Hello and welcome back to another episode of The Edge. My name is Jasper. And I’m Lutz. Today’s conversation is actually pretty cool.
We talk about stock price and we talk about that we see a potential bubble in the making, but we as well put this in perspective of what this means for founders. The technology is amazing, still has gaps as we discussed, and we will dig into it. But overall, what each founder has is still a lot to work with. And Jasper will explain later on why this is at the moment the right moment for every founder to get started.
I enjoyed this conversation. Join us. Talk soon. You are joining us today in our regular setup.
It’s pretty live. So it’s summertime. I’m in Berlin. And Lutz, where are you?
Still in the Bay Area, close to San Francisco. I’m actually in Mountain View, like the home to Google and LinkedIn. Which is the perfect setup for us talking about the gloomers and doomers in the world when it comes to AI. Although, as you might have heard in the past, we are more on the positive side.
But I think there is something we definitely should address at this very moment. Let’s be realistic. We have seen this before. A lot of hype leads to an overheated market.
Are we in an overheated market? We had some new things that happened in the world. There was crypto. There was a little bit the internet coming.
I think in between, there were people giving loans to other people, giving loans to other people. And then you would structure that. When things get very complicated and complex, and then some people start making money and talk about it. And I guess then things just pile up.
Everybody jumps on it without understanding. Disclaimer, I made a course. Yay. And it’s out, actually.
It’s coming out now. It’s like a year’s work comes to an end. But in the course, I actually talk about the bubble from the internet area, right? We had internet coming around.
Everybody got very excited. And let’s look at, go to Yahoo and pull up the stock chart from Amazon. There is this huge bubble. And then comes a decline.
And Amazon is like over here, totally coming up. We had golden 20s. We had people coming from the mountains with gold in their hands. And then everybody was rushing into the mountains.
So it’s very normal. It’s very natural. It’s tried to get an easy dollar. I definitely see a lot of excitement in the market.
So there are a couple of signals that we see. And I guess talking about the internet bubble, as you mentioned it, stock market is very high. We’ve seen the Nvidia shares rising, then the drop, because obviously some people want to make some money. So you have to sell.
And then. And the price drops. But it’s definitely on the infrastructure side and the larger player side where we see a lot of excitement in the stock market because rising prices usually mean a lot more demand for the stock because people think the stock will rise in the future. So there will be a growth in revenue because on a long term basis, it’s actually cash flow.
But it’s kind of what people are paying for because I give you money and hopefully you pay it back at some point in time. And people are hence predicting there is a lot of cash coming from this new AI area. So essentially, Nvidia is about three trillion US dollars. Essentially, that is sure there is cost for creating the chip.
There is other things you need to do in order to get the chips into the market. But essentially, the amount of money people believe we will create by using generative AI applications and AI applications. But last time, as I checked, generative AI is, oh, yeah, you can like ask ChatGPT to write a poem and you can search. Meaning there is a gap between those three trillion dollars and what we really see as applications.
And I think the interesting point here is Nvidia is not a consumer product anymore. It was when I used the graphic cards. Actually, I couldn’t afford them. I used others in the 90s.
But nowadays, it’s really an infrastructure provider. And obviously, you only pay for infrastructure if you think you can use that infrastructure in the future. And it was once the most valued company in the world a couple of weeks ago or a month, depending on when you listen to this podcast. And usually up there are more the huge ones, the oil companies or the Apples of this world or the Microsofts that are pretty old.
They’re very sizable. Nvidia is also pretty old, but they weren’t as large as they have become in the last 12 months, actually. So now that’s a kind of predictive indicator. People are probably going to be more likely to buy from us.
And that’s a kind of predictive indicator. People are probably going to be more likely to buy from us. And that’s a kind of infrastructure. They’re building infrastructure.
They’re investing in infrastructure because they think this infrastructure will be used. Whereas before, we had the whole market cap push more on the Googles, not so much, but more Microsofts of this world that are already in the application space. So the question is, is like looking back to the gold rush, it’s not everybody, but really buying pickaxes and shovels. Do we have enough gold for all these people and pickaxes and shovels out there?
And that question will, at some point in time, get different answers, I would assume. So while I still see a lot of excitement, I personally would expect that at some point in time, people saying, really, do we have really so much gold for all those shovels out there? And my guess is we are underestimating currently how much gold is out there. Plus, the second thing that happened, at least in my VC world, there were VCs, where you are living, that started offering AI chips for founders because you couldn’t get them, the famous A100 chips from NVIDIA, for example, because they felt if I have that chip, I can compute better, faster, and I can be on the forefront.
I can build my models, my applications, maybe, but it was mostly on the infrastructure side faster. Now we have the B100 or the B200 chip. So there’s really, there is innovation happening, but it was more on the constraint side, buying hardware. Obviously, now we have this people still buying hardware because they’re afraid of they wouldn’t get more hardware in the future.
But we hear from the market, this is also cooling down. There’s actually way too much hardware. There’s way too much compute, which usually leads to effect where we are currently not at the moment. You have too much supply.
What happens next? Prices dropping. Which is something I think we spoke about it last year, at least a year ago. We were kind of expecting in this hype.
Going forward. So there are a couple of calculations. How much over excess of hardware do we have for offering of compute? Some people say it’s 600 billion.
Some people say it’s 300 billion. I think you can do all these kinds of calculations, but it’s actually pretty cool because people have invested here. They have to get their money back from their investments. The stock market is expecting revenue.
That’s why all the stock prices have risen. So now we might come into the phase where you as a founder, you want to ask, you can get access to all this hardware and compute power for a much better future. So that’s a really cool thing. Yes.
So this is actually good news for the founder. Also, if you have sufficient funding, you come into those crunch times that often gets along with you getting better access to talent and so on and so forth. So there’s good news in the correction of the market. Now, note, I believe there is a correction will happen.
I don’t think it’s the right correction. I think the actual value, which we will see, is way bigger. The second thing that happens, and we also hear that from various sources, by the way, it’s the B100. I checked it.
Sorry, NVIDIA was preempting something there. People realize you don’t need these amazing new chips to do all the calculation. There’s a lot of, and I think we spoke about it two podcasts ago, a lot of deep learning models you can train on the old GPUs. Maybe it takes a bit longer, but it’s fine.
So while we get more into the application of AI models, we’re going to talk a little bit about the AI models. So let’s get started. use cases. But we as well said that those models can’t plan, those models can’t reason, all of this is still true, right?
So having more capacity essentially allows models to have a bigger context windows, take in more data, all of this is good. But it has, I suspect, in fact, diminishing returns at some point in time. So that’s a little bit why we could see the market actually reacting. Now, it’s not only the financial markets, right, Jasper?
It is as well, we see a backlash from the industry as well. Yeah, I mean, on the VC side, you probably have heard all the mega rounds for the foundation models, we spoke about it. We still have teams focusing, for example, on orchestration layers, that get a lot of funding. We see teams that go on to the next generation of AI agents, or maybe even models that can reason and plan and which are more research projects, crazy funding rounds, to my opinion, very impressive.
Don’t talk about my idea here. We don’t talk about your idea. But essentially, what happens is it’s a bit I mean, you’re mostly investing in research. And we spoke about, you need the data first, you have to figure out what the model is doing, you have to figure out how you control the quality of the output.
Same for the application. So same as the stock market, people were preempting kind of future rounds and future value increase without knowing if that would actually happen. So the backlash that comes here is that we see some realism, oh, wow, it takes longer. It’s a research project.
Oh, wow, they don’t get the data. Oh, wow, the model can’t actually do it. We thought it could. So the follow on funding without revenue traction gets very tough.
And we hear also from the valley that many of the amazing companies with amazing data, they don’t get the data. And so we see that there’s a lot of money that’s being invested in. And we see that there’s a lot of money that’s being invested in. And we see that don’t have the traction.
And you’ve seen very early exits to Google to Microsoft, if you just look it up with the teams, I don’t want to blame them, and they probably got good money for it. So we see become more and more careful investing at an early stage in teams that a have no practical application experience with AI. So I’ve done research a lot, but never used it in practice and production. And secondly, be teams that don’t have commercial understanding here.
Because just look at the data. And you see that there’s a lot of money that’s being invested in. And you see look at Jensen Wang of NVIDIA. The guy is amazing.
He’s really smart, but he’s an amazing salesperson as well. He’s just standing there and selling this product and telling the world that this hype is happening and getting into something real. That’s why everyone is buying the chips. It comes down to application, right?
This is when we say it’s a time of the founder, it’s a time of the PM. It’s not so much anymore the time of the research currently. Yes, research is still a wide field and we will see innovations. But at the moment, all the innovations we have seen, they need to put down into applications.
Yeah. And I think the interesting question is, and I don’t have the answer, maybe you have an answer. Do we really need more innovation? Or is it fine to say, hey, large language models are very good.
You can chain them, you can combine them with traditional machine learning techniques. And then you actually build an application around it. And I don’t want more innovation. I just want more practicality.
I want more tooling around it. And I don’t need the next foundational model innovation. You said it very nicely. I think in the VC world, we actually see that the appetite to go on risky, more researchy projects has gone down.
It’s the VC world that says, I want more applications. Well, I, as a professor, obviously, say let’s push the boundary. There should be more in it, right? There’s more to come.
Create a model that plan, create a model that can reason. And there’s a lot of very, very good efforts ongoing in that space. But I think from where do our audience, meaning the founders should focus on, it is really take the existing capabilities and apply them. Now, the application actually sees as well as the application.
So, I think that’s a really good question. Well, a backlash. Rolling Stone had this article about brands don’t really jump on, for example, the gen AI creation. I recently was in an airplane and the guy next to me, he created a platform.
Things happen when you’re in an airplane, right? The guy next to me, he created a platform for image generation or three-dimensional image generation. And he offered this and worked very closely with meta. And I was like, wow, this is cool.
You probably see a lot of generative art or generative storylines used. And he’s like, no, most people do this by hand. I was like, why? And he’s like, well, because the customers, they’re risk averse.
They don’t want that new stuff. Yeah, at least they don’t want to see it because they’re also afraid. And I think it’s what we see from the startups, but also if you look at Firefly Adobe, which I hear mixed feedback, it’s a different way of working. It’s a different way of being iterative.
You get results, you want to change it. Whereas when you go from the scratch, I have something in my head, I want to start drawing it. That’s my process, right? That’s what I’m used to do.
And now if I tell that in a prompt to the AI, the AI comes up with weird stuff. I have to do it again. It kind of makes it up in my head with my, maybe I don’t like it. So, but I could probably utilize it, but I’m just not used to it.
And I think we now come to this point where it’s a bit the discussion why this did cars happen? Did you ask the people that would sell horses how to build a good car? Are the people using horses? Obviously not.
I mean, we all know this now, but there were so many industries dependent on it. And I have to say, I listened to a quiet podcast, two episodes, Hermes and VMH, amazing of the guys. Amazing guys. I really have to say awesome podcast.
Louis Vuitton started with something for horses and carriages, and then they started the square cases because there were trains. And obviously you could put that in a car and then MS started with saddlebags and then you would put them over your shoulder and they got smaller. And then they realized, Hey, I can do handbags. People want to carry around their stuff.
So this kind of innovation has to happen now. And then you can sustain as a company, even a very old one, but we’re not there yet because, and I think we now come, we spoke about different layers. We now come into this phase of, we know what the models can do. We know where the limits are and we really figure out the pick axes and shovels, how to utilize them.
And then we can do that. You remember our discussion when we said we looked at so many startups last year in 23, and we just didn’t know if these pick axes and shovels were useful. We didn’t know about applications. And now when you read about what bedrock, what AWS is doing, I mean, they were betting on several models, meters going more open source, but all these people, I think they were very smart saying, we don’t know yet, but if we know we will definitely build the whole tooling space around the infrastructure.
But let’s wait. Which essentially everybody kind of starts going a little bit slower and trying to saying, how do we actually build this out? I have an interesting story. As I created my course, I built a image model with the students online.
So that’s in my course. It’s a Laura. It’s kind of like you take a normal diffusion model and then you adjust it so that everything becomes legal. So you legalize your world.
So imagine a coffee machine, you get a coffee machine made out of Lego. Imagine a house, you get a house made out of Lego. And I was like, okay, who would need that? So I kind of Googled on the, or I went to YouTube and looked for Lego.
And I found, I found somebody who actually made complete trailers from famous movies in Lego. Whoa, this is so cool. But in real Lego, right? So I talked about it and said, look, we can do this with AI.
And I pinged the person and the person very nicely wrote me. A long email back saying, you know what? I have heard about this AI thing. I worked a year and a half to create it in Lego.
And I would not want to do it any other way. I have so many questions about why you should do it with AI. So there is this user interface, as you just described, the iteration that is hard. And for when we talk about one of our principles, like investment principles is like focus on the workflow, right?
When we say workflow, what do we mean user interface? We mean figure out a way to make the human today more effective. Yeah, you have to make the value of it accessible to the everyday person. And obviously, crossing the chasm, you start with the early adapters, and maybe not the Lego guy, but the guy who does it with molting stuff and sculpture it.
And that’s your early adapter. And then you, you deliver the tools a bit later to make it more accessible for other people. And I think we’re getting there now. It’s just still a very much experimentation.
What I really like when Apple launched their Apple intelligence. Yeah. Right, the AI. Amazing storytelling.
Amazing storytelling, but as well, very good integration. And that’s like what we all would expect from Apple, be able to make an interface, which feels not like the AI, pin, like weird, or like the early GPS system, which whenever you took a wrong terms and wrong turn, like no, but integrated that you can actually work with AI. And I think there is so much more to be told in this story, find out what is good UX, find out what is good interactivity. So that’s what I’m looking forward.
And if you’re out there, if you’re looking for a good interface, and you have a nice interface, tell us we want to hear because we see the market is starting to cool. It’s too early, but starting to cool. We see that people have invested heavily in the shovels. But the interface fits a huge space.
Apple is a nice example. Think about how you can make these applications easy. We had in our previous podcast, we spoke about product with AI. Remember, the first Apple photos, when it was recommending photos to you, and you would click on them and say, Yeah, that’s really me.
It’s computer vision, right? Or we invested in companies that would do email automation, it was first recommend an answer, and then it would give the full automation after I click the same answers various times. But you transfer this experience of I trust the AI into a different workflow, but you don’t do it on the spot. We are getting there.
Now, I think that’s really interesting, because the AI is more and more predictable. And people know how to control it. What is your favorite? If I may ask, what’s your favorite generative AI use case at the moment?
So for me, for example, scribes, medical scribes, for me, this is an awesome use case. But it’s a product. The product is you go to the doctor’s office. And normally a doctor looks down on their keyboard types away.
If it’s a doctor on a like, if you go to Stanford, they have them the silently sitting students. And in the corner, which types away. But the actual impact of a doctor is that they talk to you that they spend time with you. But the weird thing is, I get it.
But the weird thing is we have that product. It’s nuance. It’s the nuance product, but they didn’t use it in a live setting, right? It’s kind of talking about your interface.
Again, they used to make notes or whatever. But they never used it in this live conversation. I invested in a company called verbat where they did it during the position, recorded a position. And then they would transcribe in parallel.
So to your point, but the AI wasn’t as accurate enough. And I think that’s the point here. We’re getting to a level of accuracy that you can trust it on the life and you don’t need the transcriber correcting it in the back, but yourself checking it. Totally.
We all had been joking about AI stupidly going wrong. I definitely did a lot of times. But now we are at a quality level. We talked about before about the minimum quality product, not the MVP anymore.
It is the MQP. And we reached now a level where this is feasible. So I’m very curious, like for me, the scribe thing is really something which going to change the way we interact. What’s your favorite application?
Is reading papers with JetGPT and getting summaries. I love it. My second favorite application is I have a favorite application. I have a favorite application.
I have a favorite I found a meeting. I’ll talk to the 4.0. So GPT again, I’ll kind of make notes and then it summarizes my notes for me in a structured way. That’s pretty cool.
I actually think one of the changes which we will see the whole summarization, right? We need to find a good way to control it. As I wrote my course, right? I obviously use JetGPT or you.com or other tools.
So I think that’s a good way to actually get to information. And lo and behold, right? What happens is it is okay to tell you the typical structures or the typical things that works. What does not work if talk about dangers of AI, talk about bias, you just get whitewashed information.
So as soon as you want to be very specific on an area, it becomes more complicated. We are at a point here where the future has to, prove us right or wrong again. To sum it up, at least my view, yes, we have a lot of investments piling up, which is great for founders. It’s might not be great for the investors, but that’s, I mean, that’s the risk investors are taking.
So people are expecting a lot of applications in AI. And at the moment, this might still be very expensive. But since there’s depreciation happening, utilization happening, same as for every factory you’re building, over time, this will get cheaper. So this should be something for founders to be able to do.
And I think that’s a good way to sum it up. We see an impact in the second half of the year, prices dropping, all these hyperscalers are looking for applications. So you as a founder, this is the time when you can make good deals. Even don’t be too concerned.
At least that would be my estimate. Don’t be too concerned about the compute cost that you see at the moment. They should drop massively in the future. All right, everyone, there is a thunderstorm coming in Berlin.
So I think we have to end the podcast. And your chickens sound very hungry in the background. Let’s really enjoy the conversation. Thank you so much.