The Hidden Patterns

Lutz Finger: Welcome welcome to yet another Cornell keynote. I’m your host, Lutz Finger. I’m a faculty at Cornell where I teach courses on AI, data, product strategy, and I spend a lot of time helping leaders like you to think through what AI actually changes in product development, in our markets, in our strategy, and then customer behavior. My own work for years now has been focused on how digital shapes the world.

I built a company called Fisheye Analytics and sold it to WPP. It was in the social media era. Then, I worked for Google, LinkedIn, and others, and I recently built a company called r to the side. It’s a platform where we offer fine tuned models for ecommerce and AI workflows.

Recently, sold to XGen AI. I’m now the head of XGen AI, and I teach a couple of courses at Cornell. As well as a couple of workshops. So if you wanna learn about AI and agentic workflows, we offer two workshops for you.

One is about personal use of AI, how do I get AI to read my email, set up my agenda, and organize my travel. And the other one is more focused on AI meaning how do I make sure that those things are scalable, safe, secure, and set up. Both courses are very hands on. You find the links in the show notes like always.

So would love to see you there. Now today we wanna talk a little bit about what gets built. We have seen amazing changes in the world. We have seen OpenAI becoming a consumer product, and lately, they are struggling about maybe they wanna become more an enterprise product.

We have seen cloud becoming an enterprise product. We have seen enormous amount of fundings going both in the consumer as well as in the enterprise world. And behind that engine are venture capitalists. I’m associated with Cherry VC where I’m a venture partner meaning I help Cherry to decide on investment decisions.

But it is a huge industry that funds essentially the growth in Silicon Valley and around the world. And it funds the AI growth. So the question is, how does AI work with VCs and what gets funded and what doesn’t get funded. So that’s the question of the day.

Looking at what is fundable. Where is AI going? For that question, I brought in a dear friend. I worked with him back at LinkedIn, as well as a stunning general partner and somebody who has used AI quite nicely to get to good investment decisions.

So I’m delighted to welcome Ben Ortlieb, Co-Founder and General Partner at Blue Moon. 6,000 startup pitches in an automated way, and that’s a cool thing. So we are going to talk all about this one, but for now, welcome to the show. Ben Orthlieb: Thank you.

Lutz Finger: Now the way we do this, we normally start digging in a little bit into your background, where you came from, and so on and so forth. Tell me a little bit about yourself, and tell me about Blue Moon. Ben Orthlieb: Maybe Blue Moon is partially a reflection of some of my experiences, so I’ll start with a few highlights. I am French, as you will probably pick up from my accent.

I am by training an engineer. But very quickly got into other areas including finance and strategy. We crossed paths at LinkedIn where I stayed for over ten years. Leading corporate development M&A.

One of the premises for Blue Moon comes from my time at LinkedIn. Lutz Finger: Where we met, by the way. Right? Ben Orthlieb: Working with folks like we met.

And had our first conversations about data and what does it tell us about the world. Part of my curiosity and part of the initial discussions with my cofounder around Blue Moon is VCs don’t tend to use any data. They don’t tend to use any technology. Agriculture uses way more technology than venture.

And so years back, I started. Lutz Finger: I would settle down on this, agriculture uses way more technology than venture capitalists. Tell me. Ben Orthlieb: I will maybe drill down a little bit.

At the early stages, right? The late stage at this point is a data problem, and frankly, if you’re doing this, you’re losing out. Most growth stage VCs have seen hedge funds coming to their turfs. And had to react.

Growth stage is mostly a signal and projections problem. It’s basically science at this point, with brackets around sort of error bars. Early stage venture is still very much a craft in the good way of a craft which is necessary, and we’ll get back to it, I’m sure. But it’s also, frankly, flying blind mostly on gut instinct and what people have learned over the years without really even having the datasets to figure out is what I’m doing true or not true.

And so, of course, VCs are starting to use Claude or ChatGPT but mostly to reproduce whatever they’re doing. It’s a sustaining innovation. They say, oh, I want Claude to evaluate a company like I’ve been evaluating them. And the question is, like, are you even evaluating them right, which no one is really asking.

So if we go back to LinkedIn, one of the things I had the chance of doing given the dataset, I worked with the data science team, your old team, to evaluate are there signals in the LinkedIn data to predict future company success. And not surprisingly, the answer is broadly after series B, there are signals. There are market signals, and there are signals internal to the types of data that LinkedIn could put together that are predictive. Things like the quality of where you’re hiring from, things like not just your hiring velocity, but actually how many people do you win compared to how many people you lose to other companies.

Those types of things, very predictive. In the early stages, no signals really work. The only thing that works is data around founders specifically. No company level data.

That’s, of course, a place where LinkedIn itself will never go because they do not want to have a shadow profile on anyone. But it was truly revelating for us thinking about how that leads into Blue Moon to say, look, at the early stages, it is true that it’s all about the founders. And so we’ve built a system that effectively tackles two questions. One, how do you see as many of the founders as possible?

Today, your top seed funds, because the market is so fragmented, only evaluate 25 to 30% of the top deals. And then two, ultimately, you need to get conviction about the founders. That part is very human to human. And we’ve built all that tech so that, ultimately, I can have a conversation with Lutz about his upbringing, his family, what happened in his life.

It’s a little bit the secret sauce for Blue Moon, we’ve been able to develop a very effective workflow, very high volume, very high signal. With data, with AI, of course. All of this so we can have at the end of the process, a deeply personal conversation. That is the last step in us making a decision.

Lutz Finger: I like this approach, because when you look from a product manager lens to it, your main focus of a VC fund is obviously to bring return to your investors. From a product manager lens, we very often are faced with very interesting tools. ML tools, workflow tools here and there. And there’s always this tendency from bad product managers to throw tools on it.

Now the way you described Blue Moon, you separate technology actually in two different groups. Group number one is can I use data to find interesting profiles? Can I have a signal? As you said, the signal around the founder is the most important one.

Group number two of the product is once I have identified targets, can I use, in this case not machine learning data, but AI workflows to make outreach and nurturing easier? These are two very different structures. One uses ML. The other one uses generative AI and agents.

And that’s actually very neat to see how you build up this product. Ben Orthlieb: Yeah. And there’s an interesting feedback loop. The fact that we use so much technology broadly actually makes us stand out with founders.

Which means we win more because we are, in fact, a tech company, not a VC. So to explain a little bit more the key phases of the process: We source potential candidates through data. For us, we invest in B2B, broadly defined, North America, early stage. So we use about 15 different pipelines, sources, to try to identify new teams, companies, people, that we then qualify.

At that point we use generative AI. To give you a sense of volume, so far this year, we sourced 7,600 teams that fit the seed B2B North America box according to our system. Lutz Finger: When you say sourced, you mean scored. Ben Orthlieb: Yeah.

Meaning, we have found them. Prescreened them as being potentially in our areas, and then sent them to scoring. Lutz Finger: And when the audience thinks about AI, because that’s the hype of the day, today is, hey, we do everything. Ben Orthlieb: This is not ChatGPT.

Lutz Finger: This is not ChatGPT. It’s actually machine learning and it’s pure. Ben Orthlieb: Yeah. The scoring is 100% machine learning.

The scoring is not us telling the machine what we look for in founders. The machine is learning from founders’ backgrounds what are the types of teams and people that tend to be successful founders, so that it helps us prioritize our time. Because, ultimately, 7,600 companies, the biggest traditional funds will see about 3,000 in a year. We are a third into the year and we have seen some 600.

It’s completely unmanageable. My calendar cannot accommodate for that. Lutz Finger: But hold on. This is the second part.

This is the calendar. It’s super important. You’re jumping over, and I would like to make this explicit because even if you are in the audience and you are any kind of manager or executive, you are faced with that switch between machine learning and generative AI. And it gets very much mingled up.

And what you describe is an extremely neat product case of separating those. First, identify, machine learning. And then do the outreach. Ben Orthlieb: Yes.

My point is because we generate so much deal flow, I will never be able to meet all these companies. And, frankly, most of those meetings would be a pure waste of time. So I need to be able to spend time with the highest potential teams. It is not my taste or me saying, Claude, I like people that have a background that looks like Lutz.

No. It is me saying to the machine, tell me the people who are most likely to succeed. But, ultimately, it is a time constraint. I can only meet, in reality, three to 400 new companies a year.

How do I select 300 to 400 companies from about 15 to 18 thousand that I’m going to source this year? Lutz Finger: Now to give you an idea of how this looks in reality, after I sold my last company to XGen, my last company did fine tuned models based on media as well as ecommerce. I sold it to a company called XGen. Shortly after that, the news got out.

Now all the models picked up the signal. Oh, Lutz is a successful founder. What did happen? First, I got into the machine learning system and got flagged.

And secondly, then there was a workflow for outreach. Ben Orthlieb: I have to say my system scored you well before you had the exit. Because we had this conversation ahead of time before you sold. A lot of VCs will use sort of one kind of criteria and assume it is predictive.

I’ve had fascinating discussions with very seasoned, very experienced, successful VCs who tell me, especially when we started five years ago, using machine learning for venture was sort of, you can’t be serious. Then people would proceed to tell me, oh, here are all the signals I have found to be true for myself, that I would have liked to have tested with data over time. But my instinct is those are the right attributes for a founder. And then I point out that this is exactly what machine learning does at scale as opposed to your instinct, probably biases, taking a couple ideas and then running with them.

But most people are not comfortable culturally to give up that part of the decision making. There are things that VCs internalized as being part of the craft, which I think are mostly urban legends. Lutz Finger: Which is amazing. Let me double down on this.

We see investments live sometimes on urban legends. Because you were an early investor in Google, your fund looks amazing. And now everybody thinks you are doing well. And you kind of have the feeling you do well because everybody tells you you do well.

Right? Instead of having the data driven approach. Now interesting part, how many sources do you score? Where do you get those sources from?

Ben Orthlieb: Any idea we can think of. Truly. So because we’ve removed the constraints of having an army of associates do the screening, the marginal cost of finding more people is actually not that much. So the big buckets: of course, all the accelerators.

When the data is public or semi public, we capture it. We score all of them. So before demo day, I’d already reached out and started adding meetings with all the companies. There’s all the public, semi public sources of information, Crunchbase.

Crunchbase has some interesting trending metrics in the back end that we analyze. So we see which companies are getting more visits than usual. This is how we sourced Mercor, the biggest company in our portfolio. We used extended networks, but not in your traditional VC way of, oh, my network will source me information.

We found ways to figure out every founder that you’re connected to. We also have a list of about 41,000 people that we proactively track. A lot of people are there because they’ve been previously founders funded by good funds. Senior product people, senior software engineers at all the scale ups and unicorns.

All those people have more tendency to become a founder. So I’m not reaching out to you now. I need to know in sort of between six and eighteen months period of you starting a company. Lutz Finger: There are two points.

One is a public service announcement which I should have done way earlier. For everybody listening to this live, you are welcome to type in questions below and ask them. I will see them here on my screen. If you say your name, I will state your name.

Second thing, if you’re a student of mine, like many of you came to me with a question in the world of AI where everything is organized, what should my job be? Where should I go? I always recommend you to do something very similar to what Ben is doing. Use data to score companies.

If you are listening, I would challenge you, we need a kind of open source repository in order to score what are good companies to work on. Ben is doing this because he wants to find good investments. You, if you’re not an investor, but somebody who is looking for a job, you want to do this because you invest your own resources on it. Now let’s go to the AI agent part of it.

What is this AI agent thing doing, and how does it fit into your world? Ben Orthlieb: The next phase is to talk to companies. To understand what they’re doing and ultimately talk about the people, understand what they’re about. At that point, it’s a human decision.

But the best way I can think of this is I need all possible information to make sure I’m trying to take the best decision possible. And so instead of flying blind, we have built what you can equate to an associate researcher. Throughout the process, we have Blue, our associate, that constantly updates its analysis of the company. It doesn’t give a recommendation.

But it learns first outside-in and then through conversations, in detail, what the company is doing, and it gives us very detailed notes. Blue is a RAG. Highly optimized. A RAG is effectively the idea of you use a language model but instead of being general purpose like ChatGPT or Claude, you give it a very specific repository of knowledge that it should draw from to make its analysis.

The selection of sources becomes quite important. Because for us, early stage venture is not the traditional GPT learning set. We scrape about 150 different sources of information. All the top VC blogs, market analysis, founder and VC interviews, podcasts, YouTube, those typically are not in Claude or ChatGPT.

So think of Blue as something that has, through various techniques, very good recall of about a million pages today of information specific to what we do. The analysis that it does are mind blowing. First outside-in, it will give me a review of the company. Key strengths and weaknesses, potential questions to ask the CEO.

We also have built a full taxonomy, so it analyzes the space that a company is in, and competitors and other VC investments. The taxonomy is generated every couple of weeks. We end up with 900 segments in B2B venture. This is extremely helpful because now there is no meeting that I don’t show up prepared.

And this is kind of a dirty secret, but most VCs frankly don’t have time and don’t have the resources to prepare. Lutz Finger: Now here comes the dirty secret about RAG. Retrieval augmented generation. It’s a dataset, and based on that dataset, ChatGPT, Gemini, and all the other LLMs grab pieces from that data and structure them together into whatever output it is.

The problem with RAG is the following. How many pieces do you take, and how big are the pieces? If I would ask you to summarize the book Harry Potter, it makes a huge difference whether I tell you to read 10 pages from it and then attempt the summary or read a thousand pages from all the books and then do a summary. So RAG has a technical structural definition, and then the next thing is how do you put this together?

There is a whole discussion about how stateful, how much memory you need to put in. As soon as we had LLMs come about, there were so many companies saying, oh, I do Porter’s Five Forces for you. I’m going to rip McKinsey. This is far from it.

Because McKinsey cannot be so easily replaced by an LLM. What it comes down to is which information do you use for a given strategy? How do you cut this information? And how do you stitch it back together?

And that’s a process question. A process which you, Ben, have solved for yourself. The question to you is how did you solve this IP question? Ben Orthlieb: I think you’ve said it well.

What we are building is a process from the beginning, and it has evolved. We have augmented the process with AI, not said, oh, this is a cool tool, let me spend time on it with no particular goal. We are very clear on measuring every step of the process and drop offs and efficacy. So we are actually able to tell you what part of our process contributes to the ultimate outcomes with pretty good accuracy.

Which again, if you think venture is a craft, you have no idea, because you don’t track any of this. And Blue will never even make recommendations. Lutz Finger: If you say venture is a craft, you have no idea, I actually believe venture is a craft. The way you score is a craft.

And you can automate this now. But if you don’t track, which many VCs don’t do, every partner is doing their own thing. If you don’t measure, you will not manage. Ben Orthlieb: That’s what I meant.

People have not standardized how they do things. One of the main problems for VCs trying to do AI now if they have legacy is they have four partners, and they probably have six or seven ways of doing things. So to normalize that and then finally augment that is incredibly hard, which is where a lot of VC firms are failing right now. The two challenges we’re seeing: the GPs are not the product managers, let alone the tech people putting these things together.

And b, the poor software engineers trying to please four different GPs who don’t really do the same things the same way. Versus for us, we have the blessing of it’s the two of us. We started from the get-go with we’re going to build a process augmented with tech. We started in 2021.

But it was always very clear that we were building for automation and clarity of process. Lutz Finger: Let’s shift gears a little bit. What’s investable in this world today? The general question is everything will be taken on by OpenAI and Anthropic.

Why should I even start a startup now? Ben Orthlieb: There are many different ways I can go with that. The first thing is, yes, things are changing. People are talking about SaaS being killed.

Probably overblown because you still don’t want just to rely on, like, vibe coding IT to give you your assistance and be SOC 2 etc. But the main point there is what used to be somewhat hard is not hard anymore. Your pure workflow systems. But there are still hard problems to solve.

The classic startup of, like, I’m a new SaaS with a view on replacing your finance process, that’s gonna be tough. What I’m seeing more and more is people are going after harder problems. Because harder problems require deeper product insight and market knowledge. So I think, ironically, in the world of AI, the clearer choices actually are worth it.

Lutz Finger: I think that’s actually very fair. All this bashing on SaaS was based on this average workflow. People say, well, we can do now any workflow, so SaaS is dead. But that’s not true because SaaS still has the database, the privacy, all the enterprise scalability, customer support.

Now on top of this, you can plug any workflow, but that means you have to understand the industry. As much as I would be very bad to create an outreach workflow for a venture capital firm. Because what do I know about outreach for capital? You understand how to do this outreach, and you build a workflow for it.

And you iterate around this workflow to figure out what are all the exceptional cases. Ben Orthlieb: Think about this. The proof of what you just said is we’ve been building for ourselves for five years. We now have other VCs asking to buy pieces of our system.

Lutz Finger: Very often people have said, why don’t I have my Agatha, like you call it, for finding companies. Why don’t I have my AI to do my emails, to do my calendar? There are companies out there who offer this, but they are extremely bad because they try to force you into an existing workflow structure. While what you really want is the workflow that you came up with.

The easiest example is tagging from email. What are the tags you want from your email? Depends on you, the moment in your life, what you are doing. I have as tags customers, universities, VC funds.

You probably don’t. So that structure is an inherent structure for you. The next big thing I expect to see is a whole vibe coding of AI agents. How do you get an agent to do exactly what you want?

Ben Orthlieb: Interestingly enough, there’s a company out of Andreessen and Speedrun last week that we’re looking at that’s doing something similar to that. But, again, they come with extreme insights on the product. It is literally the CEO’s thesis at Oxford around organizing agents and knowledge in the context of a company. Somebody can come up with a mediocre answer.

But things now become hard, especially in the sense of they’re going to be extremely flexible and adapt to you. You can vibe code part of the tech. You can’t vibe code how it actually works. The architecture is where it’s hard.

Lutz Finger: When I ask my students to build applications, they build SaaS interfaces. Only because you can vibe code a SaaS interface does not mean that SaaS interface is the future. The future is adaptive workflows. But at the moment, even Claude Code and OpenAI and Gemini, they are extremely bad in understanding how to design a workflow so that it is adaptive to you.

We are missing this. We are missing the Facebook or mobile moment. It’s about to come. And whoever cracks that nut will be a winner.

Ben Orthlieb: I think the interesting point here is also it’s probably going to be headless, meaning faceless. People put a SaaS UI because that’s what they’re used to dealing with. But most of this doesn’t need a UI anymore. An interesting market trend I’ve been seeing, I’ve done 90 conversations with first companies since the beginning of the year.

One of the things that comes up a lot more is that the data is becoming the execution layer. Everybody, every largish company, has a version of Snowflake. And what they’ve built on top of it is visualization. Companies are coming in and saying, I will directly from your company data tell you how you need to change your workflows.

We go directly from the data to here’s how you start working tomorrow. No interface. Lutz Finger: Fascinating. EVALs might become the new workflow.

If I look at Agent Garden from Google, now you can connect your Gmail and then you get a little window where you can put a prompt in. Later on, you look at the evals and try to figure out, did it work? What I actually wanted to do with my workflow. And if the answer is no, to have the lineage, to have the traceability where it went wrong is hard.

There are tools to do this. But still, it’s very hard to actually dig in. If you go to Claude Code today and you’re trying to optimize a skill, you end up discussing with Claude for hours, and you would have been faster just looking at the workflow in code. And that’s a problem we need to solve.

Now tell me your big outlook. Where will we go? Ben Orthlieb: There’s a trend that builds from what we were discussing, which is deeper than previous generation in terms of industry knowledge and tech insights. If I look at my portfolio in the recent fund, all these companies have very deep insights.

The other trend is teams are smaller and more high powered. I’m seeing more and more companies with one CEO who maybe is more product or more commercial, and call it three CTOs. Very legit CTOs who five years ago could not be at the same company because it would be a waste, you could not get the leverage on them. And they’re basically saying we’re gonna run this team of four until probably our series A because we don’t need anyone else.

I’m thinking of a company we just invested in, SenseMesh. Founder of Mesosphere, an ex-CTO at Google X, and a third person. Each one of them could be a founder. But they would be limited by translating their output into actual code with large development teams.

They don’t need any of that. Three of them plus Claude Code means they can go after very hard problems with this leaner team. Lutz Finger: What’s the best investment you did? Ben Orthlieb: The first one, it was Mercor.

Mercor is a company that effectively matches people with AI companies for the purpose of building reinforcement learning datasets. They would come to you, Lutz, and say, hey, we have a project, we need you to teach the model to be better on something you’re an expert in. Every builder of a model uses this to some extent. And that extent in terms of turns out is increasing over time.

We were part of the seed in October 2023. Lutz Finger: Data quality for models is a huge thing. If we say that the new interface is the evaluation, then somebody needs to actually ensure that the evaluation data is there and constantly reviewed. At Google, I was one of the PMs in the Google health team.

So we had hundreds of doctors helping us with our various problems because somebody needs to look at data and say, this is the right information or not. Ben Orthlieb: And given the ChatGPT moment, the company went from the seed to a $10 billion valuation in two years. This is the fastest growth company in history so far. And it is a textbook application of how we work.

Lutz Finger: We see a lot of amazing companies got built and had difficulties to establish the footprint. What do you tell the startups you work with? How to stay resilient? How to win in this game in the long run when so many things are happening at the same time?

Ben Orthlieb: Going back to our process, because you used the word resilient, that’s a people thing. We actually prescreen. Our ultimate decision is how resilient are these people. And we go deep into their personal lives to evaluate where it comes from.

So I don’t pretend to teach them to be resilient. I try to invest in people that are by nature already extremely resilient. But to your question, I think it goes back to this point of deep insight and point of view. Because what worries people like Cursor and most middle ground solutions is, will your OpenAI, will your Anthropic, because they aggregate everything, also come with the average solution.

Maybe that limits my market to the 10% of health professionals that really work on this thing. But venture success is not a trillion dollar company. You are still seeing, like in the case of Mercor, companies that can go to tens of billions with a very specific angle. Tech has been gaining percentage of GDP over time, today being somewhere between fifteen and twenty.

I think with AI, this is going to explode. You can take the point of view of every company is doomed. But you can also take a point of view that tech is going to permeate so much, including venture, that you end up with, by definition, a whole bunch of winners that are going to find large niches in those markets. Lutz Finger: I actually believe if you take the idea of AI being an enabler, then you take today’s core value of a company and there’s an ability to make them better.

If you are a marketplace, you have better matchmaking. Deezer recently announced that a huge percentage of their music is now AI generated, which gets launched over the platform. People consume that music. It doesn’t mean that the music industry is dead.

It means the way we create music is about to change. Why do I need somebody to play live? Ben Orthlieb: Because live is a very different and very human based experience. This is why I play music live.

I love the thrill. The connection with people. Lutz Finger: Yes. Totally.

That’s the reason why we do this show live. Ben, thank you so much. This was fun. I could go on forever, but we should cut it to an hour.

Thank you for everybody on the show. Thank you for listening. Remember, the next show, we are going to talk about drones and AI and warfare. So it’s gonna be a blast.

Hope to see you back soon on this side of the channel. Thank you, Ben. Ben Orthlieb: Thanks, Lutz.