Welcome, everyone, for another episode of the edge podcast with Lutz Finger in the Silicon Valley and myself, Jasper Masermann here in Berlin. Today, we talk about what the bot. So a lot happened in the bot space, different companies, different stories, but all pretty exciting stories behind those companies and bold ideas, a bit struggling, some of them. And it’s also opening new ideas.
So we talk about where the struggle is, where the challenges are, but actually where the opportunities are. And I think Lutz has also a lot of practical knowledge and experience there. So I’m pretty excited for this one. Hey, Jasper.
Good evening. Hello to Berlin. Hello. Good morning, Lutz.
How’s your coffee? Awesome. It’s black and white this morning. Our podcast will hopefully be more differentiated than that, but the coffee is black and white.
I guess it will be as hot as the temperatures in Berlin, 30 degrees, and you probably have some Fahrenheit over there. It’s still temperature. But Jasper, tell me what the bot is happening. Yeah, it’s crazy, right?
We have Meta with their celebrity. Bots. We have Grog, the Twitter bot, which is the Grog with a K at the end being accused. And then Character AI is being sold to Google, which is one of the most famous bot companies.
It’s huge. It’s huge. I mean, like a lot of stuff is happening. Last year, like one bot after the other came to the market.
Everybody made a personal bot. We had Pi from Reid Hoffman, right? Get sold as well. Then we had Character AI.
Everybody loved it. Get sold as well. Then we had celebrity bots. They got taken off, shut down.
But at least some celebrities got millions, if you read the press. But not millions in followers, just millions in dollars. So Meta actually paid celebrities to be represented as bots. And I think they got some weird names like Brew and the Dungeon Master.
I think that was Snoop Dogg. Well, I paid for my course. I got the celebrity, Kenneth Cookie. You probably know him from The Economist.
And I paid him. I paid him $404 because, well, he’s not a bot. I just needed his voice. No, but he actually, I mean, he was super kind that he is doing all of this.
He could have done the course with me in person. I mean, he is a technologist celebrity, essentially. I wanted to use his voice because he himself doesn’t really have the time. Sure.
That was a lot of money. Well, another thing, but I’m not Zuckerberg, right? Yeah, but I think the interesting one is also here. When we look at what happened.
So Zuckerberg said they didn’t get enough traction, the celebrity bots. So now they have a tool or kind of a, let’s say, a suit where you can set up your own avatar, your own bot. The Twitter bot, that’s a bit different because they’re accusing that of misbehavior. So that’s something we can maybe tackle later in the podcast, how you deal with that.
But all the others, Pi and Character AI, it sounds more like very expensive. With Equihires, I mean, Character AI, they were part of this attentions or you need paper. Two of the authors, I think one of the founders of Character even was part of Lambda, the Lambda bot that was built internally in this 20% extra time. So really reputable people and also the investors and Jason Horowitz, you mentioned Reid Hoffman.
But then what we read is that the investors get bought out. The people go back to the companies. So Character AI, they go back to the Google team. And yes, they get 2x, 2.5x, I think, the money the investors bought.
All the other shareholders, so essentially the employees, they go over the company or they actually get fired more than 100. So it all sounds like I’m not acquiring a business model. I’m not acquiring working use cases. I’m acquiring smart people.
So there’s something not working with these bots at the moment. Maybe. Let’s dig into this because you’re right. There is a lot of happening.
It’s kind of like reshuffling cards, but everybody is excited. And money is exchanging hands. This is something for the edge, right? Can you actually explain a few of those bots?
What was Character AI? So Character AI, essentially, they started the whole idea because they came out of Google. They built a bot internally at Google. There were a lot of safety concerns at Google.
And that’s why Google never wanted to make it a whole product. So they left, got a lot of money. And they wanted to have bots for everyone representing certain characters. That’s what’s.
The name is. So I actually used Winston Churchill as a bot. It was interesting with interesting conversations around wine. Also ask him a bit about, you know, the Second World War, obviously.
You could also have your own superhero that you would chat to. And all these bots would get content that is relevant to the character I would define. And it’s just to have a conversation. What we hear from the founders when they started the business and were a bit also later into more heavy usage.
It was mostly for people that wanted to flirt. People that wanted to kill time, but also lonely people that just wanted to have someone to talk to. But it was very broad in usage. And they didn’t monetize it yet.
So that was basically Character AI. Fascinating study. I might not get the numbers right, but there is an average percentage of how many people are depressed in the US. And then you take that average and you compare to clinical depressed.
And then you compare to the percentage. Of clinical depressed users of Character AI. And that was more than tenfold. People actually use Character AI because they wanted to have a relationship.
And they used that relationship with a computer. We saw as well from a clinical perspective, Character AI and other of those chatbot types helped people to not commit suicide. Very interesting. But it’s a very different use case.
It’s a case that what we originally saw with the customer support, customer care bots, which were definitely not to kill time, but rather more on speed. And these bots were basically to spend time with you. So it’s enjoyable. What was with Pi?
Was that similar or different? So Pi was very similar. It was announced as your friendliest bot. So there was obviously a race to who has the best bots.
And Pi, Reid Hoffman actually. I presented it once and I saw it and I was like, it was this bro style kind of communication. It was fun. But, you know, when you talk to Google or Apple, Siri or any of your home devices, you very soon get super annoyed that this tool is so chatty.
So this bro style, very friendly, very conversational, but a little bit too much friendliness and too much chattiness. I just want an answer. But yeah, it was a fun bot. And I remember from our conversation with Retu from Ultimate last year that if they labeled the chatbot as a machine, not as a person, then people would be much more friendly with the bot.
And what you read from the feedback from Meta about, for example, the Dungeon Master bot from Snoop Dogg, people got very annoyed that it was a bit like Snoop Dogg, but not really Snoop Dogg. There’s definitely different kind of bots, either the ones representing someone or the ones. Solving a problem for you. That is essentially, I think, the core to our discussion here.
If we look about value correction and we said AI is just a tool, it’s a service here to help us. And if we are looking for value creation, then this is the differentiation. Is it a tool or is it something to entertain us? And if it’s something to entertain us, what is the entertaining part?
If something. If something talks like Yoda, chatbot you must be. Is that really entertaining over a long time? Am I looking more for a function?
And this is actually super good. Like I use a Google Home. The typical use case from a Google Home is to set a timer. I have not used it for anything else.
I don’t know. It’s a box as big as two boxes of bread, but it’s only a timer. I could have. I’ve used a way better with this.
But anyhow. So if I tell Google Home to cancel a timer, then it will answer something like, consider it canceled. There was way too much friendliness. Just say, OK, so that I know that you heard me.
By the way, that’s also my only use case for Siri is setting an alarm clock and a timer. And it doesn’t reply at all, by the way. I talked to a product manager at Google who initially brought this back. And he’s like, we started with timer.
We are still at timer. So that’s the use case. That’s a functional use case. We spoke when we kicked out our podcast series last year.
We spoke about this interface that is not happening thanks to large language models and also triggered by chat GPT. So what we both basically just discussed is again an interface. So I either have access. To a conversation or to a character in a conversation or to information.
That’s more the customer care case because I want a solution. Where’s my parcel? I want to book a flight or something. I want to access something in a different way than maybe just clicking on a graphical user interface of a website or having a phone call with a person.
Now, this is super fascinating because let’s for a moment. Take the a bot helps me to do things away. A bot helps me. To do things.
To do it. Set a timer and I get a new interface. I get this. This is a use case.
But there is the entertainment use case. And I believe the entertainment use case. Why did. Zook have to call off the celebrities.
Is because there is this uncanny valley. Maybe we should explain this term to the to the audience. What’s an uncanny valley? It is the valley.
If you have the curve. In terms of on the bottom axis is how much is the computer like a human. To how much do we like them. The more.
Natural. The conversation becomes. The more natural somebody copies us humans. The more creepy it gets.
So there is this valley between. Hey, hold on. You’re not. You’re not supposed to be that human.
And that’s called the uncanny valley. And we have seen. That. In the celebrity bots.
Do you really want to talk. With some AI that sounds like. But it is not really Snoop Doggy dog. It could be fun at the beginning.
I guess right. I mean if at the beginning you’re like oh this is like Snoop Dogg. But obviously you never spoke in private to Snoop Dogg. You only know Snoop Dogg’s fun interviews or maybe now the Olympics.
Right. But you know it’s not Snoop Dogg. And I. This might be like the start of non-alcoholic beer.
Right. In Germany as non-alcoholic beer came people said oh that’s not the real thing. But I actually believe it’s something slightly different. It is.
I’m going to Snoop. Not me personally but like the fans go to Snoop for the authenticity. We are in this podcast live. Because of the authenticity.
So. Getting to authenticity. I can ask chat GPT to explain code to me. In a gangster rap.
90s gangster rap style. I can even ask it to speak like Notorious B.I.G. And explain code to me. But I cannot make a chatbot authentic at the moment.
Maybe. Can you give a brief explanation. Technically to the listeners. Why we’re not there yet.
Or why is it so hard to get there. Yeah. So if we differentiate between those two areas. Entertainment.
And function. Or style and form. And functionality. Functionality is hard enough.
But if I say set a timer. There is only one timer to set. So that’s the structural flow. We have seen companies like Ultimate Get Sold by ToSendDesk.
Or other structural companies like Flunk AI. That say. Yeah. Ada.
Forethought. Typical ones. All of those. They say there is a truth.
And there is only one truth. And how do they work? They take the LLM just as the interface. This is what we always said.
There is an interface. So they use a RAG model. And we have a whole episode on RAG. Where we say retrieval augmented generation.
I have knowledge. In the weights and biases. In my model. And if that.
Those weights and biases are not sufficient. I actually go to database. Retrieve information. And then I formulate an answer.
That works. If there is one. And one answer only. As a while back.
We talked about. How I tried to convince ChatGPT. That I am Captain America. And I failed.
But I actually tried to make this very clear. Now today ChatGPT will say. I save it in your memory. That you feel.
You think. You are. Captain America. Essentially it builds a mini RAG.
Where it kind of saves this information. But in aggregator. Have the problem to define. What is correct.
Is it correct what is in the weights and biases. Is it correct what is in the RAG. Or is something else correct. There is no feasible way.
For me to describe it. The earth is flat. Is that a correct sentence or not. Dear listeners.
Scientifically. No it’s not correct. However. In a conversation with Jasper.
Where I’m kind of saying. You know. The earth is flat. I have a different idea.
Jasper. We have seen this. We have seen the complexity of. Aggregators of choosing right or wrong before.
And I think the interesting. Is also to your authenticity. The model knows what happens in the past. So when Snoop Dogg got a question.
He gave that answer. But he got maybe the question 20 times. And he gave 10 different. Answers or hopefully just three.
But to decide which one is more authentic. Right now in this kind of context. That’s hard for the model to decide. Snoop Dogg just does it out of his brain.
And out of his experience. And because he’s Snoop Dogg. But the model doesn’t know. What’s happening in Snoop Dogg’s brain.
Just has historic conversations. And that seems like. When we talk about Snoop. He is an entertainer.
He’s an artist. Let’s take something where you’re saying. This should be more black and white. Like my coffee.
Let’s create a bot about Biden. Biden has said a lot of things. But if you ask the bot. The Biden bot.
About China. Well. What opinion do you want to take? The opinion the Senator Biden had.
20 years ago. Or the opinion the President Biden has today. And who is taking the decision. Of how to.
Leverage this information. And who is taking the decision. Biden if you ask him personally. Will have an opinion.
Based on his own. Weights and biases. But the Biden bot. Has a database.
Has a trained model. And it cannot create. The same. And we are not even talking.
The same information needed. But if I look now. At the future. I would want to get to these kind of bots.
I want to have. A bot that gives me the opinion. Maybe of the US government. I want a bot.
That gives me. Some insights of a famous researcher. Without reading through all the papers. I mean to the interface point.
It’s an easier access. I can ask deeper questions. And it’s just faster and more convenient. To access this kind of information.
And for many people. It might even be more fun. Than just reading a dry paper. Or historic press releases.
So maybe we race a bit forward. And look into the future. What do we need? What will lead to this kind of supremacy.
With AI. And what do these bots have to be. Very good at. So I think the time.
That we get 20 links. And we as a. Human decide. Which of those 20.
We want to click. That time is over. Because we have the ability now. For a bot to summarize.
But that creates. The problem. To decide. What is the right information to give.
From all the information out there. And I believe. It’s the age of conviction. We want conviction.
We do not want. Average. And large language models. Are averages.
They give what the general. Public says. They don’t have any other choice right. Exactly.
The quote which comes to mind for me. Here is from Angela Davis. About racism. In a racist society.
It’s not enough to be non-racist. You must be anti-racist. So it’s no longer sufficient. To just give you.
One information. Which is the most. Likely in the internet. I want to give you the information.
Structured the one. Which you want to see. Or I want to give you. The joke or the entertainment.
The one which is most suitable. For the moment. And it’s very hard to do. I get your point about conviction.
But for me it’s really more personalization. I don’t even need you. As a bot to be very convicted. Maybe you mean the same thing.
But I don’t want 20 search results. From Google. Even if I have an account there. I want maybe one.
And if I say I don’t like this part. Then it gives me another one. And if I don’t like that part. It gives me another one.
But that’s kind of for me the interaction. And there they can have conviction. On what to show me next. But I definitely don’t want 20 search results.
And then help Google finding the relevance here. Now funny. This is how Google has worked before. Right?
Like we think about 20 search results. Google has in reality. A conversation or had in reality. A conversation with you.
That’s the reason why. And this is obviously no investment advice. But this dark pride crash. Which we saw.
Because people saying oh my god. ChatGPT is now doing GPT search. Don’t underestimate Google’s knowledge. Google tried to give you one answer.
Because the first answer. Was the one answer. I tried to give to you. Now if you click on the second.
And the third. You still had a good user experience. But Google learned something from it. So Google understands the long tail.
And we will need to get into. A chat conversation. Figuring out this one answer. And one answer only.
And that is conviction. Now it’s funny. The US state officials called on X. To address issues with Grok.
From Twitter. And I’m not a fan of Twitter misinformation. But the fact is. That we had.
Misinformation all along. And we have now. But summarizing the misinformation. So I believe.
The world will be way more. In a branded setup. Or. Companies like Google.
Will figure out the long tail. The judge is still out. That leads for me to the personalization part. If I want the information.
Or certain opinions. Then I can have it. If I don’t want them. I can even filter them out.
I think that’s one. But the other one. We think about bots in the future. It’s entertainment.
You mentioned that. We also spoke a lot about. Bots assisting me in the workflow. Because I don’t need my graphical interface anymore.
So I can. Have a conversation via voice. When I’m in the field. When I’m in the car.
Or just you know. Traveling from A to B. So even accessing old clunky SAP systems. But back to your point.
I have a certain user account. I have a certain history of usage. With for example my ERP system. Obviously I want the bot.
Being a bit more convicted. Of what I want. Versus asking me too many questions. Because then it’s getting annoying.
It shouldn’t be a long conversation. If I was just one quick access. Versus if I want to be entertained. And I think the last part is.
Then if I’m just looking for information. So not even interacting with software. This is the old search. That you just discussed.
Then it’s really about relevance. And it feels like that’s something. That’s being solved more and more and more. I think you also mentioned some patents going on.
So access to information. But the workflow part. And the consumer entertainment. Part.
That seems to be at the very early stage. Because of what you just said. Around conviction relevance priorities. What do we have in both areas?
In both areas we have information. How relevant is what I need to say. Because if I go to the ERP system. Workflow.
And Microsoft talked quite a lot about it. How they believe the future workflow will look. A co-pilot for everything. So the workflow needs to have the right information.
So does your entertainment bot. We need to figure out the right information. Then we need to figure out the style. As we just discussed earlier on.
Style seems not to be the real thing. However. And this is fascinating. Because we talked about all the opportunities.
Generative AI brings. And we came always back to the point. Most likely. Companies with existing access.
We need to have the right information. We need to have the right tools. We’ll have the majority chunk of the market. And here we see.
Exactly that’s happening. So while we’re still figuring out style. Entertainment value. While we’re still figuring out long tail answer and answer quality.
What we do see. Is TikTok. Google. Meta.
X. Snap. Everybody. Who has access to users.
Puts out a bot and saying. I’m monetizing this access. I give you a bot. But there is in that part, there is actually an interesting opportunity, right?
Because as much as we had suddenly social media pushing and saying, oh, I give you an image application, or I give you a job placement application, or I give you a blogging application, we always had the problem with cross platforms. So bots that want to talk about either your knowledge or the news or the ERP system, no matter whether it’s entertainment or workflow, needs to operate across platforms. You cannot say I give you only an SIP bot if your system contains HubSpot. You cannot say I only give you a Snap bot if you are placing information as well on LinkedIn.
So we will see cross platforms coming. Yeah, the race is on. That would be actually pretty. If you think about how to disrupt old software landscape, and it’s kind of, you have a certain role, you have a certain workflow, and this assistant follows you and access those old, maybe system of records, relational databases for you in a different interface way.
I like that part. We want to see some ideas. Definitely. Cool.
Thank you very much. Jasper, awesome discussion. We covered it all. We covered style, we covered correctness.
We talked about… We talked about access and convenience, right? So this ecosystem, like the new technology changes the way how we interact on a business part as well as on an entertainment part is actually a nice micro case of the market opportunity we saw in AI. Pretty fascinating.
I could go on forever. Yeah, but unfortunately, we still have something to do, right? So yeah, let’s… Thanks so much.
Thanks so much. Next time, different coffee, maybe different weather. Hopefully, it stays warm. Have a nice day and a nice evening.
Thank you. Talk soon.