You asked, we answered

Hey everybody, for another episode of Just Bandits on AI and questions you had. This was awesome. You guys asked us so many questions and we’re trying to address them. We go into quite a lot of discussions.

We discuss Hollywood strikes and we discuss the backlash of society against AI. We talk about high-skilled knowledge workers versus frontline workers, which I thought was a fascinating discussion. We talk about VC hype and why so much money flows into it. We talk about investments, US versus Europe.

I mentioned a little bit that we had somebody from the Bundestag today here and we give an overview of is EU an AI region. We talk travel and business cases. We do talk about our loved area. Is there a moat in large language models, which is pretty awesome.

And we even cover infrastructure. It’s a mouthful of things, which we had today. Amazing. And it was so much fun.

So I’m pretty sure you guys will enjoy this episode. Let us know. Keep the questions coming. We will.

Hello, everybody. Hello to another episode of Just Ben Looks about all the questions about AI you never have to ask. And this episode is about your questions. But first of all, Jasper, how’s your coffee?

It’s actually pretty bad hotel coffee. Salt is actually pretty good for bad coffee. You can just make it a little bit salty and then all the bad tastes go away. So first of all, thank you so much for joining us.

Thank you. Thank you so much for the comments, what we should improve, what we can change or where we’re good at, actually. But we also got a lot of nice questions from you. And we did a shout out to the community asking you what we should cover in the episode.

So some of the things we will use for future episodes, continue, please continue doing that. And some questions we thought we should answer right now. Yeah, we will just read the question and then we will have an opinion. And thanks, everybody, for all the questions.

Pretty cool. Most of AI applications are currently designed for highly skilled knowledge work and are tailored to their specific use cases. Where do you see the potential of AI applications for operative workforce like frontline workers? What must be given that an AI application can create long-term value for the frontline broker segment?

Yeah, and that question comes from someone building something there. So that’s great. It’s a start-up question. So thank you for that question.

I think to start off, the question is what’s different here. Frontline workers move around. They don’t sit in front of a computer. They might not have a university degree, which might not be an issue.

But still, it’s a different type of persona doing a different type of job when a knowledge worker, maybe a researcher, journalist, however you want to define them. I think what’s interesting there to kick it off with is our debate around the interface. Because a knowledge worker is very busy, maybe at the machine, maintaining something, maybe at a shop, talking to customers. And then you don’t have the time to sit down on the laptop and do stuff.

But you might have a quick second minute to speak to something or type in a question and get an answer. So you could support these people, for example, in their maintenance work because they have a question and they would ask them in their language. The AI translate, not just put the question in and ask the knowledge database, but can probably also even transfer that language into technical knowledge and find relations. That’s something you could imagine.

Obviously, you have to make sure that the answer is right. So coming back to our feedback discussion, you don’t want to use just let’s get take chat GPT, put it on your FAQs or your technical documentation, and then let people just ask questions. They would have to test this a lot and make sure that the knowledge workers are happy. So probably a human in the loop at the beginning, a little bit what Reto explained from Ultimate AI in our interview, is something that would be very helpful here.

But I think it’s a good idea to have a knowledge database. I can imagine this could work. I thought about it. And interestingly, for the German audience, we have Anna Christmann from the Deutsche Bundestag.

She actually came to Silicon Valley this week today. And I had, there was a group of Silicon Valley folks. We talked about AI and we did a round, round robins. We did a round panel and asked what everybody in the room, this was venture capitalists, professors, startup people.

So like a very good group. And we asked about, what do you think about AI? And they said, what is important for you? What do you see?

And very often the idea of summarization and logical reasoning as the core idea behind it. There’s one question later on for education, but that’s exactly the key when we come to this later. But the exact idea is summarization and logical reasoning. So the person who ever asked the question about a tool, which is currently used for summarization and logical reasoning.

And essentially, it’s a tool that’s used for summarization and logical reasoning. And it’s a tool that’s used for summarization and logical reasoning. And that has a participants participants Hold on, that is for highly skilled knowledge workers. Yes, because summarization and logical reasoning is a task, which is very often for knowledge workers needed.

But if you are a frontline worker and you have to take a certain decision, you might pull up the manual to look it up. You might call the next level. You might get permission. You might need to ask.

I mean, there are so many processes which somebody might need to follow, which is knowledge retrieval. So for me, LLMs are knowledge retrieval thing, or like a summarization thing, an interface thing, however you want to call this. And at the moment, because of the limitation of the use case, which is mainly chat GPT or very often chat GPT, it became something which is knowledge worker. So I was completely surprised.

I think it’s a fair question, but I think we will, we’ll see way, more use cases. And I’m very curious to hear what company is out there who is thinking about other workforces. Very cool. Very nice.

Cool. Then let’s go to the next question. Okay. Could you please explain the reasoning and economics, wow, behind CBC’s participation in very large free product rounds?

For example, Mistrock. Don’t they mostly just finance CapEx GPUs? Is the dilution of resources, the!! I think it’s obvious, as we discussed, you need a lot of money, compute power just to train those models, the foundation models, if you want to build a foundation model.

And there is a lot of discussion around it. This cost will go down and so forth. And by the way, it’s not just compute, right? It’s also the supervised fine tuning that you have to do.

That’s also pretty, pretty expensive. You need a lot of resources for it. We spoke about it when it came to the output of the models, the biases and so forth. Now, so you need that money if you want to be competitive.

And I listened to some nice discussions. There from Stanford professors and other people about the open source models and closed models and the closed ones have an advantage there because the open models are never that precise. Technically, they might catch up, but the output itself, because of that money that goes into the training, they just can’t afford it. So yes, it’s true.

You have to invest in that. And that’s kind of CapEx because you do an investment and you hope for a return. And we know that these corporates out there are heavily investing and so giving it away, Yeah. And that’s a huge participants participants So it makes a lot of sense to do that.

And if you want to catch up, yes, you need a lot of money at a seed stage already because some other companies have already created quite some value and with quite some money. So that’s the first part of the answer. In certain areas, if you don’t have a product and people pouring money in you, you know, these things might be just wrong at some point. Let’s say.

So for full disclosure, we’ve never invested in a foundation model at Sharing Ventures. So to the first question, so what is actually happening there in the ICs? So we discussed it in our IC. Obviously, I spoke to other VCs.

At the end of the day, there are not many examples that you can earn a lot of money with this. And it’s also very hard to foresee the future. I mean, we tried it also with our predictions and in our discussion, what people are hoping for, that you create a value of this foundation model with great output because they have seen it with JetGPT. And then some corporate, large corporate.

So we just snap it up because either they need it for the equity story on the stock market because they have no other idea what to do with their money internally. So they hope this is something. And then the third part is because at the end of the day, this is a hype, as Luke said, and people just don’t want to get left behind and they don’t feel they can build it themselves. So that’s why they would just buy it.

I think, like for me, these are two things. One is there is something happening in the world and we can have later on. There’s another question. This is a question to business values.

And to the person asking this, we don’t know yet how value is created. We don’t like we don’t like, you know, you have always a discussion with shovels and gold. And so we don’t even know what your shovel looks like. We don’t know which is actually becoming dominant.

Pinecone attracts a lot of money. NVIDIA attracts a lot of money. So certain things we know, other things we don’t. Right.

So this is a little bit the area. In these situations, venture capital looks for teams, making things happen. I can give you a good example. Element AI.

At one point in time, I had a discussion with them and the name behind Element AI is Joshua Benin. Joshua, the guy who is one of the three essentially of machine learning and AI. So he’s the only one who is in an academic setting. Everybody else went to one of the big corporates.

And he created Element AI. And overnight people poured in money because they said, look, he’s doing this. He is attracting smart talent. He’s going to figure it out.

Sad story, he didn’t. But that’s a different story. There was this hope. So Mistral actually said, well, we will make AI useful.

Well that’s something worse. I don’t know how you do it, but if you attract a lot of smart people, then let’s do it. So yes, there is a hype, but there is as well a huge opportunity. And because we have this huge opportunity, there are bets on people.

And if those don’t turn up, you do a fire sale and you sell the people. And maybe for the listeners who don’t know Mistral, that’s I would say basically Anthropi or OpenAI in Europe. So mainly French team. There are obviously other teams out there.

The idea is a little bit that you have the European version compliant under regulation and so forth. Let’s see how that turns out. But it brings us to the next question, Lutz. So talking about the West Coast, we got the question, Hollywood strikes because of AI.

What else can we expect? Will it be harder to create work that is worth paying for? And as we all know, there’s a lot going on. Lutz, maybe you have also some details, but people are on strike.

People are demanding to be paid for work that the AI is doing. People are obviously afraid that the AI will take away their jobs in Hollywood. As we all know, there are some people who earn a lot, earn a fortune and have yachts and jets, but a lot of people do hard work and don’t get paid that much. It’s actually interesting.

So this question actually came from the East Coast. Somebody from the East Coast asked it, but it’s somebody from my Cornell students. Now I as well had, just by incident, by chance, I actually posted about that we see the advantages of AI. And somebody in Europe, I don’t know if you know, but somebody in Europe, I don’t know, but somebody in Europe, has suggested that AI has a huge participants participants participants participants participants participants participants participants participants participants participants participants participants Whenever new technology comes onto the scene, it changes existing workflows.

It changes existing work. And that creates, obviously, friction. And that creates fears. I think the speed of what we see here at the moment is what creates a special level of fear.

But that we see people complaining about it, it’s nothing new. I looked it up. There was a strike from the writers and actors, actually, in the 60s, as the onset of TV. And then there was another one as we got the home video onset in the 80s.

And based on that strike, actually, I think it was like Ronald Reagan was very active in the strike. Actually, we got the whole idea of royalties for second screening and so on. I think the interesting thing there also is if you just imagine how films, like one made 100 years ago and what evolved over time. I mean, I’m a big Star Wars fan.

If you look at how they did the first three movies, it’s basically somebody holding a little model of a spaceship flying around. This guy is probably not doing that kind of job anymore because it’s done by 3D and computers. And if you look at the whole sceneries that have been built, people really building sceneries, they still do. But it’s much, much less because you can also have sceneries made in the computer.

But I agree with you. I really agree. It’s extremely fast, the development, as we know, with generative AI and the people are not prepared for it. And I think that’s the big fear here that you don’t have the time to adjust.

However, overall, it is like I’m not super worried about the actors because of the way we have royalties played out. So Carrie Fisher is later in Star Wars, right? And if I recall, she had a dinner speech. Where?

She compliments her look every morning when she looks into the mirror and said, I wish this would be mine. Essentially making a joke about the fact that her look are actually sold off to a company. And as we then saw at later Star Wars movies, she gets reused her image. So I foresee that there’s a future where actors brand like this is a brand and you carry on a brand.

The Barbie movie is now in this theater. I haven’t seen it. But the Barbie movie is in there. Barbie is a brand.

And you recreate and you repurpose the brand. Yeah. And what we should not forget is AI can only repeat something from the past. So even stable diffusion generating pictures, it’s a mishmash of old things.

That’s why I don’t know if you realize when you use mid-journey a lot, it’s always kind of the same theme. You try to change it, but pictures become similar. But an actor, and if you also have a story. That’s about timing.

That’s about the right mimic pace that you show at the right time. So this is something more difficult. But the people working in the back, the people also supporting the movie, the long list of people names that are coming. There’s probably something that will be automated and something to prepare for, I think.

I would challenge this. So I think you bring in two ideas, right? Idea number one is the human creativity is still key. I’m not sure this is the case.

We still we see in. There’s a lot of development in AI, even if they’re mixing things where there is creativity as we would define it as humans coming through generative AI. Now, same with mid-journey. You reference mid-journey, but mid-journey is just one generative model trained for a specific purpose.

If you go, there are certain platforms and we should do it one point in time, an episode where people train their own platforms and they create a generative AI on a Japanese manga or on only one certain stuff. I agree. I agree. We also saw.

Creativity. I agree. We also saw some actually a couple of startups, which I think are great ideas where people would say, this is the corporate identity of your company or this is the style you like. And we have different models.

It comes back to the fact of fine tuning using racks, but controlling the output of the model in a way. Yeah. And we talked about it in the future of Indie, right? That Drake and Vance.

The weekend, which was AI generated stuff and Drake complained about it. But I think there will be a trend where Drake says, this is the identity I represent and let it live on and let it create in his approved name, sort of thing. And then if we have such a model, actually Hollywood showed that it works perfectly fine. It’s secondary screening, screening on the home devices, screening on streaming devices, and you have royalties.

So it’s a lot of different things coming in. So for that area, I’m not super concerned. Yeah. Still there are people and again, to the example of the guy holding the spaceships in front of the camera to imitate them flying, there will be automation coming.

We know this. But it brings us to the next question, Lutz. So how can we prevent European startups from losing out to the US in terms of investment sizes when it comes to AI and thus falling behind? Maybe to remind everyone.

Because of the previous discussion. If I got it right, it’s a bit confusing, but Mistral raised for a kind of a pre-seed idea and team. That’s what the press says. More than 100 million, maybe 120.

Maybe there was a lot of compute. But then we have other companies like Anthropic. I think there are more than a billion that they have raised now. Inflection AI, we discussed in the podcast, 1.3 billion that they raised.

That’s a huge difference. So this question really makes sense. So 10x and then plus there are many more startups out there. So what can you do to make that?

First of all, remember, I would never invest in just a pure large language model because we need to make business with it. And as we saw in the last few months, everybody and their dog would publish large language models as open source. But that is not it. It’s a fun question.

If you say US large language models versus. I don’t know. If you say US large language models versus, and now you’re kind of trying to complete that sentence, nobody would come up with the word Europe. So the question, Europe isn’t currently not on the list for the AI companies.

People think about China. There is a lot of discussion for how there is a competition, but Europe is not on it. And the question is why? And we have talked about this in this podcast before.

The machine learning idea. Like if I have 10 points in a line, I can put a line through it. If I have two points, I can put a line through it. Right.

There is a mathematical idea of creating a line. That is not protected. That is out in the open and people know how to create a line. Machine learning is obviously very simplified, but is something close.

What you need are the points. No points, no line, no luck. Now Europe. Now Europe has protected data from their citizens.

And I do not want to comment whether this is a good or bad thing. I actually think protection for citizens, protection for users is a very good thing. But in line with that, they made access to data harder in many respects. Which leads again to more people not being able to create on top of this.

The setup systems. And for me, a very telling discussion was while the US very much thinks about how we make money with large language models. Europe started to come up with the discussion about how do we do privacy concerning large language models? Or how do we ensure that only IP rights, correct IP rights setups are used?

All of the large language models fail then at the moment. I am not saying that. I am not saying that this is wrong. I am only saying that at the moment there was a group that did not give a damn about all those models.

And they were not in Europe. And then there is a group that now thinks about how do we do this correctly. And therefore, it is not on the map. So adding to Luc’s opinion there.

We should just be a bit careful asking ourselves why this is happening and what are the benefits. So how can Europe catch up? My opinion. We saw it with the cloud providers.

So a lot of discussion around is this secure and so forth. But it is there because there is a lot of value for businesses in the cloud. And there were the European Union was, yeah, there was a lot of activity but it was not around business. And we see where we ended up nowadays.

So apparently, the European approach at least for the cloud was not the best one. So that is one thing to consider when we have now this discussion. So I think the important point here is yes, there is an advantage in the US. But it is not the best.

There is an advantage in the US. They have more money for venture capital. We all know that. But that does not mean you need more venture capital money in Europe.

You definitely need more of it. But if you create that from the state, the question is always what kind of people are deciding who gets that money and who is not. And we tend to work too much with subsidies. That is my opinion.

And we could ask the people who will make money with that where they need this kind of money. That could be something. So we have a lot of money. We have billions for AI now dedicated.

We have billions for other projects. The question is where does that money go and is it productive? Is the output great? And we alluded on many examples.

So maybe there the US way was more successful. So this is one thing. How can we catch up? Yes, we need money.

But also we have to decide where this money goes. And we have great universities, great researchers, partially created stable diffusion as a project. It is just a question how we can help them to make businesses out of that. That is my opinion.

Yeah, I agree. By the way, the last one is complicated. As we all know, the underlying large language models, BERT was first pushed from Google and they completely missed that opportunity. And they are obviously, as we know, a US company.

So it is not only being the ecosystem. It is not only being money. It is having the right moment and right idea. And I think OpenAI was a little bit surprised by the effect chat GPT had in the open market.

They did not really know how to deal with it initially. Next one. Travel. I love this question.

So which travel startups with real AI approaches, whatever real means, do you have in mind in Europe? So I think it is tough to answer this because there is a lot of at this moment in time because obviously companies are developing a lot of things with AI right now. So we should not talk about concrete companies. But what we see and we discussed.

And we discussed many things around automation, AI agents, content creation and so forth. If you just think of the value chain in travel, there is a lot of cumbersome manual process steps that are being done. I am just talking about booking a flight. If you think about the Amodeus engine, how you type in information, how you retrieve information from that engine, how you then present it to your customer.

The whole interaction with the customer. A lot of call centers still. Even the modern companies, the startups that have been started. So all these process steps are currently being heavily investigated.

What you can automate without changing the systems too much. That is great about large language models. Information retriever. We spoke about this knowledge worker approach.

That is the same thing. And also our interview with Retu Ultimate AI. I think it is on the website so I can disclose it. They have travel companies as customers in their portfolio.

So that is the, I would say, more efficiency part. How to make this whole process. How to make this whole model just better. Which also leads to better customer experience because of the interface.

You get the information faster. You have more information that is more customized. So this is what we will see more and more coming up with the right companies. That from the front end side you just have a better experience.

But then also in the back end side you just reduce a lot of cost and you are faster. This is actually a super interesting discussion. And I am pretty sure we will see their startups. Whether it is.

Planning for a vacation. I mean you can already today ask your favorite large language model question about. Should I hike Mount Fuji in winter? Or like I just was in Japan.

And if I look for, should I hike Mount Fuji or Fuji sun as the Japanese call it during the rainy season. The answer is no. Right. It becomes cold.

It is cloudy. And it is really not a pleasant experience. Now. A large language model can have that information already.

The. What you need if you think about travel and you think about information retrieval. There are travel companies already trying to make your travel experience nice and trying to sell you experiences. Now this is now a sales job.

You will. The sales job is a job which is information retrieval. Plus some. Argumentation.

Argumentation logic. Okay. Information retrieval and argumentation logic. That’s large language models.

That we can actually automate. It’s awesome. The only problem is large language models at the moment do not have an agenda. And it’s very hard to get an agenda in.

So what we saw for LegalOS by the way they just launched the new feature. Awesome. But for LegalOS they talked about creating guardrails to make sure that the information they give is safe. Okay.

the information they give back is appropriate. We talked about routes last time and how they are, like how the technology looks like. What we haven’t seen so much, and I’m pretty sure that like many startups at the moment trying to do this, is creating an agenda, a patient flow. This is how you convince somebody.

Now fill in the points and help me convince the person to do the Mount Fuji hike. For example, right? For whatever it is. So the sales LLM, the LLM targets, whatever you want to sell, that will be the startup I want to see.

So let us know if you’re building on one, but that’s a good one. And we know this from sales, there are playbooks. So yes, some of the part is interaction between human beings and empathy, but there are also a lot of playbooks so you can convince people. Good.

Next one. It seems as though many ventures you are… you see are using the same foundational models or combined of such. With this being the case, what are some effective approaches to building competitive modes around the solutions utilizing these AI models?

And how do you think about these protections as you assess early stage ventures with commercial promise, but are at risk of replication? Yes. That’s a long question. It’s a long question.

Now, we talked about this many times. Open. AI in itself is not protected. This is as Google said, oh, I’m not concerned for OpenAI.

I’m more concerned about all the open source models out there. This is now you see so many different open source models, as we said earlier. So I don’t think it’s protected. However, let’s talk chat GPT.

Chat GPT is actually a huge consumer application though. Underline is OpenAI’s large language. The! which compared the different answers from ChatGPT on prime numbers and other things and saw that it’s degraded.

But still, people stick to it because only a little bit of degradation isn’t too bad if this is your trusted consumer service. Yeah, plus what we should know is that it really depends on your use case, what you want to do with it, and so forth. It’s guaranteed for a certain output. That’s what you pay for.

They worked a lot around those guardrails. They worked a lot around the output. And at least what we hear is that other models are also not too bad, but the output there is still better. Plus, OpenAI is working a lot on reducing the cost.

They also say this openly, that this is one of their targets. Yes, it’s marketing. But people are obviously also hoping for costs going down. That’s why they still stick to it, which makes sense.

No, but what we are looking for is the application. We’re looking for the solution. We are trying to understand what people are solving there. And if there’s a need for it, and if AI makes sense.

And then we look and we discuss it in one podcast for the technical teams, being able to switch between models, being able to assess the quality of the output of the models, the feedback loop, and then the whole package makes sense because you’re not dependent then on one model’s performance and cost. One of the listeners, a person I highly respect, wrote us the following question, which actually feeds into this. When we talk about Google and other big tech players having no mode, this is like, yeah, they are like the model that’s not protective. How are we factoring in their attempts, progress in developing in-house AI chips that are tailor-made towards executing their models?

You tell us. Well, I mean, the question is great. The approach is great. Nobody wants to be dependent on video chips, although they are great chips.

But at the end of the day, if the compute power is the worst thing here on the cost side, and it is, and then comes the cost. The supervised fine-tuning, then this makes a lot of sense. And I actually think what we see here is we are working upwards the stack, right? We see that there is an application like ChatGPT.

We see all the potential promises. We don’t know exactly now how this works, but we do know chip power is needed. The next thing we know, storage power is needed. The next thing we know is, hold on, we need observability platform.

So we go in stack level up. And you see in those stacks, slowly, one winner after the other is decided. And while on the actual value creation top layer, currently only existing businesses like Microsoft, like Google and so on are making inroads. And we see a few new players like LegalOS or others to actually push stuff out.

Yes. Yeah. But we saw the same in the cloud development. It makes a lot of sense to go on that.

And we saw in the hardware development, we had some approaches around the world with chips. Some companies went bust, some companies are pretty successful. But I also think we spoke about AccelerA from the Netherlands when it comes to edge computing. This will open up the field for a lot of companies, at least to try and attack the big ones, because it’s a different kind of application.

Cool. Like LLMs on the edge will be super interesting. And you saw Apple now have their own LLMs. And I don’t know whether you realize, but Google, hey Google or Alexa or Siri, they all sound so stupid.

Is it just me? Or did they really degrade in quality? Or do I just expect more? I heard from some friends, which I don’t want to name, that at least in some of the big companies, essentially the development on those bots have stopped because people say, well, let’s work on the LLMs and use them.

So Google is the best. Google is the best. And hopefully, Alexa, Google, they will come back and start being your wishful help hopefully. Let’s hope that.

So for now, thank you very much for listening, everyone. We already come to the end. There are so many more questions that and we said we would talk about education. I think that we have to do that in another episode because I have to head into the mountains now and I guess you want to go to sleep.

It was awesome, as always. Thanks so much.