally when a technology can reach that level of penetration, it reaches a sustaining level. So if you look at this go-ahead, for example, electricity, telephone, automobile, all these original analog infrastructure technologies took decades. And it's actually not that difficult for society to respond to things that happen over decades. Digital television PC cell phone these all still took more than a day so again relatively easy for people and companies and society to adopt as we start getting the internet we start seeing things a little faster but still trending towards about decade and then chat shooting team came out and took 60 days to reach basically a market level tipping point so we're seeing a pace of change and this will be consistent in every single thing I talk to you about is I'm going to show you a snapshot of how quickly things are going to change Thank you.
Mostly so that you as leaders of companies understand what the rate of change looks like and what are some of the challenges that you're gonna face. Both in terms of the talent you have in your company and the technologies you decide to adopt. And I can, we can talk about that in a little more detail but it's important to understand this didn't happen in a vacuum. There's a series of combinatorial things that were occurring adjacent to this.
But the way we did it at the time, and it's important just to understand what was going on, what changed, and what that ultimately allowed. So if we were processing a sentence like, I crossed the road today, the early forms of Bayesian basically would look something like this. We would train a model. We would then be processing a sentence so we would get to the and we would basically try and compute what statistically is the next word that we're going to see.
So we use things like technologies like back words, there's a bunch of things we could do, but basically we would create a hidden state of the past, a computational formulation if you will of all the words that preceded this, and then we would guess what the next word was. And then based on the next word, we would guess the next word, and the next word, and the next word. So one, it gave us no long range prediction.
We could only see one word ahead. And then two, it wasn't computationally efficient because it was sequential. So every time we wanted to ask the next question, we had to ask the previous question. Then it was also subject to competition of drift.
What you may not realize is all of the work that basically led to that insight was actually based on a mistake. that happened within Google research. So if you step back and I'm here as CEO of Google, what were the things Google was in the process of internationalizing? They wanted to be in every country in the world. So what did they have? They had mail. and search. They have maps. So what did they need to knit it all together?
So many months later, a research team, actually a group of interns, PhD interns, were like, why the hell was that thing yellow? So they went back and did a forensic analysis. And it turns out the language was a South Indian language called Canada. Anybody know what Canada in India starts with? What letter? So the DevOps engineer actually back fingered the training file and trained it on what was effectively a Cyrillic derivative language.
The only thing they couldn't figure out was why the hell did it, why was it yellow and not red? So what they realize is there was in the neural network when they put a huge amount of energy into training the model something happened that they couldn't explain They couldn't quite understand. They were seeing other instances where this was happening. In fact, if you step back and you look at who were the progenitors that went to OpenAI that did the work at Meta, that did the work at Google, many of these people were actually party to what had actually happened.
Because this basically instructed us in that there was something else happening in large model training with high density of data when we were running massive computer density. And that really was the precursor to what ultimately became this paper, which is called Tension is all you need. So tension is all you need was based on a non-Bayesian based approach. It was based on a transformer architecture. So let me explain to you, there's a few things that happened that basically made this work. So the way that they looked at a sentence was very different.
And once they started processing it, the first thing they needed to do was understand the relationships, not just between this word and the next word, but all words. And this was the concept of a transformer, where they looked and they created a mathematical representation between this word and then its statistical, its vectorization and embedding of relationships between this word and all other words.
And they just started doing this across massive amounts of data, and that was this concept of cross-attainment for a transformer.
They started processing everything they had. So something unprecedented in computer science history happened. We saw these models grow, I mean Moore's Law is doubling every year. We saw models grow over 15,000 X in three years. That's unprecedented human history. We went from processing a little bit of data and processing all of the world's, basically the internet, all the world's web of data. So we went from millions of parameters to trillions of parameters.
on their answer, we would go back and computationally reinforce the model. The problem is if you look at how large these models got, there's not enough humans in the world. There's not enough hours in a day. So someone came up, I still can't figure out, I've checked every piece of research, I can't find out who did this, but they came up with this concept of a self-supervised model. So think about it, they've already processed the world's data. So now when it comes time to train it, they took a sentence like John went to the mall and bought a hamburger from Johnny Rockets, and they would statistically begin removing words. So going through computation. When you think about training a model, this is exactly what they're doing.
I spent my entire career from the time I was 23 working as a technologist. And my entire career, I never used the word emergent behavior. But I don't know another word to describe this. So, I mean, we've been doing algorithms for a long time and this is across major, major, like arithmetic, natural language understanding, words and context. We were, and the context of this is the bottom is the amount of flops that went into training models.
So think how much energy, how much, how many dollars were put into training models. So 10 to the 22nd, relatively large model. The models were not very good, not very good, not very good, and then all of a sudden, Something happened, and they were great. They weren't just great, they were shocking and good. And we didn't know why. We kind of had an idea. In fact, we still don't exactly know why. We'll explain-we'll give you some of the details. But all of these people sitting in labs were suddenly seeing this crazy behavior.
Yep. This was right around, call it 2017, 2018. What you may not realize is Google had a working version of a generative model in 2018. What? I mean, it wasn't perfect by any means, and it really did horrendous things, but like, Google's really smart. Eric had stepped down. He wasn't executive chairman at the time. Why the hell didn't they release it? Yeah. They had it five years before anybody else had it. They seem smart.
Yeah. Think about it. I mean, basic innovator's dilemma. So Search was this amazing business where you spend 0.0001 cent to not actually give you an answer, to tell you where to look to find the answer. And on that, they can make 0.0002 cents. It was like this drug business. They could just make unlimited money. This was not that cheap. The other thing was, it did, because it was trained on the world's internet, it was trained on horrible data.
Turns out that didn't work as well with images. So at Stanford, we've been doing work in general adversarial networks for a number of years looking at computer data. Stefan O'Rumon in 2019 wrote the paper on diffusion, which actually took a similar but different approach, where we looked at pixels, and rather than looking at an autoregressive, they looked at a bidirectional context. So effectively, diffusion is an error reduction algorithm.
that's never existed in computer science. And at Stanford, we have kind of our big five CS people at Stanford. Stefano Irma is one of them. So Stefano did this work on images, and then the top ICML paper, the top PhD research paper two years ago, it may have been three years ago, was actually a paper on leveraging diffusion, not for images, but for language. So it was Stefano Irma, Vladimir Khrushchev, and Aditya Gupta.
So they were all-Aditya and Vladimir were all postdocs under Stefano. Volo went to Cornell and worked under Mike Hoganweiner, who wrote the last book, The Age of AI with Eric. He's now the head of CSAIL at MIT. And then Aditya went to UCLA. So they put forth this paper that we can do bidirectional context on language using diffusion, which was pretty radical at the time, and they spun it out. We funded that. So Eric and I gave them the money. That's Inception Labs.
So very, very interesting, because one of the challenges in traditional, and I'm sorry, we're getting deeper than I expected to get, but the problem with traditional transformer-based languages is a couple of things. One, the token processing speed, because it's an X token, can be very expensive, it can also be relatively slow. So one of the reasons why you haven't seen really radical breakthroughs in voice and other things is because they require huge amounts of token processing.
Diffusion can do that. So in fact, the inception model that was put out, it's called Mercury. was actually could process tokens 10 times faster. because it's bi-directional. 10 times faster than a traditional transformer architecture. So in the beginning, or I would say in the beginning we thought this was kind of monotonic, because after all the model monies were going, now you're beginning to see there's a plethora of things.
It's actually this concept of better data. So data that has signal. And this is particularly relevant when you have frontier labs that have effectively ingested all of the world's available data. So think everything on Reddit, everything on the internet, they're basically running out of fuel.
There's not more data.
So we started with a 100 million parameter model. And you'll notice that It was at 38%.
In fact, if you took-Anybody that knows something about medicine and you gave them access to this test, just by guessing, they can probably do better than 50%.
So not only was it not as good as an average human, it was actually worse. It was statistically worse.
because the general data that they were being trained on didn't incorporate that information. So they kind of asked the question in a very reductive way, since we can't get more data, What's the least amount of better data that we could use that would allow us to pass this?
So they ended up, they came up with 1,800. 1-8-0-0. training pumps. And they were very specific. This was pre-reasoning models.
So one was instruction fine tuning, like teaching it things about physics, like what's the boiling point of nitrogen. So it activated portions of the periodic chart and things like that. Then it taught a basic chain of reasoning. A clinician doesn't make a decision on one piece of data. They make a decision based on a longitudinal set of data. What are all the things that I see, and how do these fit together to give a clinical diagnosis? And that's exactly what this test is.
So they ended up, they came up with 1,800 transcripts.
constitutional layout and make the-they make it to teach it how to behave, to teach it what are the rules for how we want you to respond. And they came up with four zero examples. But these were exquisite examples. They took questions from the exam and they just labored over how would I process this data? How would I think about the present? How would I describe this to a clinician so that they could use this effectively? So they came up with 1840 training examples. I think that's 92.6%. And if you read the subsequent paper to this about six months later, they got to what I would say is super intelligence.
They got to more than 95%. And I don't know if they-they haven't published anything else, but I'm guessing they're probably continuing asymptotically to get better and better and better over time.
So Why did I pick this example? I picked this example because largely the frontier models have run out of the world's data to work. If you go back to the fall, one of the biggest things that M-PROFIT did was they created something called MCP. How many of you are familiar with MCP? Model Context Protocol.
What this was, was a way of facilitating the model's access to your data. Everything that lives behind the firewall. Okay. So You can think about this in terms of material science, medical knowledge, biomedical information, all of the world's most valuable data, high-stakehold data is not in public domain. It's actually sitting behind firewalls. So what's interesting in the first generation of what happened in the MCP, because most companies were not very sophisticated in how they allowed their employees to use this, There was a massive sucking sound that happened in corporations where all their data was getting hollowed out.
And models never forget. When they get access to your data, that becomes knowledge that they incorporate into their broader data set. So there were a number of large companies, like I'm working on a project for Ramco, They were, they aired that. They literally prevented any of their data from getting out of the organization. Because the problem is, if you're the biggest player, you have the most data.
What happens if somebody that's not as big as you gets access to that information? What if the models suddenly know that all of the historical intuitive knowledge you have is now available to everyone. This is true in every single domain. So now you can see like a lot of these model companies are trying to create these relationships because they want to look behind the firewall as more of that highly, highly specialized data.
So a lot of what's happening in applied AI now is really taking the core capabilities of a frontier model and then doing all of the post-products Multi-modal.
So for a number of the portfolio companies we have, It's not uncommon for them to orchestrate 50 or more models. in order to do something very specific. There's no one frontier model that can do everything. So you're seeing this become a federated approach of fine-tuned models that are trained on this information that lives behind the model. This is going to be one of the biggest Frontier battlefields that you're going to see is who can get access to this data.
So they had one of the major consulting firms like Shogo School Buses, young people and they just took all this data and organized it in a way that we could use it to train a model and they did. They trained an adhesive model on both the world's data and on the specific information they had and three months after they trained it, they found a replacement for the PFAS additive into their adhesive. Turns out it was cheaper It was more performative than the one that they had.
But the likelihood that they would have been able to do that on their own, I would ask them probably. They would have, I mean, they would have been able to do that. But it's unlikely they would have gotten lucky. Now they're working on a flame retardant. There are both in computational biology, particularly in industrial, we've reached the point where we're well beyond what humans are capable of doing.
campus at the time, but there were these annoying robots. Like, you'd be standing to get a coffee, and there'd be a freaking robot sitting in front of you. And the robot was teleoperated, so somebody else in the building was actually looking through the camera through a 3D HUD, I think like a meta-type headset and hands, and they were basically controlling the robot and doing things. And they had 13 robots.
They identified 700 tasks, and these were pretty basic robots, just a mobile-based, six-axis arm, a little wide-arm camera.
where they actually had the robot complete 700 tasks. So some percentage of each of those 700 tasks were done multiple times. And then they trained a robot.
It didn't work. There were limits to what we could do. This is a robot that was trained on this model and then put into an environment it had never been in. The only thing they gave him was a ray of hope. And that's it. So it was using a language model and it was using a world model. So watch what happens. So you'll see the prompts here. So, "Robot, go bring me the rice chips from the drawer." Didn't tell what rice chips were, didn't tell what a drawer was, didn't tell it how to open a drawer, just gave it the basic layout, like a CAD drawing.
Okay, and maybe it would have started to move and I need to create an image of that robot. So now I don't have an image because it's purely synthetic. I used diffusion and they said, okay, I want you to make an accurate physics based image of this robot at this precise point in time based on the physics of what the robot can actually do. And then they would say, "10 seconds." Eleven seconds. 12 seconds, 100 seconds, 1,000 seconds, and you'd be creating thousands, tens of thousands, millions of images.
What happens when you take images and you run them together? What do you get? You can do it. So they found a way to create synthetic data because this is rooted in core physics of the world.
So we were able to generate massive amounts of synthetic data using a computational approach to generate the images.
Now once we have the images, we can feed them into a model and train the model. This is why if you go back and listen to Jensen's GTC, what he's talking about is Omniverse, which is their world model.
And in fact, the first version of it was trained on 1.4 million hours of video.
What we've observed is that when a model observes people moving and things moving, it actually begins to understand the physics of that environment, such that you can now begin to train a model to say, you've never seen this, but based on what you've learned, do this task. And then they can fine tune it after they've done it. So this is a little more than a year later. So this is a much more complex robot.
This is actually a bimanual robot and it's trained on both synthetic but it's also trained on reasoning data. So Chelsea is going to interact with this robot.
If you-I had to cut it. But if you ended up watching it further, she built the sandwich and went, oh, really? I forgot to put a tomato on the sandwich.
So we remove the bread. The lettuce, went and got a piece of tomato, put it on the sandwich, put the lettuce back out and put the bread back out and said, "Here's your sandwich." This was...
Okay, so this is by Google. How many of you have seen Genie 3? I can't imagine how much money it costs to do this. It must be a shocking amount of money. You watch it. This is about a minute. I'm going to let it run. Peace. Every single thing you're seeing was generated by a language prompt and there's no image data, every single thing, and this is synthetic.
For transportation. Or even something totally unexpected. You can use Juni to explain real world physics and movement. actually grind. You can generate bullets with distinct geographies, historical settings, fictional environments, and even other characters. I'm excited to see how Genie 3 can be used for next generation gaming and entertainment. And that's just the beginning. We can help with a Body research. Training robotic agents before working in the real world.
We're simulating dangerous scenarios for-Disaster preparedness and emergency training. -New pathways for learning. Agriculture, manufacturing, and the world. We're excited to see how GE3's world simulation can benefit research around the world.
So historically, post the Industrial Revolution, if you look at the rise of what effectively was industrialized productivity, Generally speaking, We've had labor, the concept of labor, we've had productivity, and the more technology we've applied has increased labor productivity, and that's always been basically correlated with jobs. Wherever we go, workers become more productive and generate more money.
The problem with this is robotics doesn't enter the GDP calculation through labor. It enters it through capital.
So it is entirely possible that as robotics begins to take hold, where maybe the next source of labor is not the next human in a different country, it's the next version of a piece of hardware, a robot, that actually can do that task.
And if that's the case, it's possible you can see a significant increase in abundance, GDP, but it could be uncorrelated or even negatively correlated to put jobs. So you can already see some of the early telltales of this. Like for example, United Auto Workers went through a recent, last couple of years, went through a negotiation with the big three automakers. One of the things, one of the clauses they put in there, it's very specific, but they don't want robots, unstructured robots, within 12 feet of a canal.
We don't need to talk about that. I'm like, okay. They said, actually, through our union negotiation, we are not allowed to use generative tools in the media side of our business. I'm like, what? Are you kidding me? I said, did everybody else in the industry agree to do that?
So what's happening in the world right now is you're seeing organized labor governments that are like putting their finger in the hole in the dike and going, let's just prevent this from happening. This is Pandora's box. It's open.
Here, watch it. If you think how fast I just showed you what happened over a little more than two years We went from having no data in a completely different model to suddenly creating all the data in the world We need and actually creating robots that allow away mode to drive around in a level four way to have my Tesla I sliced the hell out of my finger on Thanksgiving and we have a bunch of cars in our driveway My wife got a brand new Tesla.
that was working on an autonomy solution for forklifts. So we partnered on this with Toyota. So Toyota is the largest industrial manufacturer in the world. When we started, the first forklift that we instrumented, the equipment on the forklift cost a million dollars. We have eight GPUs in the United Kingdom. We had five lives, each of which cost $180,000. We had stereoscopic cameras, we had radar systems, we had a huge amount of data.
Morphology they're working on uses one black well check, one color, from infinity. One. And then the rest of the sensors, the entire bond cost for that is about $25,000. and will likely be half that given to you. A million dollars down to $12,000. By the way, have any of you seen the cybercaps running around? You see the little gold things that Tesla's making? Two seats. If you drive around, you'll see that.
Um. Tesla with Waymo is using a Geely device. That vehicle costs, let's just say it costs $40,000, $45,000. The sensor pack they're putting on it, conservatively, is probably $30,000, $35,000. And the energy efficiency of that device is a certain number of miles per kilowatt. If you look at the Tesla, the cyber cab, the cyber cab, the building, it's estimated that at scale, they can produce it from between 16 to 18,000 dollars.
I think we need people to drive a forklift. This is just the tip of the iceberg. What I'm trying to tell you is that all of this capability is going to push itself down into the industrial workforce and the industrial supply chain. So you can imagine the next best labor you can get might be the next version of a robot. This is going to have a profound impact on This is why computer scientists talk about things like universal basic income. What do we do when we're living in a world of plenty, but it's not evenly distributed? How do we make sure that we don't have civil war and all the other-you already see it. My partner Eric got booed.
at a commencement speech. And if you read the text, you'd be like, this is really good. But there's just a bunch of students that feel disaffected, that feel anxiety.
Just like those 80-year-old women that I started the discussion with, they feel the same level of anxiety for the world they're going into. And if you go 50 miles outside of Silicon Valley, this bomb went off. We're all singed. Most people haven't even seen the flash of it. I hosted a team of 80 grad students from a big university in Switzerland. If I give my students, I'd say the ones that I've trained this year, probably a good seven or eight out of ten in terms of their capability. I would say that team from Switzerland was a one. They're not even thinking about it. It's not part of their daily journey to think about that.
In the vast majority of organizations, this is true within Google. So take the autonomy example. If I roll this out, and it fails. In this case, like someone dies, something really bad happens.
What happens to me? I get fired. In fact, that person from Uber got fired. So the asymmetric risk is if I take that risk, the penalty is a death penalty. Therefore, they will never ever take that risk. That's why you've just seen companies that are all running this in a sandbox because there's no benefit to doing it.
The other side of asymmetric risk, if I do this and it is profoundly disruptive, do I benefit disproportionately in my organization? And generally speaking, the only person that can typically absorb that risk is typically the CEO. They're the one that can make a profound risk because ultimately it's shareholder value that might be increased. So for many reasons, but one, the technology wasn't stable, but Tesla is moving faster because Elon doesn't seem to give a shit.
But at the same time, Google, you could imagine, If you got hauled up-let's say Google ended up on the cover of the Wall Street Journal, and Sundar gets dragged in front of a Senate hearing. Do you think that engineer is going to get fired? Hell yeah! You're gonna sacrifice somebody. So the problem that's happened in a lot of these is they risk reward. It hasn't been structured that people are going to put it out until it's perfect.
And we're already at level four, I would argue today, statistically speaking. In fact, Google published something last fallWeird. orders of magnitude better than it gives. Like if we just ran the numbers and everybody did autonomy, we'd save a huge number of lives. Is it perfect? No.
You can never be perfect in the kinetics of the physical world where weather changes, other people are not autonomous, you're going to have accidents. And the reality is that this is going to be safe. So what's happening right now in the bulk of organizations is asymmetric risk is not set correctly. Either you don't have the right talent or you don't have the right risk.
So they asked the first question, which is, what's the first thing you should do? I said, first thing you should do is, you don't have a technology problem, you have a talent problem.
I said, how many people are you competing for are also competing for jobs at Google or Meta or OpenAI, one of these? They're like, zero. I was like, would you spend a million dollars to get a world-class AI engineer? And the guy's like, oh my god, no. We're a heritage company.
I said, I can get those people. If you give me all of your damn... I'll do all the work to compute it and I want your big company, I want 30% of any value we create by either reducing the cost of your product or creating new products that allow you to win in the market. How many of you want to do it? What do you think happened? Zero. Zero? Why? Yeah, so, I can't do this. I'm-name your company. We're a big company. If I did this, we could be paying you billions of dollars.
Here's what's going to happen to you. Two years later, They're going to pay me a ton of money. I guess there's a ton of value to be created. Your board is going to go, what the hell? If you were asleep at the wheel, you're going to get fired. You know what the guy said? He said, I'm 63 years old. I'll be retired in two years. So that's the inertial amount of... Friction that you have to overcome is that the people that are in the position to make the decision likely don't want to make it because, one, if it goes wrong, it's on their watch. And if it's right, they won't benefit from it.
So you have people that don't have appropriate incentives to be able to do this.
This is why you see private equity waking up to this opportunity, why Blackstone just did a billion dollar deal with OpenAI, because they're looking at industry going, you all aren't gonna do this, we're gonna do it too. I gave a talk in London last summer for Bain. They acquired the largest outsourcing company in the world, broken in two, and the group I talked with, 86,000 employees. So I literally flew out there, and one of our alums is running the company, and I know him really well. I'm just playing for fun.
So I got there. I gave the talk. I thought it went really well. There was a roll-up of these founders that had been acquired and then went public. And then a week later, I met the CEO. We were in London. I met him in Amsterdam for dinner. I was like, how'd it go? And he's like, oh my god, it was a freaking bar fight after you left. He said, my team's like, this is awesome. We're going to do it. And the Bain people were like, if you fucking do it, if you do it, we will fire you.
If you do the plan, you're going to make a lot of money. That partner is going to make a lot of money. If you do this and it doesn't work, that partner is going to get fired. If you do it and it generates 10 times more value, they don't get 10 times more value. So Bain basically stepped in and said, no.
And Tomer Cohen, the former CPO at LinkedIn, talks extensively about this. He wrote a number of papers on something FSB, the full stack builder, and he talked about even in the context of working with Microsoft and Sundar, how hard it is to introduce technology into a company because all the antibodies in the company exist. So I would say for you, you almost need to create a sandbox.
that's separate from the normal things that are, I mean, most of your companies are designed to prevent bad things from happening, which basically don't allow good things to happen.
Because you know investing in your health today and tomorrow and the next day is critical because if you're not healthy, you can't be productive in the world. You will not have a high quality of life. He said, I encourage you to think about AI like you think about your health. If you don't work out for a day, you're OK. If you don't work out for a month, the next workout's gonna suck. If you don't work out for a year, you're not healthy.
If you don't use this technology every day, if you're not continuously pushing yourself, If you're not learning and you don't do it for an extended period of time, you are not going to be relevant in the world in the future.
You need to wake up and say, what is the thing I'm gonna do today? Before you leave here this week, how many of you built something with AI?