Transcript
But today I have lots. I'm happy to be here. You have a master's degree in quantum physics from TU Berlin. And you have an MBA from INSEAD. And then you have built products at Google, LinkedIn, Snap. You have actually written that the winners won't be those who use AI.
They'll be those who redesign work around it. Certainly. There is what I mean with winners won't be those who use AI, the ones who redesign is how I Business hasn't changed. We still have needs. We still travel. We still consume.
We still build. The only question is how can I do this business more effectively? And yes, there will be people who then use ChatGPT to write an email. But the winners will be those who either build the platforms or rearrange the workflows in order to be effective. Like what does accountability look like in this AI native workflows? Very fascinating question.
Now, if people write an email and that email is wrong, then who is accountable for it? Should I say, "Oh, it was my AI agent that was accountable for it?" Not quite. It was me who trained and set up and focused an AI agent. But you have said that the website is dead. It's hyperbolic. The website is not dead, but the newspaper is not dead, right?
The homepage is not dead. Nothing of this is dead. It's only it's reduced in importance. Then an agent can go, look up websites, read those websites, do the research, compare it, create a metric for themselves saying what is the most explosive or most fruitiest, and then give me back an answer. What is the biggest gap between hype and reality in AI right now? Implementation.
You argued that the next AI battleground isn't really prompts or agentic systems, it's memory. Yes. And so, memory is essentially part of our context window. I'm pretty sure some companies created some AI versions so that in the interview you get an AI support. And I was like, I wonder I wonder whether that works. So I asked ChatGPT about it and it scolded me.
It says, "Lutz, you as a faculty member you should be extremely careful using those tools." And then so it had a very clear opinion of what I should do. But even worse, it noted down in my memory that I wanted to use that tool. That's pretty much a good idea for a big disaster. Hello everyone. Welcome to yet another episode of Build AI podcast. Today I have Lutz joining from Mountain View.
Thanks Lutz for joining us. Happy to be here. Great. So I'll just jump in. I'm very I'm fascinated by your journey and your background. And so I wanted to ask you like you have a master's degree in quantum physics from TU Berlin.
And you have an MBA from INSEAD. And then you have built products at Google, LinkedIn, Snap. So can you walk us through the arc like how does a quantum physicist end up you know leading product strategy and building products in Silicon Valley? I'm faculty at Cornell as well right at the moment. And I very often get the question from my students, should I still study X engineering or X business or whatever it is. The world is changing very strongly and for me physics was the base of learning complex matters and structuring complex materials.
And I've done this essentially all my work life and I've done this for my own startups. I built two as well as for Google, LinkedIn, and Snapchat. Right. No amazing. I have been sort of reading some of the stuff that you have published or, you know, the podcast that you have been, you know, part of. And you have actually written that the winners won't be those who use AI, they'll be those who redesign work around it.
I think the workspace is changing so fast, right? So, I wanted to ask you like, can you give us a real example of a team or company that that made that leap and and companies that that didn't? And and what kind of a difference do you see between them? Absolutely. There is what I mean with winners won't be those who use AI, but the ones who redesign is our business hasn't changed. We still have needs, we still travel, we still consume, we still build, right?
They are the human needs or the needs which we in businesses have hasn't changed. Not because there is AI, there is no business. Far from it. The only question is how can I do this business more effectively? And yes, there will be people who then use ChatGPT to write an email or use a product to do your customer service more correctly. But the winners will be those who either build the platforms or rearrange the workflows in order to be effective.
If we talk about agents doing something for you, it sounds so neat. We can like relax and we can all think, "Oh, this is so cool. The agent The agent thinks, the agent plans, the agent takes over." It ain't that easy. Many things can go wrong. There are many risks which we are producing. We essentially have to design the workflow for what we need.
And all of us can actually easily relate to this. Why do we not yet have an agent that answers all of our emails? Because it's too complicated to do this in the broad set. You actually need to figure out for your specific emails, what are typical categories, and then have for your specific emails an agent put into the workflow. So, the focus is on the workflow, the surrounding, the setup, the how do we empower you in your environment, and not can you use it? True.
We actually have built an email agent which uh understands the email, tags them intelligently, and then creates a response using the tone and style of the person. But, I would agree with you that still like 95% there. It's not It's not 100%. So, yeah. It's And it's It's all of us know this, right? You take an email and you copy-paste it into your favorite large language model, Gemini, ChatGPT, Anthropic, whatever it is.
And the output is okay, but not [clears throat] what you would have written. And that okay and but not what you would have written is essentially very annoying for people, and it is a context question. It's not that this is not solvable. Good structure for setting up an email client would be some in the background, the agent reads all your emails, creates a categorization of how you normally respond to which person you normally respond in which area, and keeps that as a context. Now, the context windows are large, but we cannot throw all my million emails into the context windows in one go, figure it out. No, we have to condense it, structure it, make [clears throat] it observable, and then allow the agent to work over it.
Now, that is already a little bit more complicated than just using it. So, when I say the future belongs to those who actually structure that work, then that's what I mean. Or, put it into historical terms, we got the worldwide internet at some point in time and everybody was like, "Wow, look, I can now can now use a browser to order things." The world did not belong to the people who typed in an order into a browser, but to the ones who built platform to allow shopping or to organize workflows in around shopping. Yeah, very well said. I'm switching gears. You talk a lot about accountability being the missing piece.
I wanted to ask you like what does accountability look like in this AI native workflows? Very fascinating question. When we see people get excited about open claw, then accountability very often is missing in that structure. People believe that Asians can do everything for them. That is not true. Now, we might say, "Hey, but look, there is recursiveness.
At a later stage, we just set a target and then the AI learns everything we need and context and you get it." All wrong. And yes, open claw is terrible in terms of security, but this will be all fixed, probably. But again, accountability means who takes responsibility for something. After we took a decision, there is a process happening. I had very amazing discussion yesterday with the CEO of Schwarmär. So, Schwarmär builds autonomous drones that goes out in a swarm.
They recently IPO'd for Cornell, small little like announcement here. For Cornell, I'm running as well a podcast. I hope that's okay that I talk about this, which is called the keynote, and where I'm interviewing people about AI and applications. And I will have the CEO from Schwarmär there. And in the most important part, if you think about autonomous warfare, is accountability. It's not that you say something went.
It is about who is accountable for something happened. Think about at the moment have conflicts. Think about some Tomahawk missile. The Tomahawk missile once underway cannot easily be recalled. It's not like oops, I made a mistake come back. That 10 12 minutes the missile is on the way.
It's essentially autonomous. Nevertheless, we would not say no, it wasn't me who shot the missile. It was the missile was on the way, right? So you have a very clear idea who has accountability. Now, if people write an email and that email is wrong, then who is accountable for it? Should I say oh, it was my AI agent that was accountable for it?
Not quite. It was me who trained and set up and focused an AI agent. We very often talk about the new world that all of us become managers or all of us becomes CEOs from many many agents. I like this picture. That is actually a very good description. But if that description is true, then obviously the accountability lies in you the CEO.
Because when something happens at a traditional corporate, the CEO might lose their job if somebody of their employees did not do the right thing. Because ultimately they are responsible. Right. I I really like the point that you're making here and I I think I also feel that this is a good way to also describe what is happening on the consumer side versus the enterprise side in AI. Like there's a so much hype on the consumer side with open claw and clawed launching things every day. But when you go to the enterprises and you see the adoption crisis, if I may call it, there are, you know, like 25% companies which are doing something with AI, but most of them are actually doing pilots.
A very few of them have actually gone into production and I feel the the lot of reason is reason behind this is what you're talking about, right? Accountability. For example, we do a lot of work on things like anti-drift engine for agents. The agent was supposed to do something and it had like, you know, this outcome but then it started drifting in between. Are you like are the companies that making these agents measuring at every moment that are the agents drifting? And then, you know, some sort of an audit trail like who is responsible if something bad happened, wrong emails were sent, wrong contracts were signed or or whatever it is.
And then there are things around security and so enterprises actually would go into mass scale production when all of these things are solved. So, I think that accountability point is really good. It's accountability, it's observability, and it's essentially it comes always down to quality and control. And we can now have an agent coding for us. I made a test. I recently was asked by a company to look at the effectiveness of their search.
So, what I did is I built a small scraper and searched the website. It's not a big deal, right? You have the web URL and you do a small question like question mark plus Q equals a query, so you go and you slowly but steadily retrieve all the content from the database. I had two possibilities to do this. I could do this with a loop and just code this by myself in Python or I actually thought why not ask Codex, the agent from OpenAI that codes for me. I used Codex.
47 lines later, I had a scraper. Which was very cool because I wrote everything in human language, but 947 lines is a lot. It's a massive text. Meaning, is there a control equality control in between? Far from it. Now, if I I recently built an agent, one of the things I do for Cornelis, I I run this workshop for everybody can join there.
It's It's a workshop about um how to build agents. If you or any of your startups wants to to join, let me know. It's a work like we are like it's 3 hours, we code agents, and we understand the complexity of agents. Let's say I do build an email checker, and I go and well like I I can again use any kind of coding tool. Let's say I use Claude this time. It will build me an email checker tool.
Um however, it will probably not use the best model because on average, like email checking used to be keyword recognition. So, it will find personalized emails because it said my name Lots, and not use a large language model that has context. And if it uses a large language model, it will use one of the older ones, smaller ones. So, the quality is an issue. Moreover, observability is an issue. It might not use LangGraph and Lang like LangSmiths in order to understand what's in happening in the machine.
It just writes it all up and says, "Now you have it. It succeeded. It ran." So, evaluation, testing, quality, all of these are extremely important points, and those don't go away. The interface changes. Meaning, for all my CS students, when they come to me, it's like, "Is coding still a thing?" Well, coding isn't, but architectural design and quality still is, and therefore, yes, you are all very much needed. Right.
Completely agree. I'm I'm going to switch gears. I'm very glad that we talked about this these very important topics and let's talk about something which is about customer acquisition and traffic and so you this is about geo AO, you know, answer engine optimization and it's a very big topic because for some websites and merchants from 5 to 15% of the traffic is already coming from LLM searches. So now, you know, like it is going to impact the commerce, the transactions that a merchant might see or a website might see traffic. And I think you made a very interesting comment that you have said that the website is dead and I wanted to first start there. Like what do you mean?
Like please explain. It's hyperbolic. The website is not dead. By the way, LPs still exist, right? So we stream music, but we still have those LPs which are now in fashion again. So the website is not dead.
The newspaper is not dead, right? The homepage is not dead. Nothing of this is dead. It's only it's reduced in importance. If um if we think about how do I find a product, let's say I want to buy a good coffee. I love coffee, right?
So I want to buy a good coffee, something which is extreme fruity and tasty or explosive and so on. I can actually go and search and Google it. And then I have to describe it and I have to go to the website pages and have to read it and have to take a decision. Is that the right thing? Now, there is already an agentic workflow here which is so encapsulated into what we call search that we as a consumer don't realize this. The agentic workflow is if I say I would like fruity explosive coffee taste, then an agent can go, look up websites, read those websites, do the research, compare it, create a metric for themselves to saying what is the most explosive or most fruitiest?" and then give me back an answer.
That whole process of research, estimation, summarization, bringing back to me, that is what we currently see in search and discovery. Now, to our initial discussion, who is going to make the money here? The ones not who use suddenly Open AI, they have to pay $20 for it. No, the ones who build that search and discovery, the Googles and the Open AIs. Now, why is the homepage or the page of the the front end actually bad bad? Because now an AI agent is doing all that investigations, and because you do all that investigations, I don't need the front page anymore.
I don't need the front page to be made for humans. I need a page for agents to be able to understand what's here, and I need that agent to be like getting to the information easily. But that agent does not necessarily need funky images or or like large-scale videos. It just needs the information truthfully in a machine-readable format. And then it takes a decision, helps me to make my decision, and I come with a very high intent then to the website. So, there are two things for brands now to realize.
A, yes, your traffic goes down, but your revenue probably doesn't because you should see more high-intent consumers. Consumers that used an agent to be informed, and then they come to your website. Later on, they won't even come to your website because an agent might be shopping for you. But that's still um for many people far out. But the future is actually for those who manage now that process, who manage to inform agents about how to summarize, how to research, how to do all of this work. And that's what GEO is about.
SEO, search engine optimization, was about how can I ensure that I'm seen by the search engines. And GEO, a gener- a generative engine optimization, or AEO, answer engine optimization, or LLMO, large language model optimization. We have loads of those. Freaking time shows you it's a new concept, or relatively new concept, is more about how do you help the search and discovery process so that the customer, the end user, then comes to your website. It's fascinating the times we're living in. Uh everything is getting affected.
Like I was reading uh that websites like Digital Trends has lost 97% of the traffic from peak to Jan 2026. ZDNet has lost 90% traffic. Wired has lost uh lost 85% and so on and so forth. Like TechRadar, Wired, all that is happening. And then, you know, all of this started because Perplexity started this AI search thing. And then Google started giving, you know, AI search results.
So now nobody's going to these websites because the AI is actually giving the summarized answers. So the agents have visited these websites. And similar thing is happening on the ad side. I think the ad engines are also going to get affected if agents come with a lot of context and what they want. Uh so as you said, like the if you provide that answer on your on your site, if you have the right schema, if you Like, you know, we we had a we had a guy who was building this AEO GEO tool. And he said, "Today, unlike SEO era, you actually need to build almost encyclopedic content on your website.
You have to answer every possible question around around your industry, your product, and and so on." And that's that's something that agents really like. So with with so So happening, I wanted to ask you like if you are a founder of a company today Let me jump in there because I disagree that you have to answer all possible questions. You do not have to answer stuff which is common knowledge. You do not have to answer like for example if I tell you Amit the earth is round you will be like yeah, okay. Why? Why do I have to listen to this?
Because you know that the earth is round. I don't tell you anything new. If I tell you typically example which I always love to use is life is like a box of chocolate. Why? Because we are all brainwashed individuals who has seen Forrest Gump and that's the meme and it's chocolate. Now you are Amit currently at a place in the world where the temperature is so that chocolate is not actually that delicious.
You have to eat chocolate in an air conditioned room coming out of a fridge to make it actually even delicious which in the most like in the area where you are not everywhere is air conditioned. Meaning chocolate isn't that delicious and chocolate is a sticky kind of substance substance which gives you diabetes if you consume too much and is not amazing. So why is life like a box of it? Life could be like a box of surprises, love and happiness. Now when I tell you those things that's novel content for you. That is content which is not the main general average and that makes it special.
So when you are a brand you do not have to answer all the questions. You have to answer the question what sets you apart and what makes you special. And that was always the case. It's only that marketers lost sight of it and kind of thought if we do keyword stuffing maybe it works better and then we are back in the SEO game. But the core principle is customers want to buy something and they why to buy something which is differentiated and special. And that is still the case.
I have to answer in GEO, how am I special and how am I differentiated? And I have to do this in a trustworthy manner so that the model can pick it up and realize that's what LUTZ stands for. Right. So, it's it's very interesting, like today it's very like nobody can say with certainty whether it should be encyclopedic or whether it should be very specific content. I think we'll have to see like how how these LLM searches actually pan out and you know how how the sort of the analytics looks from, you know, once once you know, the searches have materialized. But I I think the question that I was going to is that if you're a founder today of a company of a startup, how should how can you make sure that your product show up when somebody's searching in ChatGPT?
Are there like are there techniques or things that you would do? Yes, there are different levels you can make sure that this works. I run yet another thing. So, if you go to my website www.lutzfinger.com and sign up for the newsletter, I actually for people I have a following, like for people who sign up for my newsletter, I offer workshop on GEO. It's a very small workshop. Again, it's it's about 3 hours where we work through the different levers somebody has.
So, if you want to join, you're totally welcome. And if you say that you heard this in this podcast, then I am happy to give you a discount for the workshop. Now, very high level, the workshop breaks like the the impact breaks down into technical and content and and workflow. Technical, there are things you can do to make your website being easier possible. It's similar to the SEO game. You can have question and answers, you can make stuff easier to be understood.
Content, you want to have um content to describe what makes you special. And um so that the websites can parse that and kind of come back and take a decision. And you want to do this in an authoritative way. That's the reason why people love LinkedIn and Reddit so much because the authority is there. If I write it write at Forbes, the authority is there as well. So, these are good tools to booster your uh recognition.
And then the workflow is you do not want to think about what are typical questions my brand should solve. You want authenticity, meaning you want your customer say this or you want to go out in a broader spectrum. And that authenticity you only get when you reach out to a community, ask people. And all of this is a process and therefore suddenly you're not down to a workflow question. I have seen Profound and other companies going around and building measurement devices. I by myself, my last company, I've built queryedge.com.
You can test it out, it has a free version, which does measurement of how well is your brand positioned in Op May I and other search engines. But measuring in itself is not helpful because measuring is just one of the effects you will see. So, you need to fix the technical part, you need to have the right content to be picked up. Which content? Is it the earth is round? Is it life is like a little box of chocolate?
Probably not. You if you like what am I What does Lutz Finger stand for, right? I stand for product and AI. I stand for geo, I stand for workflows. Those things are things I know about and I can help. I don't do everything.
I do a specific things. I have a very specific skill set where I can help people. And all of them hopefully know this. Why? Because I posed, I do my videos, I kind of explain in short snippets the content I know about. So, that's what brands should do.
And then in order to scale up that content, you need agents. Right. I'm going to switch gears and this is very helpful and interesting and talk about something which is, you know, very, you know, a lot of fun as well. So, in the research it came out that you literally replaced yourself with an AI bot in your own Cornell course. So, I wanted to know like what happened. Like did the students actually learn better or worse or differently?
It's a It's an interesting question. If you are like me, faculty at an at a university, what's your task? Your task is not to summarize content. And students come, they want to learn content, but honestly, they could learn this content probably way more effective if they talk to ChatGPT or to YouTube or TikTok or who knows what. Why? Because every student has a different baseline where they come from.
Every student have a different learning style. I cannot in a class I currently I teach 100 in my my current court 180 people spread over three classes, right? I cannot adjust to every learning style. So, I do something, students like it, but that doesn't mean that it's the optimum for each and person. Put it differently, there are better tools out there for students to learn than listening to Lutz. And if somebody needs to learn the base basics of machine learning, data, observability, they don't need to talk to Lutz.
This is known. Like pick up up a book or ask a website or go to medium article. I can help them to guide them to correct information because there is still, like, obviously BS out in the world. I can help them to to curate. But, that's about my value. If that's the case, then you don't need a professor anymore.
Then you don't need a faculty anymore for this specific part. So, I replaced myself then. I made sure I had a curated list. I based on that curated content, I trained my students. Now, students can ask the AI lots all the things they want to have. I have a couple of videos I explain to them, and I flipped the classroom.
Students do not need to read anymore. They need to talk to my AI agent before they come to class because games learning is happening on the side of the students, not on my side. And content aggregation is I collect the content and I let the student engage on their pace, on their complexity with the content. Summarization is not a job anymore that a faculty has in the future. Therefore, I replaced myself. I replaced myself as well to actually show my students how to do this.
So, um how do you replace somebody in my liking? It's It's really about who looks like me, even has a accent like me, and is based technically it's a rag, meaning there's a database that where the agent, my fake lots, takes information from, summarizes, and brings it back because summarization is not a task of a faculty member anymore. What is a task though is ensuring that people understand where things go wrong. So, my core task is not summarization anymore. My core task is to challenge the thinking, the logical thinking. meaning let's go back to the agent design that we had earlier.
If I have an agent design, I want to have observability, and I want to have an input control. One of the typical problems of agents is that you can do code injection. Like there's an agent, and code injection is and we see this in the news, like people making fun of it. I Like I think Sephora had a had an agent, and somebody says, "Write me Explain to me Python code." Oh, no, no, no, Chipotle Chipotle had a had an agent for explaining the menu, and somebody says, "Okay, write some Python Python code for me." That's prompt injection. You go in, and you tell the agent something to do that the agent can't do, right? So, to understand why the agent Chipotle created is not functional sufficient.
To think about those risks, to think about the architecture. That is my job, to make people aware of the nuances, to make people aware of the risks, to make people aware how to deal with them. So, and that's This is so easy to show. You go to any of your favorite coding LLMs, ChatGPT, Claude, you name it, and you ask for Create for my agent, create an input control. Again, it will use keywords because that's like the sentence Life is like a box of chocolate. The common, most generic way how people did it, but that's not the way how you want to do it today.
So, understand when you then see the code, when you see the structure that the LLM created, that this is wrong. We need to be able to challenge. We still need to have How do you say, like the critical thinking element in our approach. And that's the core value a faculty member can bring. That That's what I focus on in my courses. And I replace the stuff which I believe should be replaced.
Very interesting. So, my partner at GI Ventures, he actually takes a class at UC Berkeley every week and he invited me and I can totally relate to what you're saying because what he does is like he actually asked them to read a book a famous or you know something very useful every week like let's say hard things about hard things and they have to read it and come and then they discuss like some very and he wants to sort of challenge their critical thinking during the discussion. And it was like very interesting like you know how that entire class goes. So, I can totally relate to what you're saying. Maybe we you know this last part of this part we I I want to do like a rapid fire. So, you know quick like four questions, you know quick answers.
[snorts] So, are you are you ready for it? Okay. Uh biggest AI myth that needs to die. AGI is close. And I think did you recently explore or did a podcast with Martin Casado from A16Z? Yes, Martin Casado and I we talked about it um the Cornell keynote.
It's a very good podcast. So, highly recommend it. What is the biggest gap between hype and reality in AI right now? Implementation. Right. One prediction for 2027.
OpenAI will focus. Right. One piece of advice for founders building with AI right now. Don't underestimate business value. Right. Cool.
Actually, I have you know I you know because I have the mic so I can ask you probably one more question if you're okay with it. I just thought about something. I think I read somewhere that in one of your like you know post or something you you argued that the next AI battleground isn't really prompts or agentic systems, it's memory. Can you can you talk a little bit about that? I think it's an important topic. Yes.
So, memory is essentially part of our context window. Let's go into politics. If we look at our current landscape, political situation in the United States, you find very clear opinions on both sides of the aisle. And both sides have a complete utter misunderstanding of the other side. And the reason why we have this missing discourse or this misalignment is memory. It's context.
What do people believe to be true? They put in and that makes the context. At Google, remember I worked for Google, one of the main parts of search before LLMs was keep no memory. Because when you go to Google and you type and you will not believe how many people type into Google search facebook.com, then they expect to get facebook.com so that they can click on it. They do not expect something else because, hey, I showed you where facebook.com was last don't ask me again, right? So, memory is something very, very tricky which becomes the next battleground.
How do you as an organization ensure that your memory is kept in the correct form? What is the memory you want to capture? This is a problem for us humans as well. In a medical world, PTSD is essentially a problem from the memory you acquired. You had very negative memory and those are stopping you from functioning correctly. If you, I mean, if you look into venture capitalist, I can point to a handful of venture capitalist that had bad memory because at one point in time they failed in the space and therefore they're not touching it again.
It's a memory and context problem. So, models now start to learn about our memory and they are keeping that information. Now, how do I deal with it? How do I manage this memory? I give you an example. So, I built for my students I built an bot that researches jobs and then helps the students with the outreach to alums.
And then I was like thinking, can I build this bot even further so that it helps you in training and so on and so forth in terms of the interview. And I was like, I think I'm pretty sure some companies created some AI versions those that in the interview you get an AI support. And I was like, I wonder. I wonder whether that works because I I I cannot imagine working it if you are in a life discussion you can see where the eyes are, right? So, I asked ChatGPT about it and it scolded me. It says, "Lots.
You as a faculty member, you should be extremely using those tools." And then so it had a very clear opinion of what I should do. But even worse, it noted down in my memory that I wanted to use that tool. Okay. I had to clean up the memory. So, that is like for one person. Think about you are now Morgan Stanley and saying, I built a huge AI system based on all of my reports, which Morgan Stanley tries, probably.
And now suddenly that reports become part of the current memory. That's pretty much a good idea for a big disaster because they don't have yet control over memory and what to include and what not. Super interesting. I think that you know, we'll have to end this but this is super interesting. I can go on for another like probably an hour or so. But thank you lots for your time and sharing like really amazing insights.
Uh We will and I'll also tell the audience to look at some of the courses that you're taking one on the agentic side, the AEOGEO side and also visit your website. But again, you know, thank you so much for your time and coming to the podcast. Thank you. Have a good one. Thank you.
End of episode · Ep #16