Investing Billions - E210: How Startups Can Avoid Being Disrupted by OpenAI w/Eric Olson

Episode Date: September 8, 2025

What does it take to build an AI-native search engine for science? In this episode, I spoke with Eric Olson, Co-founder & CEO of Consensus, the platform making peer-reviewed research accessible through AI. We covered the company’s journey from Series A to millions of users, the realities of competing with tech giants, and what truly creates defensibility for AI startups. Eric shared his perspective on the “AI talent wars,” building products at hyperspeed, and what truly creates a moat for AI applications. If you allocate to or invest in AI, you’ll want to hear Eric’s frameworks for product strategy, market sizing, and execution speed.

How do you keep yourself from getting your lunch eaten from the Open AIs and GROCs and Gemini's of the world? This is a question that is on every single founder and investor looking at the AI application layer of where is that line between horizontal and vertical? I think we're getting some pretty interesting data points start to come in because Open AI has over the last year started to launch a few tertiary products beyond their just general chatbot. An example of one is they recently launched a media recorder transcription extension of chat GPT that directly competes with your granolas of the world that are built into your Zoom meetings. That's done basically nothing to disrupt a company or product like granola, and everyone still thinks that granola adds all these additional features of delight and some additional depth to the product. If that's the case, I think there's so much room for vertical products to continue to exist.

if a product that is as simple as a median transcriber in your Zoom window can't be fully disrupted by these horizontal models. There's so much of a product that is. Congrats on your series A led by Union Square. You also had Nat Friedman and Dan Gross, fabled AI investors. Tell me about where consensus is today. Yeah, no, appreciate the kind words. Super excited to get that round done. USB's obviously got an incredible track record and, uh, specific.

specifically an incredible track record at our stage. And then as you said, Matt and Daniel, you know, biggest names and some of the biggest names in AI investing today and obviously been in the headlines quite a bit lately as part of the AI talent wars. As far as where we're at today, yeah, we're kind of in this track from Series A to Series B. We have about 20 employees now.

We have about 5 million users now worldwide and, you know, tracking towards some of those series B revenue metrics trying to grow in the 20% month over month range. You mentioned these AI talent wars. I've never seen anything like it. Somebody turned down a billion dollar offer over four years. I don't know if that's true or not, but these absurd numbers. How does an AI company compete against the METAs and the open AIs of the world today?

Yeah, I saw the same. I think it was a billion over four years and maybe even won like $1.5 billion. I mean, it's crazy. We're talking contracts bigger than any sports contracts ever given. And these folks are kind of the modern day, modern day athletes in some way. you know, I think for a company like us, you have to know the game that you're playing, and I think fortunately to some extent we're not exactly playing in this same game that some of these big labs and hyperscalar companies are.

What I mean by that is a lot of the folks who are getting these astronomical numbers are this group of, you know, 100 or so folks who are really at the cutting edge of frontier model research and know the magic sauce of how to get these language models from soup to nuts out the door, and that isn't what an application layer startup like ourselves is really trying to do. We are obviously using language models and using AI in our products, and we need to have people who know the models in and out, but it is much more of a software engineer classic with some AI, you know, sprinkled in experience that we're looking for,

which are not necessarily the folks getting the billion dollar, billion dollar price tags. That's really concentrated to a very, very, very small amount of people. So we're more kind of competing, you know, we're still competing against big companies for folks like that. It's more of like what startups have always had to do when competing against big companies. Maybe you could double click a little bit about what consensus is and who your product is for. Yeah, yeah. So consensus is an AI search engine for scientific and academic research.

Think of us like if anybody's ever used a Google Scholar or a PubMed, at any point in their academic or professional lives, we're trying to build the 2025 AI native version of those tools. Another way to think of it is, like, super verticalized perplexity for a specific document type and for a specific user and use case. So people who are trying to do academic and scientific research. So what that means in practice is the folks using our tool are lots of students, academic researchers, academic faculty members, those in industry. A lot of clinicians use us to do some scientific research, sometimes also answer clinical questions. And then folks across industry, we have a lot of R&D workers at biotechs and pharma companies.

even R&D workers are like CPG companies anywhere where real scientific academic research is being done, we usually will have quite a bit of pocket of users using consensus. How big of a TAM or market is this? Yeah. So, I think there's, you know, any way you kind of slice up a TAM, you're making some things up, but to give you a few ways to kind of look at it. So the number quoted a lot for knowledge workers is about a billion users. And I don't think that academic or scientific research applies for all billion of those users.

But if you look at some of those reports of like what are the personas and roles within that billion, about 500 million of those billion have some use case for academic or scientific research. So that's folks in academia, folks in healthcare, and then some miscellaneous industry jobs like R&D pharma, even in financial services if you're in the healthcare or bioscience world, there is a use case for a tool like this. I think the total addressable market is about 500 million end users, and then as a good proxy for that number, Google Scholar, based on their traffic, does about 50 million monthly active unique users, pub meds somewhere in the 20-ish million number. I think of that being some fraction of that 500 million, using it on a monthly basis, that about makes sense. So I think without even really expanding our market, which I do think that we can do, of making it more accessible to use research, meaning there's more people who might benefit from using research in the work. I think we can't extend that 500 million number.

But I think that's a good starting point of people who need insights or need to use these papers today. I think it's about 500 million ad users. And I used Google Scholar when I was in grad school. How do you guys improve upon Google Scholar and maybe talk to me about a couple of use cases? Yeah. I mean, I think this is one of the most exciting things about our business is that we are competing against a product that's been frozen in time for about 20 years now.

Google Scholar was super innovative, and I think there's a whole interesting story about vertical search. If it was one of the first real used vertical search products that split off of a general purpose search engine back in the early 2000s, when Google split it off, which speaks to the need of a specialization in this use case. But because it was split off, they actually demonetized and they no longer put ads in it, and they don't make any money off of it. So it's really kind of just been maintained and continued to run by a very, very small group of people within Google and not really prioritized by Google. because of that, it hasn't really changed. So if you use Google Scholar, it's actually kind of like a fun way to see what Google used to look like. It's the same interface. It's the list of blue links.

There's no summary put on top. There's no real great interactability with the results. It's a list of blue links to your query. It doesn't do that well with the natural language query still. It's really built for keyword searching and finding papers. So I think the simplest way to just improve upon that is what we partially do, which is take all of these new modern practices of building search and analysis products. exist with the advent of these language models. So that's pulling information out of those papers

to give you into this nice engaging, synthesized way. I think there's a huge, just thread to pull on there beyond just the summary within line citations. Like if you use our product, we give lots of different visuals of, you know, visualizing the results below. So whether that's showing them in a table, showing this like aggregator count of papers that agree with a certain stance. Key papers or key authors will use models to pull all that information out from those

papers and give you in that kind of like summary analysis. section up top. That's just like kind of table stakes in AI 2025 searching analysis products, but that's something that luckily our main competitor does not do. And I think there's a lot to do beyond that as well, think of them more as like workflow oriented features. So Google Scholar really just is, again, a list of links for you to then go interrogate yourself. But there is a lot more depth to somebody doing a literature reprocess than just that and a lot of actions that need to be taken following a list of search results. So that's integrating that with a reference manager to store the papers to go into further.

That is potentially diving into one particular paper asking a bunch of questions of that while still not losing your place on the search results page. So think of it as like post-search, post-getting a high-level analysis, what happens next? We can keep streaming together features and workflows to make that more seamless. Google Scholar's done none of that. So again, all the way back to the top, we're super lucky to be talking about a product that is pretty frozen in time. So anything we do beyond a list of links is differentiating between Google Sky. And it's intuitive that you're not competing against the same engineers as Med and OpenAI, as you mentioned. What's not intuitive to me is who in AI is going to win from a vertical and horizontal approach.

These LMs are every single day they get new capacity. They're able to do deep research. And how do you keep yourself from getting your lunch eaten from the Open AIs and Crocs and Gemini's of the world? Yeah, I mean, I think this is a question that is on every single founder and investor looking at the AI application layer of where is that line between horizontal and vertical that makes sense. And I don't think anybody really knows the answer in the world of AI products today. I think we're getting some pretty interesting data points start to come in because OpenAI has over the last year started to launch a few tertiary products beyond their just general chatbot. So, like, an example of one is they recently launched, like, a meeting recorder transcription extension of chat GPT that directly competes with, you know, your granola's of the world that are building that built into your, your Zoom meetings. From what I can see, that's done basically nothing to disrupt a company or product like granola, and everyone still thinks that granola adds all these additional features of delight and this additional depth to the product that can still be there.

If that's the case, I think there's so much room for vertical products to continue to exist. If a product that is as simple as not to trash on granola, people freaking love granola, but a product that is as simple as a median transcriber in your Zoom window can't be fully disrupted by these horizontal models, I think there's so, so, so, so much room for vertical products to live. And then I think to the other part of the question of what can products like us do to not get our lunch eaten by them? And it's, you know, I think it's staying focused on the problem that you're solving because everybody has a finite set of resources. Everybody has a finite set of focus, yeah, finite amount of focus they can give. And even the most capitalized, smartest people in the world can really only truly be great at a finite number of things.

So your moat against big players is your focus in getting into every nook and cranny of your problem of what your users are facing. There's never not going to be a market for that if you do that incredibly well, even if intuitively some of these products should be swallowed up by a capability of a model. There is going to be someone. I'm not sitting here and saying that the new capabilities and as these models keep getting better won't make some products obsolete. We have seen that happen with some.

I think people generally overestimate how much that will happen and how much surface area there still is to build vertical products. And I think we're seeing evidence of that still some today as open AI continues to launch out new product lines alongside chat chit that don't seem to be ripping successes yet. Yeah, the meeting recorder versus granola use case is an interesting one. Why is Granola able to delight users in a way that the OpenAI product does not double-click on that? I'm not a Grinola user myself. I have teammates who are, and they love it.

You know, but I think I can still answer the question without even being a power user. That it's back to what I said before. It's when you truly focus on something and you can show a user that at every, step of the way you are there to solve the problem that they want you to solve, that is a product that is. You can feel it when you use a product that this is, you know, the fourth priority of a company. It just doesn't feel the same of the whole stepwise process, even if sometimes that core function is the same, all of the tertiary stuff and the product messaging on onboarding on the emails they then send you after you sign up in what happens after a call and how

they deliver you back that information. If all of that is done maniacally detailed, focus on the particular problem, it will feel different than somebody else working on that same problem when it's their fifth biggest priority. There's so many little things that add up and everybody wants to know that you are there to solve their problems. You mentioned him earlier. Nat Friedman has a great quote that he said to us, you know, I could hire somebody off the street to clean my house and they would probably do just about as good of a job as somebody who has a cleaning service, but I still go to the cleaning service because they might just be 10% better. I know they already have all the supplies because they're there to solve my problem.

They're marketing towards me. They communicate in a way that is designed for this exchange. And I'm willing to pay a little bit extra for that for me to know that that person is there to specifically solve my problem, even if the core capability isn't that differentiated. And that I really think does exist in software products. It's kind of interesting. If you take it from the framework of scarce resources, the scarce resources, the scarce resources that startup has is caring a lot about a problem and having very smart people going after

that product. So yes, in theory, OpenAI could bring in some of these people that they're paying $100 million a year to focus on this side project. But in reality, they're focusing on the LLM. So what you get is kind of this effect of having the B minus players focus on these products. Or maybe like in the case of Google Scholar, they have just a couple of people doing it as their 20% project. So there's something around that focus and around that having like the very top engineers in the company focusing on that one thing. That's that's so critical. Exactly right. And if you don't feel that massive inertia and energy of the whole company behind you too, even if the people that you break out to work on that product might be incredibly talented and

just as talented as some of the people we might have in our our doors. But if it's all we focus about, we will have an advantage over you. And I really do believe that even the best, biggest and most capitalized companies in the world can really only truly be exceptional and best at a pretty finite number of things. And that number is usually smaller than what people think it is because you need that inertia of the whole organization behind you to really truly build great products. It also really answers the, you know, generational, it also answers a question that so many people ask in the startup world, which is why can't Google come in and do this or Facebook? It's that concentration, that focus that could only be done within the concept of a startup

and within the incentive structure of a startup where people really care, people have that equity. Exactly, right. I mean, this is the same trope that's existed in startups for decades. It's just faster-paced and more on display as we're in this new world with language models, but it's really the same debate that people have always had. And the case has always been that there's room for disruption and there's room for startups. And we didn't even talk about the tolerance for risk, too, involved in all of this and the advantage that startups get of that, of the ability to launch and,

and put things out there and have a risk tolerance that big companies just don't have. I mean, Google, one of their demos, whatever it was, one or two years ago had, like, a wrong answer of one of their models and their stock price dropped by 8%. That's $100 billion. And that's by making one, you know, that is the risk that they have to deal with, that, you know, if they put an LLM into Google Scholar and it summarized the paper that said vaccines cause autism and somebody took a screenshot and it with viral on Twitter, they might lose $50 billion of market cap.

Is that risk worth it for that? like their risk tolerance of doing innovative things just is lower than what you get to do as a startup and allows you sometimes do more interesting things and build products that they'll never build. Ever wanted to explore the world of online trading but haven't dared to try, the futures market is more active now than

ever, and plus 500 futures is the perfect place to start. Plus 500 gives you access to a wide range of instruments, S&P 500, NASDAQ, Bitcoin, gas, and much more. Explore equity indices, energy, metals, 4X, crypto, and beyond. With a simple intuitive

platform you could trade from anywhere right from your phone. Deposit with a minimum of $100 and experience of fast, accessible futures trading you've been waiting for. See a trading opportunity. You'll be able to trade in just two clicks once your account is open. Not sure if you're ready, not a problem. Plus 500 gives you unlimited risk-free demo account with charts and analytics tools for you to practice on. With over 20 years of experience, Plus 500 is your gateway to the markets. Visit us.plus500.com to learn more. Trading in futures involves the risk of loss and not suitable for everyone. Not all applicants will qualify. Plus 500. It's trading with a plus. So outside of competing with Meta, OpenAI, and Alphabet and all these companies,

you're also developing products in a hyperscale market, meaning by the time you've released your product, the AI market has gone to its next iteration. How do you develop your product in such a hyperscale environment? Yeah, it's a great question, and it's crazy, a crazy, crazy world we live in. I think, number one, there's absolutely just no substitute for complete urgency, really, really hard work and trying to move as fast as you can be possible, because we are in the fastest moving space in the world. People know how big of an opportunity there is.

We are not the only people that have thought of using out of lens for scientific research. And if you are not truly pushing the pace every single day, you are going to fall behind, whether it's your contemporaries doing similar things or, you know, the models will just be good enough that people could just end up using them for your use case if you don't truly win mine share. And I think back to the earlier point of how do you compete against these other products. Like I think the bare case for startups is that if you don't capture mind share and you don't move really, really fast, it isn't that they're going to launch these super competitive specialized products, but their products could be good enough

if you don't build a great product. So I think it does raise the bar that you have to cross. I just think it is an attainable bar if you move really, really fast and build really, really great products. So I think speed is just, it is more important than it's ever been before in startups today with the moving markets.

And then I think the second part is kind of what you were, you said some of that you were alluding to some of this in your question, but I think it's also, you know, having any mindset of knowing that the models are going to improve, they are going to get cheaper, they are going to get faster, where they are going to get to have longer context windows and building your product with that in mind.

And maybe sometimes wearing, you know, eating a little bit of costs for a few months because we know GPT5 is going to come out Thursday and it's probably going to have the bigger context window. It probably will have small iterations that will be cheaper than anything on market. And we can get something out faster if we just ship it with GPT40

and know that it's going to be a little expensive than a little soul for a bit, but we'll be able to swap in whenever new model comes out in a few months. You have to be willing to make some of those tradeoffs. You have to do it, you know, thoughtfully. and we do have finite resources of cash too, but I think startup generally taking some of those swings

is the right move just with an understanding that things are going to get cheaper, faster, and swimmer. How much of your product development is driven by customer feedback and customer demands versus internally deciding as leadership or as a product manager, this is what the customer should have in the next iteration? How do you balance those two forces? Yeah, give a shout to my co-founder, Christian,

who's our chief product officer. He's a product manager by background. He's a freaking incredible product mind and leader. What he says he shoots for, and I think we do a decent job of this. There's obviously no way to say it exactly, and it's always some combination of the two. It's about 70, 30, 80, 20,

with the 70, 80 being driven by users' requests. And then you layer in, you know, just like general bets that you want to take to give in the direction of your company that users aren't saying. You want to layer in some intuition about where the market is going and things that we should bet on. And then also, like, you never want to fall into the trope of just building exactly what users ask for. Sometimes you have to take what they're asking for and distill it down into a problem and kind of crafted in a slightly different way than maybe they asked for it.

So that all kind of goes into that, like, 30% bucket of what we're doing internally. But the foundations, at least the general directions, should be very, very user-guided, the exact specifics and some other bets you kind of sprinkling. A lot of that can come from your own synthesizing your own market observations and your own goals and desires as a company. But I think 70, 30 is roughly a decent heuristic for what we try to do. So another way, customers always know their pain points, but they're not oftentimes technical enough to understand how to solve those pain points. So sometimes you listen to their problem and not their solution. Exactly right. It's a classic, like, user interview best practice is like always take a little with a grain of stuff. The question is usually never, what do you want us to

build, it's more like, what are you feeling to use the product? And what are you trying to solve for? What is your problem? That is more insightful than strictly just asking what would you like their debate. There can sometimes be really good ideas, but it is more universally applicable when you look for problems, not solutions. As you build out consensus, how do you think about building a moat? And is that even possible in an AI consumer product? Yeah, good, good question. I think for startups, moats are kind of a myth. I think the only like real, real moats that exist are usually distribution and brand. And those are not typically things you have the advantage of as a startup.

I think your quote unquote moat as a startup is your focus, as I said before, and your ability to be narrowed in on a certain set of problems and your speed and ability to innovate and take risks. And I think you just have to rely on that until you truly have. scale of distribution and brand, and that truly is a moat against, you know, upstart competitors. There are obviously exceptions. Like, I'm not a hardware expert, but I know Nvidia has some, you know, multi-year lead on its competition, technologically speaking. There are some exceptions where there's this, like, special technological breakthrough

innovation that you have internally that others don't. Usually that's not the case. In the history of software, it isn't that, you know, Salesforce has some incredible technological breakthrough that some other company doesn't have that gives them this moat. What is their moat is they were they executed incredibly well. They focused on a narrow set of problems. They built a great product and eventually they had the scale of this brand and this distribution that really is defensible. It really is hard to crack through if you're an upstart. So I think as startups like a moat and you having something that nobody else could really do is kind of a myth, but you can

protect yourself from getting crunched by being really focused and moving really, really fast. And eventually, you build these more durable modes over time. Is that a yellow or red flag when a VC asks you that question? It's a reddish, reddish yellow flag. Reddish? Isn't it kind of an interesting thought experiment, though, to think about these things, even if it's kind of a misnomer? Yes.

And that's why I didn't say it's a full red flag. But if all they do is they just look at you and say, hey, what's your moat? That's kind of a red flag. But if they ask you it. Yeah, if they ask it in a slightly more like thoughtful way with some like kind of other threads to pull on, I think it's a perfectly acceptable question of like, how do you want to develop, develop a durable business over time? It's like a perfectly reasonable question. Or like, what do you view as how you defend against some big players?

What do you think your unique advantage is? Like, I don't know. If all they do is just ask for the moat, you're probably talking to an associate who's just on tech Twitter and it's just kind of asking stop questions. No offense to associates. So I'm going to ask you a kind of difficult question, which is to take off your startup founder and CEO hat and just look at it, not even from a venture side, but from an asset alligator side, let's say you're a family office, your institutional investor. How would you play this, quote unquote, AI market? Are you trying to, like, how would you invest into the AI space? Would you do kind of like a spray and prey and know that something's going to hit very big? Are you focusing on a kind of thematic, a couple of themes, or how would you play, how would you invest in the space if you're an asset allocator? Yeah, let me caveat by saying, I'm not an asset allocator.

So I've definitely not, definitely not the best person to ask this question too. Yeah, I mean, I'd say, number one, I don't really believe that there are that particular, like, defensible advantages of any part of the stack, so like the application layer or the kind of like infrastructure layer, the hardware layer, like, I think a lot of the same things are all present across all of them. So, like, I'd be interested in having exposure across the different layers of the stack and not just investing in only one. And then I think within each layer, I think, honestly, to some of like the themes of what

we talked about before, I think, you know, history doesn't repeat itself. It rhymes. And I think all of the same best practices of how we depict the best companies and the best founders in those, each particular area are just going to be true today and look to stick to your fundamentals that way. look for really, really, really great founders, really, great teams who are solving a really important problem, and that's really the best that you can do. And there's going to be ones you miss on, there's going to be ones you hit home runs on. But if you just keep indexing on that,

if people serving really great people solving sharp problems, you'll eventually do pretty good in the long term. And I think it isn't really reinventing the wheel of what exactly that you're looking for. I think the one thing to be caution on is just how crazy some of these rounds can get in our space and knowing if that makes sense for what your goals are of a particular institution or firm like us v is a great example they they don't really do any growth stage they're pretty focused on series a mostly with some some seed some bs and they're you know they're looking for contrarian bets they always have and they're not going to be chasing this you know billion round uh raising a series b billion dollar round in fact because that's just the that is how they've

made their their their hey that's how they know they're that's what they know they're great at and that isn't the way that they're set up as a firm to win is getting into those billion dollar rounds. So I think you just have to operate with your constraints and mostly stick to the same fundamentals that have worked in software investing for a decade. Yeah, it's an interesting take because essentially it's a new market. It's obviously large, but you have to focus on your controllable variable, which is backing the best managers and backing the best founders and let them take you to the promise land, take you to the next trillion dollar business. Nothing about AI fundamentally changes the ABCs of investing, which is backing the best talent going after the best

opportunities. Exactly. And then within the constraints of whatever you're doing as an asset allocator, of what types of firms and stages you want to be giving allocation to, or if you are one of those VC firms, you know, what types of rounds and stages you're going after within those constraints, it's going to be mostly the same fundamentals. Well, Eric, this has been a great deep dive on consensus. Congrats on everything that you've done and look forward to continuing conversation live. Yeah, much appreciate. Thanks for having me on, David.

Check us out at Consensus. Dot app. Yes. How should people follow you and keep up to date on consensus? Yeah. You can sign up and create a free account if you want to check out the product

at consensus. App, APP, or follow us on Twitter at Consensus NLP. You'll see lots of product updates and interesting musings on AI and science products. Thank you, Eric, and appreciate you jumping on. Appreciate you, Dave. Thanks for having you.

Thanks for listening to my conversation. If you enjoyed this episode, please share with a friend this helps us grow also provides the very best feedback when we review the episode's analytics. Thank you for your support.