

8 Oct 2026 · 1h 19m · 24 chapters

VC roundtable on personal AI agents and “agentic” systems, sparked by OpenAI’s reported math breakthroughs and broader implications for AI-driven science, plus debate over valuations (e.g., Instinct, Muse) and what makes agent startups defensible.
Written by AI. May contain mistakes.
Discussion on recent OpenAI math results and their implications for investing in AI.
“I think, you know, the biggest news of today and of, you know, maybe I would reckon maybe the biggest news of many decades, for me at least, is this unbelievable math result from OpenAI.”
Exploration of potential breakthroughs in healthcare and understanding complex systems driven by AI.
“We spent a lot of time in healthcare, right?”
Discussion on the disparities in computational resources among AI companies and the importance of context in AI applications.
“And there's all these other domains that are non-verifiable or domains where, you know, you kind of have to go the additional last mile that maybe you don't think a big lab will do.”
Explore the potential of personalized AI agents and the importance of context.
“that has all of your personalized context on your, you know, biofeedback and whatnot.”
Discuss the evolution of AI models and compute efficiency challenges.
“Labs may not think about it going early, but there is a flywheel to be had in those areas because you're collecting the data.”
Delve into the complexities of investing in data companies amid market uncertainties.
“I am sure that you have seen a 101 different RL environment slash data companies that sell to labs.”
Examine how data is changing the landscape for AI startups and their founders.
“reinforcement learning, that type of thing.”
Analyze how the skill sets of founders are changing in the current tech landscape.
“they keep serving them until they become, you know, more recurring and deeply embedded.”
Explore how the skill sets of successful founders have changed over time.
“And how does that change how you think of what a good founder looks like today?”
Discuss the rise of young, innovative founders and their impact on industries.
“We take a very bimodal view of how we invest in some of this.”
Examine how AI influences the decision-making and capabilities of founders.
“And so I think we all have to reassess what the high quality founder looks like, but we're all learning.”
Analyze the hurdles personal AI agents face in gaining user trust and distribution.
“I was recently out of my tech bubble and I was at a friend's wedding.”
Dive into the competition between AI products Instinct and Muse, their features and market expectations.
“So Instinct, you don't think this will overcome distribution.”
Learn about various practical applications of personal AI agents in everyday tasks.
“Yeah, I think a billion users is a big number.”
Discuss the advantages of personal AI agents and address concerns regarding data privacy.
“I have to go around the table and ask everybody what their favorite use case for themselves.”
Examine the potential risks and societal impacts of relying on AI for decision-making.
“for any of cases where I would have to share any personal information like my credit card or SSN.”
Explore the complexities of investors backing multiple competing startups in Silicon Valley.
“Like how Google search was designed for humans and you would go down like one page, two pages, and that's it, like that was a search.”
Exploration of the unique cooperative yet competitive nature of Silicon Valley.
“I wonder if at some point founders say, hey, this is bullshit.”
Discussion on the complexities and ethics of investing in competing startups.
“And then there's some big personalities.”
Analysis of the changing landscape of SaaS businesses in the AI era.
“Um, we have had it happen where companies sort of pivoted into a closer competitive nature than we had originally planned.”
Insights on how investors keep current with technology and startup trends.
“where we thought those multiples made sense and held up.”
Panelists discuss their thoughts and concerns about technology and startups today.
“What are you thinking about, worried about, trying to learn more about when it comes to tech and startups today?”
Discussion on how seed funds should adapt their strategies in the evolving market landscape.
“I mean, I think you've got to go to the source, right?”
Panelists emphasize the need for investors to grasp technology to cut through jargon.
“Or are we okay investing in companies where the outcome is a billion dollars or less than a billion dollars?”

Get the gist, explore an idea or ask about something said.
On This Week in Startups, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00Hey, everybody. Welcome back to This Week in Startups. We're doing a VC roundtable, and you may notice this guy doesn't usually do the VC roundtables. I know Jason is traveling. He is abroad this week. He is not here, but I'm not doing this by myself. That would be ridiculous. We've got an amazing guest host stepping in. He is a partner at Menlo Ventures. Give it up for Deedy Das, our special moderator of the week. Deedy, thanks so much for being here. Thank you so much for having me, Lon. These are quite big shoes to fill, as you know. I am not the world's greatest moderator, Jason J. Calle is, but I will try my best to entertain all the lovely folk on stream here today.
0:37We certainly appreciate your expertise, stepping in as a guest host. That's going to be amazing. And we've got an all-star panel for Didi's big week here. First up, Kanu Gulati. She's a partner at Coastal Ventures. She's led investments in PolyAI, Wabi, Sarvam, so many others. Kanu, thanks so much for being here. Very excited to be here. Awesome. And we've got Chris Farmer. He is the founder and CEO of SignalFire, Grammarly, and Alchemy, among the notable investments I see here in my notes. So, Chris, awesome to have you. Thanks for having me. And finally, we've got Shiel Monat. He's the co-founder and general partner of Better Tomorrow Ventures, seed stage firm.
1:13They're investing in fintech and vertical AI startups and a wonderful returning panelist to the show, one of our favorites. Shiel, thanks for coming back to the show. Glad to be here. Before we go into our big stories, I do want to take just a moment to thank our friends at Plaud. You may notice I am wearing my Plaud note pin right now. It helps me keep track of everything we talk about on the show. Also, just all the conversations happening around the office, anything that I need to keep in mind. It is constantly listening in only when I tell it to. I can turn it off if I don't want to share what I'm doing.
1:43And then it turns everything into these amazing AI-generated reports and transcripts, whatever you need. You never have to like, oh, what was that person saying? I forgot to do the thing. It's just recorded. It's saved. It's in your notes. It's ready to go. So you definitely want to check it out at plot.ai slash twist. And if you use the code twist, they're going to give you 10 % off of your plot note pin. There was a look at one of those amazing transcripts. So thank you to our friends at plot. And now, finally, we can begin the VC roundtable. Didi, where do you want to go first? I think, you know, the biggest news of today and of, you know, maybe I would reckon maybe the biggest news of many decades, for me at least, is this unbelievable math result from OpenAI.
2:28You know, am I qualified to talk about math? No. However, you know, there's this meme that goes around which says, hey, did you read the latest Ryman hypothesis result? And everyone's like, yes, I did. Did you? It's like, yes, I did. Did you understand any of it? No, I did not. Did you? No, not really. But it's cool and it's a big deal. And this is what I wanted to talk to, you know, the Shiel, Chris and Kanu about is we're looking at these math results that, let me just start with some headline stuff. seven Millennium Prizes, price problems exist. One of them was solved 20 plus years ago. And if you look at this result combined with, say, the Navier-Stokes result that OpenAI published a while ago, along with some other researchers, you're looking at major mathematical progress in four of the six remaining unsolved, biggest unsolved problems in mathematics with AI.
3:29And there's many ways we could sort of take this discussion, but maybe I'll just start with Kanu, who's at Khostla, an open AI investor. When you see results like this, how do you think about investing in anything? Or how do you think about startups and innovation? And what continues to have value in a year or two years when we're looking at, hey, the smartest minds of the world can't do what AI can do today? We at Kusla Ventures are very proud early investors in OpenAI, and I'm very excited about all the results. But as a former scientist, it does make me uneasy because it is moving at such a fast pace.
4:14There's an amount of, just the pace of advancements in all of these different sciences is going to continue to be mind-boggling. I think, you know, in the past, you know, Sam have identified that, hey, people are usually not great at thinking how when things compound. And we're already at that inflection point where AI and AI's capability and generalization and being able to solve some of these unseen, unopened math problems is just compounding. So your question is very fair. You know, from an investment point of view, we often think about, you know, what can we invest in where the ultimate value is not just flowing into the token generators today?
4:58What is the next capability that we can invest in? And I can talk about some of the things that I have been investing in, which I feel could be interesting, but it is a very fair question. And the, you know, for example, even within related to the math area, right? So math happens to be a domain that is very formal. And for AI to continue to build proofs on top of proofs is you can see why it's working. There are domains which are very informal, which are, you know, connected, written in very natural language with the intent hopefully captured well, but it is not 100 % correct. And so some of the companies I've been investing in are trying to take real world, very complex domains and formalize it such that you can do this advanced math.
5:44And so you always think about what are we investing in and will it continue to get better because of the underlying models getting better and all of the capabilities that they are producing. So, you know, Pramana Labs would be one of them that is going after taxes. Logos Research out of Imperial College in London is another one that's going after financial code. But the intent of the math or intent of the code, that once it's captured in a formal domain can move much, much faster. But I see, I think, similarly in chemistry and physics, just in the software or digital land will move a lot faster.
6:18And now we are seeing a lot more companies trying to close the loop with real world experiments, with wet lab experiments, with robotic experiments. And we're seeing more of that. But it is a very, very fair question about what can we invest in early stage that will not ultimately flow into the token generators. And it is getting harder and harder for all of us. Jill, Chris, where would you agree, disagree, continue that or say, hey, here are some areas that I look at or I think are still valuable or, you know, it's a free for all. Who knows what we're all doing? And we'll see. What's the take?
6:48We spent a lot of time in healthcare, right? I think there's going to be just unbelievable breakthroughs in that ecosystem of understanding a lot of parts of the human body and life that we still don't understand well today, like the brain and complex systems and cancers and all these types of things. And I think, to Kandu's point, it's hard to anticipate how powerful this gets as they compound. But as you see a convergence of all the different investments that are being made across life sciences, wet labs that can test experiments really quickly and run through it, you know, capturing all of the data that you wouldn't have historically kept in a human context because you only put something into clinical trials if it's been successful.
7:35Whereas in an AI world, you want to capture all the failures as well in order to train and learn from those data sets. So there's a lot on the data capture and reinforcement learning and whatnot, and all these categories. And as these things converge, where we get better at understanding complex systems, but at the same time, I get better understanding biology or chemistry or all these things as these different types of models converge and can collaborate, the breakthroughs that I think we're going to be able to have is going to be super powerful. And I think a lot of what I'm most excited about is a potential impact on human health, right?
8:10And then also planetary health and other super complex systems that, you know, there's a lot of debate on what's real and what is climate change and all this type of stuff. And it's because there's so many, you know, inputs into these systems that it's hard to sort of really understand the butterfly effect of different things. And so now I think as we get better, these models get better and can collaborate across different disciplines, there's gonna be a lot of exciting, I think, breakthroughs and better understanding of the ramifications of some of the choices we're making in these domains. Yeah, I completely agree.
8:43I feel like it's actually the math stuff is obviously the first the first results we're seeing. But that's just because it's the place where it's easy to verify that the AI discovered something actually new. But I think what's very exciting is just extrapolating that to all of science and making invention itself cheaper. and I think like Kanu and Chris said I think there's going to be a lot of implications in other sciences I think cancer therapies, stuff like that but I think the implication is that AI can create a cancer therapy quickly but I think there's still a lot in proving that it works safely in humans and I think that's where we've done some investing I think Similarly, in chemistry, you can design a better battery chemistry, but someone still has to manufacture that battery and test the battery.
9:41And I think that's why you're seeing so much in physical AI. Odoo is the all-in-one management software that brings everything you need into one easy-to-use platform. And founders know it's going to save them a lot of their time. And time is money. Time is valuable. That's all you have as a founder. Let's take a moment and we're going to explore one of these real apps you get access to when you sign up for Odoo. Let's check out Odoo Studio. Odoo Studio lets you customize or reshape any of the tools on their platform, or even build your own original productivity app, all without knowing how to code.
10:14So if your software company needs a specialized customer pilot and onboarding tracker, you can easily knock one out yourself in Odoo Studio. Or if you're constantly lending out laptops, hardware, or microphones like we do in our production facility? Well, you can build your own equipment checkout and asset tracker in just minutes. Why wait for someone else to invent the software your company needs? Build it yourself in Odoo Studio. If you're cobbling together your back office across five different apps, get started today at odoo.com slash twist. And your first app is free. That's odoo.com slash twist.
10:53You know, maybe one thing that I struggle with, and I agree with all of you guys, right? Like, so this is true. And there's all these other domains that are non-verifiable or domains where, you know, you kind of have to go the additional last mile that maybe you don't think a big lab will do. But the thing that I struggle maybe to be very honest with you the most with is you look at these big labs and well, compute is clearly very constrained in the world. And it is, you know, it's a rich person's game. I mean, there are certain companies who are able to commit billions and billions of dollars to compute.
11:30And their compute efficiency, defined by the revenue they can generate for every incremental dollar of compute that they purchase, is probably the best in the market versus any other kind of startup that tries to compete. So it's almost an unfair balance of intelligence in many ways, where how could a company possibly compete with an open AI or an anthropic who has, you know, hundreds and thousands of thousands of GPUs with, you know, a paltry, maybe 5 ,000, 10 ,000. How does one, how do you think about that? Like, does that sit with you or you still believe, no, you don't actually have to have the most compute to be able to do this?
12:14How bitter lesson-filled do you think you are when it comes to that sort of thing? I'm not at all a believer that the models are just going to own everything. I think that's intelligence. If you'd isolated that, that's great. And it's obviously a critical foundational input and incredibly valuable. But how you apply that intelligence and how you give it context and permissioning is going to be critical. There's a reason why Ford deploys engineers have become such a thing in recent years. There's a lot of change management, a lot of guardrails, a lot of trust building that needs to happen that a lot of companies are going to have to wrap this, you know, the intelligence in context and permissions in order to actually do anything agentic, right?
13:01And so the change management piece is non-trivial. Then there's tons of proprietary data sets. We do a lot in vertical AI. We used to be probably 80 % horizontal SAS and, you know, with a heavy ML element. And then we flipped to 80 % vertical AI because I think you can build data density in lots of different use cases where the model may have intelligence, but it doesn't have anywhere near the context that you have because you've built that up, you're going to outperform dramatically. And that can take many shapes. For example, we have a company called Evenup in the personal injury law space. I mean, they're the dominant player in that domain because they've seen all the different cases and back and forth with negotiations with insurance.
13:44And 99 % of that data is in the private domain, you know, the big models, like, what are they going to do against that with somebody who has so much dominance in a vertical? Or, for example, a company like All.Health, which is a clinical grade wearable, right, that has all of your personalized context on your, you know, biofeedback and whatnot. Like that, you know, the models help unlock that data, but it's that personalized data for early detection of issues that you may have in heart or whatever, that you need to augment the intelligence with context on a person or an industry or whatever it is.
14:25And so I think there's lots of opportunities to organize and unlock intelligence through the workflows, context, permission systems, guardrails, security, all of the things that need to be wrapped around it to actually get that horsepower to the ground. There are a few different things I'd point out. I am obviously a believer in large language models, the way the transformers are built today and their scaling laws because those have worked and all of the large labs are going down that path. I do think there is a ton of opportunity to improve them. You're still in the very early innings. These models are very inefficient.
15:01The amount of data they need to train, the amount of compute they need to train, the amount of compute they need to serve, they're all very, very inefficient. And there are tons of very interesting work happening outside of these labs in solving some of these problems. I think some of the, you know, opportunity that Chris mentioned is very true today because there is limited context that these models can work because the underlying compute infrastructure is still very expensive. If there were a way to have very, very long context possible, it's not limited to 1 million tokens as of today, but maybe 100 million.
15:33And maybe, you know, maybe how you surface that information and what you really need to retrieve ahead of time, et cetera, all of that changes. So I think there'll be a ton of different algorithmic approaches that will improve. And it might come out of labs outside of, you know, the labs you mentioned. And I also think, you know, already because of the current lack of HBM or current lack of resources, there is innovation happening from open labs. And it's partly driven because of the constraints in the current ecosystem, but there is innovation. So there's a ton of innovation that's going to come out.
16:11I also don't believe that LLMs and transformers will be the only way for us to get to, you know, whatever we consider AGI or, you know, SI, et cetera, whatever the current term for it is. It's going to be different types of intelligence, different types of approaches for artificial intelligence. As I'm very bullish on, like, there will be a ton of opportunity. And then there's this whole other thing about, you know, like maybe in specific areas where you need data collection from the physical world. Labs may not think about it going early, but there is a flywheel to be had in those areas because you're collecting the data.
16:48For example, we have a weather modeling company called Windburn that collects in-situ data from different points in the atmosphere. And they will just be ahead of anybody else collecting those data points. So there is a physical world data flywheel opportunities as well, and we've been investing in those. And I can go on and on about that. But I think, yeah, I don't believe that just scaling up will be the only way for us to continue winning the ultimate challenge for the next couple of years. Yes. And I also believe that the short, the constraints in compute and other components is may not be it's not like five, seven years from today and probably for the next two to three years.
17:26Yes. Yeah, I think at some point compute stops being the bottleneck in all of these things. like we talked about, like you can create a bunch of promising molecules. And, you know, right now, something that costs$50 million to create, you know, a few years from now might be in the thousands of dollars. And it might not be that compute is the constraint. Actually, it might be lab capacity, the ability to do clinical trials, manufacturing patents, all that kind of stuff might be the constraint. So I think right now it certainly feels like the big labs are where this kind of research gets done, but I don't know that that necessarily needs to be the case in the future.
18:03Can we just double click on that real quick? What do you think, Sheila, is going to change that's going to make, are we just going to be able to make a lot more chips? Is it going to be the chips get way more efficient? Is it just we're going to build a lot more data? What's going to happen that's going to make compute become so much more widely available? Yeah, I think it's all of the above. And sort of like the frontier model production is like what we're relying on today. It need not be that. It may be that today's model a year from now is going to be much cheaper, and you still may get output from it that is relevant.
18:37There's so much to be said about compute. There's another. So when I look at labs, I think of, or generally, let's talking about AI. There's data. There's compute. There's maybe algorithmic improvements. There's a couple of other things here and there. There's this tweet that I saw this morning that I thought was interesting. It's a tweet. Maybe we can pull this up. Uh, Gokul says, um, many of us may know, uh, Gokul already says, you know, the, the fastest company to go from, uh, zero to a billion in revenue run rate is actually not any of these guys that you've heard of. It is actually a data company.
19:13And, you know, Shio, Chris, Kanu, you, you guys are all seed investors. You guys see seed investments all the time. I am sure that you have seen a 101 different RL environment slash data companies that sell to labs. And I'm sure you've also seen companies that come and say, hey, you know, we're at$100 million in revenue or, you know, an enormous amount of revenue in a very quick amount of time, most of which is typically non-recurring. But how do you how do you assess this? Like, do you invest in those kinds of companies? I've heard a mix of different opinions from sort of my investor friends and various different people.
19:56Some people are like, look, I don't touch that kind of stuff. I don't believe it has long term value. Some people are like, who knows what has long term value? If anything, I think data will always be something that labs need. So, you know, this is as good of a bet as any. And then yet others caveat it by saying, yes, but I don't know what value I would pay for these things because they definitely don't trade like software businesses. They're non-recurring. The margins vary, to say the least. How do you guys assess this? Would you guys invest in this kind of stuff? And if so, what type of data companies would you invest in?
20:33I know a lot of you out there listening are wanting to start your own company. So if it's finally time for you to make your dream into a reality, you can't find a better partner than our pals over at Northwest Registered Agent. Northwest Registered Agent is now offering free identity services. That means a free domain name, open source web hosting, a business email and phone number and everything else that's going to make your new startup look and feel like a professional company, all with no purchase required. And this is free done right. They're not passing your data along to any third-party vendors, and they're never going to sell it.
21:08So if you have an idea you can't get out of your head, or if you're already building something amazing, get started with Northwest for free and upgrade as you grow. Learn more at northwestregisteredagent.com slash twistdomain. We're investors in companies like Micro One, you know, that are doing a lot of training, reinforcement learning, that type of thing. There's just so much capacity to do that. I think you have to get to the frontier. You've got to be at the front lines, understand the nature of this data, what's working, what's not working, how to apply it. And then great entrepreneurs at the end of the day is what we're backing to figure this out, right?
21:52And to see where they can migrate to corporations to help them adjust and tweak their data in order to apply it in their particular personalized use case. You know, I think it's going to move from the labs to a broader set of end customers. You know, and so we have a number of companies doing this in different verticals and whatnot. And yeah, the growth is explosive and it's very concentrated in a handful of labs today, But all of them have a thesis of how it will be applied to broader enterprise, because I do think that as you move to more open weight models, you've already seen a massive migration per the Vercel data.
22:33That was one of the sort of stunning shifts from proprietary models to open weight models that went from like 80-20 to 20-80 in a matter of months. Right. And as these corporations start to get more adept at this, build the skill sets in house to personalize, you know, for their own use cases and build their own defensibility over time versus outsource it all to proprietary models. You know, I think there's going to be a lot of need for this across any large business. And so, you know, how they do that bridge, there's huge go to market challenges and professional services type challenges. But I think there, you know, some of this will be enduring, but some of it will also, you know, come back down pretty dramatically, maybe as rapidly as it went up.
23:21And so it's going to be a matter of, you know, how adept these entrepreneurs in building lock-in and landing and expanding, just like the SaaS days, right, with these customers where they find the pain points and then they keep serving them until they become, you know, more recurring and deeply embedded. I disagree with the sentiment about data will always be needed because, as we've seen, some of these larger model companies, they're good at taking this data and then at some point they don't need that data, right? Because a model is just inherently able to solve those things. We don't know how well the open-weight models are distilling from the top labs, right?
24:06So the data, if the top labs get it, the open-weight models also get it. So if the enterprises or this existing data is your currency, then you're going to, the terminal value seems not exist. But I think Chris's point on, at the end of the day, it is the founder that they're backing. That, it's very hard to argue, because I have seen founders may have started somewhere, and the story started to work, and then they were ahead of anybody else in figuring out what the labs need next. and they're just ahead and they know how to continue to build a business and continue to be one of the most wanted data or data type or app build on top of that data commodity.
24:45So I think the right entrepreneur can still solve it. We have, as a founder of Coase Ventures, we've stayed away from that category, which we believe like the terminal value might not be there. So for, yeah, we haven't invested in anything in the data category for that reason. And, you know, But there are founders that it's been very hard to be like, oh, are we not going to invest in that founder based on everything we know about them? That's the hard discussion. I think they're super interesting companies. But I think even thinking about headline revenue is kind of like the wrong way to think about these businesses, right?
25:17You have to think about it from a gross profit basis. And then even then, there are a lot of questions. We mentioned customer concentration. We mentioned, is the demand structural or is it temporary? And I don't have a good answer for that. I would say like the models obviously continue to need better post-training data and evals. But so far, it feels like as models have improved, actually, the amount of expertise actually has risen. I don't know if that asymptotes at some point. I think there's a question of like, what is proprietary in any of these businesses? and in many cases there isn't much, but there's a great opportunity right now to generate cash and they actually don't require a whole lot of fixed capital.
26:09So it's a very interesting business. I haven't invested in any of them. Interesting. You know, I like, there's a couple of threads that all of you touched on, which I thought was interesting, which, you know, we talk about this a lot as investors, which is ultimately a large part of every investment we make does boil down to, you know, how talented this founder is. I sometimes wonder, and I've had varying thoughts about this over time, is we live in a very different era today than maybe five to 10 years ago. Namely, five to 10 years ago, you might have been, I don't know, 10 years at a great SaaS company, and then you leave and you say, hey, I'm going to start basically another SaaS company.
26:54I'm going to do some top-down enterprise sales. I'm going to hire a bunch of people. There was usually a little bit less competition, and you could be quite successful sort of running several kinds of playbooks back in the day. A big constraint for many founders was just being able to hire software engineers at scale that actually could build out the software. Maybe you disagree with parts of that, so if you do, feel free to tell me. And that's just a part of businesses. Obviously, many other great companies in that era existed too. whereas today all of a sudden, at least when I talk to young founders, the way they work is just extremely different.
27:34I mean, the size of teams versus the kind of capacity of output that they have, it's heaven and hell. You have teams of five people who are running thousands of agents in the background who do a bunch of different things. They say yes to this, no to this. It sort of seems like a wholly different skill set, at least when it comes to software. you know, leaving aside maybe some healthcare specific or fintech specific stuff where, of course, I know you need some niche skill set. Do you agree with that? And how does that change how you think of what a good founder looks like today? Is it fundamentally the same and you're looking for the same stuff?
28:10Or are the skill sets actually changing a little bit today than it was 10 years ago? Yeah, it's a good question. I think like the broad skill sets of setting a vision, selling that vision recruiting people are all the same as they ever were and and that's what it takes but um there is this new class of founder that that didn't exist before and i i think that they i would say like in our experience at least skew younger um and you have very folks in their early 20s building amazing businesses scaling very quickly um in in categories that you would have thought required expertise in accounting, legal, stuff like that.
28:55People without backgrounds in those places are doing quite well. But I think at the end of the day, what have they done well? They set a vision for their company. They recruited people against that vision and they sold that vision to their customers. And so I think those things are evergreen. You've heard me say it many times now, if you're an early stage startup or small business, every single new potential customer is extremely valuable. You can't afford to have a team that's missing calls. If your communication is inconsistent, your customers will notice. That's why today's episode is brought to you by Quo, Q-U-O, the smarter way to run your business communication.
29:32I use Quo all day long. I love it. And it's just great to have that one number that my team and I can share so that we don't lose any customers. Quo is the number one top business phone system on G2, and it's already trusted by more than 90 ,000 businesses. They're going to bring all of your calls, texts, and contacts together in a shared collaborative space. That means your entire team is going to use one shared number, so there's zero chance anyone is going to fall through the cracks. Money is on the line. Always say, hello with Quo. Try Quo for free. Plus, get 20 % off your first six months when you go to Quo.com slash twist.
30:08That's Q-U-O.com slash This is one of my favorite products in the world. It gets my highest endorsement. We take a very bimodal view of how we invest in some of this. In some areas, you go with super deep domain experts that understand the workflows, the politics, the incentives, the data sets, et cetera, and they have the credibility to get the partners they need to design and build systems to attack problems in particular verticals. But we're also very extensive investors in what S.H.I.E.L.D. described as these young, this is their first job, and they're just sort of hacker-type personas that have no idea how an industry works.
30:51Because we also believe when you have these tectonic shifts and nobody knows where it's going, you need heat-seeking missiles that are just insatiably curious, move super quickly and iterate. Right. And it's somewhat Darwinian and a lot of these things don't work. But when they do work, they reinvent industries to an even greater degree that anyone from the industry might do. So like someone from telecom wasn't going to invent Snapchat or someone from the hotel industry wasn't going to invent Airbnb. Right. So you look at it from a completely new paradigm and a blank sheet of paper. you have some of the most revolutionary advancements because you're not burdened by sort of the prior logic and maybe constraints that the industry had faced historically.
31:37So I think it's a really interesting time for sort of young founders because you can also do so much with so little money, much like what cloud computing did for people to sort of not need all the sort of Apex in order to get off the ground. I mean, what you can do to spin up and iterate quickly with these AI tools is pretty incredible on lock. I think how AI pill the founders are has an impact on how we think about them, because they're going to take the bets that they feel will be the right bets for the company in the coming years. They're making an informed approach about what the underlying capability of the large frontier models will be, and then they're basing the bet on the company on that.
32:21So being AI-pilled is probably, I guess, very, very important access now. Age-wise, I've seen all various ages. Yes, they're the younger founders, but also folks who are clearly in their 30s and 40s and 50s. But because they are able to take the ambitious bets and they think about, you know, we're investing in a chip design company that is going to release a chip with like 20 folks because they have so many more agents that can do a lot of design and exploration and all of the hard work of generating the RTL and also all the way to physical design. I think it's just a new world now. And so I think we all have to reassess what the high quality founder looks like, but we're all learning.
33:04There is no pre-made template like, okay, thou shalt look like this, or thou shalt be at this age. I think we all have to just fundamentally question our known beliefs here. Yeah, I mean, my only thing to add is like, for me, you know, taking the sort of venture hat off, I just think it's spectacular how every, you know, generation of technology seems to exacerbate the skew of outcomes of every young generation. So, you know, one way to say it is, it seems like the median gets slightly worse, but the extremes get immensely powerful. And so when I see some, you know, the other day, I was spending some time with this 19-year-old, he wasn't a founder or anything.
33:48And he just showed, he was just showing me, he was introduced to me by a founder. He said, hey, this guy's really smart. You should chat with him. And he was showing me his flow of how he did things. And to me, it was just remarkable. He had, you know, this tab of, hey, here's a bunch of my 50 agents that are running. And it's just trying to break different websites at once. And then it'll alert me when it breaks a website. And then I get to do something with it. And I can just collect bounties off of this. And I'm like, how are you doing this at 19 just boggles my mind. I wouldn't have even thought to think about this.
34:23So it is pretty exceptional how much it can amplify talent at that level. And I do want to touch on another topic, which comes up all the time. I was recently out of my tech bubble and I was at a friend's wedding. And the only AI topic that seemed to come up at this wedding was how are you using your muse or instinct? And I thought that was quite funny because they talked about nothing else model related. It was just muse instinct. And then some of the hedge fund guys were like, are venture capitalists absolutely bonkers? Why would you fund this for$10 billion? It doesn't make money. Please explain this to me.
35:04And so let's just start with that. How do you, you know, when people ask you, and I'm sure maybe some people have outside of the tech bubble, what is up with these valuations for something like an instinct? is your instinct to defend it and say, hey, you know, actually there's a world where this might actually be an amazing company for these reasons. Or are you sort of like, you know, I don't really know. How do you guys think about it? I'm a bear. Oh, of course. Look, I think these personal agents, I think this is actually going to be an unlock period for personal agents and it's going to be super exciting in the next few years.
35:46I think a lot of what I've seen with other people and certainly applies to myself is there's only so many points of entry into my private data that I want. And, you know, you know, my Gmail has, you know, social security numbers, taxes, like all sorts of things that, you know, it just gets very dangerous at some point to have too many like things plug into that. And obviously, like Facebook has just massive distribution. I was actually looking this up with a founder friend of mine, and it's something like, you know, three and a half billion people on the platform, you know, and Google with YouTube and everything is similar, right?
36:28I think it's even greater. It's like four billion people. It's like half the people on the planet, and it's like something like 70, 75 % of people that are connected to the internet, right? Right. And so, you know, the distribution reach they have, the intimacy they already have through Gmail or Apple or, you know, I think is not is a very, very difficult thing to get over. And so, you know, I think either a more specialized model versus a general sort of personal assistant, I just think it's very non-trivial to overcome getting people to give you access to their most intimate information that is dangerous from a cyber threat type of standpoint.
37:07And I think people are going to really want high levels of trust and security. And I think this is an area that could favor incumbents unless you're highly specialized, like travel or something like that, where you just have real advantage in a particular domain. But we'll see. So I think paying that kind of valuation, even if you have a super slick product, that versus distribution and the ad networks and everything that the huge platforms have, I think it's a big uphill battle from my standpoint. So Instinct, you don't think this will overcome distribution. That makes sense. What about Muse at large?
37:46Do you think Muse is a billion-user product, billion-DAO product? I mean, it could be. I mean, look, I think OpenAI ran into this situation, right? If you look at what they're building relative to like, I mean, it's Google's to lose in a lot of ways, and they seem to be foot faulting again. They seem to be on the balls of their feet for a while, and now they seem to be a little off.
38:13But OpenAI needed to build monetization to bring the cost down for a lot of folks. And that's going to be, you know, some sort of where the consumer is the product, right? And the payer is someone who has a commercial interest in selling something to or showing something to that consumer, right? So they had to build an ad network. Then they're also trying to build hardware devices and sort of the whole stack that like decades of work have gone in at Apple and Google and whatnot. And so, you know, I think it's not just distribution, it's distribution and the ad ecosystem and the wallets and the, I mean, like all of the different things that you need in order to do everything to commercial transactions, to advertising and the network of commercial relationships that you need in order to monetize this, in order to subsidize it for the consumer to make these agents, you know, free or very low cost for them.
39:08And so I think there's lots of layers of the stack that the incumbents have. And look, startups beat incumbents all the time. But I don't think paying that kind of entry valuation, to me, the risks are significant, even if the reward is significant. And it's hard to justify that risk-reward sort of trade-off in my mind. I think it was a bit of like tail wagging the dog. The company had to raise so much money because compute is expensive right now. And I think you have to believe if you're investing at that price, you have to believe that this is going to be at least one of the major consumer interfaces.
39:48And so if personal agents are a major consumer interface that we use for travel purchases, restaurants, those are the things people are using for right now, insurance, a bunch of others. You can certainly make money. I don't doubt that you can make money doing it. I think where I struggle is that Instinct will be the winner rather than OpenAI, Google Meta, maybe Apple. And of course, we've seen since the time that Instinct came out a couple months ago, we've seen a whole host of other agent companies come along. And obviously, not every one of them will be able to raise the amount of money you need on compute to be successful.
40:28um i feel like instinct did a tremendous job at their launch and um at least four vcs got a lot of hype early they were out a little bit before muse and that was great um muse is a phenomenal product uh you've seen it in meta stock price um but instinct you know i i personally use instinct all the time um and i've tried others hark came out yesterday uh sesame has a product there's several other products that I saw just launching today people are constantly messaging me hey try this one I actually have personal agent fatigue where I'm pretty reluctant to try another one and again input all of my data so it'll be interesting to see what happens I personally think that the large companies have a huge benefit here but I can understand why you would underwrite instinct to a 10 billion, which is like, what are the odds that this becomes the major consumer interface that's worth a lot of money?
41:30And if the odds of happening, you know, you sort of like build some distribution on what, how the likelihood of that happening. I think it's tough though. And then just to follow up, do you think Muse is going to be that say, say instinct, maybe struggles to compete on distribution. Do you think Muse actually hits a billion, you know, daily users who continue to use personal agents? Or, I mean, and I asked this question in the vein of like, or, you know, is this sort of a first world type problem of like, people have to book flights and restaurants in the Bay area, but most of the world really doesn't care that much about this.
42:01How do you, where do you stand on that? Yeah, I think a billion users is a big number. Um, and I don't think that there's a billion DAUs for this product, but there are a lot of use cases out there. There are a lot of like, in a lot of the world, you still have to like enter information into a shitty system like i've used it for a lot of medical things um and actually interestingly enough um like i used it to pay a health care bill recently and um muse wouldn't do it because it only has my tokenized credit card data um instinct stored my credit card data in plain text which is not great but it actually was able to pay my medical bill in a way that Muse was not able to.
42:45But I think there are certainly hundreds of millions of DAUs that this is a product people can use. People talk about it as if like only VCs have this problem. And it's true. People don't travel as much as we do. But people do make purchases. I, over the weekend, was at a bachelor party and I actually used it instead of Splitwise. It actually automatically knew what purchases I'd made. And it was just like, and it had a spreadsheet of everybody who was there and what times they were coming in. So it actually automatically split everything amongst everyone instantly. Pretty useful thing. Again, not something that happens every day, but it happens often enough for everyone.
43:26I think the category can become really, really large, right? This could become the way that people interface with anything digital, right? Like, you know, if I just work with my agent, I happen to be trying out Vajo, which is a KV portfolio company, their agent foe. If I don't need to call another physical business to, I don't know, organize my eight-year-old's birthday party, I am happy, right? Like, all the way, you know, Vajo's take is they can go even recruit the human they would need to go carry out a task. If somebody else can just take over all that from my plate, I think the value I would pay could be very high.
43:57So I do think that if their company gets the trust, gets into the workflow of our day-to-day lives, and like Sheila said, we have the fatigue. We don't want to continue to train different agents and get them connected on how we think about and what our preferences and how we would like to do work. So there is stickiness possible here. So I do think the category can be very, very large. I am not a consumer investor. and so coming in at a 10 billion kind of a price would be like just something I would not think of so very early bets on trying out something very different approach you know I did make an investment in a company called Rabbit that got a lot of flack because they promised too much but if you look back like this is what they had promised like you would have a device you would just talk to it it would do the task that you need and not worry about individual apps so the but coming in at a very early price to say like hey this is a bet worth taking I would do that a 10 billion price as an entry point, I would be worried about that.
44:51I have to go around the table and ask everybody what their favorite use case for themselves. I know, Shield, you already had a great one for a personal assistant was because everyone keeps asking me this and they're like, look, I can only think of flights and, you know, these really stupid sort of basic things I can already click buttons to do. My personal favorite one so far has been, and I guess it's with an asterisk as it's still yet to be completely done, is I was flying back from this wedding, from the EU, and two of my flights back to back were delayed. So I ended up being six hours late and I was just frustrated.
45:30I find out, thanks, thank you, I'm not going to say which, I use instinct. So I ask, thank you instinct for telling me that I'm eligible for a 600 euro refund because this is how the EU works. So I said, great, I don't want to fill any forms. Can you go and figure out how to get me this amazing refund because my flight's six hours late? And it went ahead and did a bunch of stuff and chatted with an agent and tried to escalate to a real person and then did that and then asked for their manager. And it was just phenomenal and they sent a few emails and it said, look, your case, here's your case number.
46:08We're going to get it for you. I was very impressed. I said, look, if you can automatically get me the money that I clearly rightfully deserve for my flight being late, this is a good use case of an agent. So with that, you know, Chris, Kanu, Shio, anything exciting that you have used instinct or muse for? There's a lot of stuff like that. There's a lot of stuff where there's a little bit of friction and using an agent makes it easier. You do wonder, so many business models are built on that little bit of friction, preventing people from doing it. So you do wonder what changes. I do a lot in fintech and I think a lot about finances and credit cards, They charge you more interest.
46:55If an agent can automatically move your debt from a credit card somewhere else, what happens? There's a bunch of things like that. So I think that's another interesting thing is like, how do businesses change when friction is reduced? Yeah, I haven't used Instinct for the reasons I mentioned, mostly because I didn't want to give it access to, you know, all of the data. But I have built personal operating systems and cloud and whatnot that I find sort of really helpful as like a chief of staff for big, big complex projects that sort of take all the inputs from, you know, all the different, you know, whether it's granola or, you know, calendar and email and all this type of stuff with, with, with all the research I do around a project.
47:36On top of it, I'm also really bullish on personal finance use cases because it's like the bane of my existence of like, got to aggregate all this stuff and then give it to an accountant or whatever it is like just the paperwork on the healthcare side, like that type of stuff is the stuff that not only is it time consuming and requires some level of expertise, and there's tons of intentional bureaucracy and friction in that, but it's also like, very painful, like sort of time on all of the stuff. on things that I don't enjoy doing, that I would have a lot of value for those types of use cases. I'm most bullish and excited about sort of personal finance use cases and just all the administrivia that we deal with in our lives that hopefully we can agentify and not have to deal with insurers that deny things for stupid reasons when they really shouldn't.
48:28I haven't used them for any of cases where I would have to share any personal information like my credit card or SSN. I mean, there was an incident yesterday. Somebody reported that their agent shared their personal bank account information on a company's group, Slack, etc. So I don't trust it. But for, you know, like I mentioned, organizing my eight year old's birthday party, calling up different places who can accommodate. You know, these businesses are open between nine and five. I have other things to do between nine and five, Monday to Friday. I don't want to be on a phone call calling up who I called, what information I gave them, what was the next follow up step, what's I need to fire up a court.
48:58Like all of those very annoying tasks can be now taken care of by my foe. I'd pay for that. Right. So, yeah, I've been using it for very lightweight, but just stuff or areas which are not exciting for me. Fascinating. Fascinating. I guess one thing that I was thinking about hypothetically kind of to shield your point is there was this really viral Twitter post that I saw maybe about a week ago where it was some senior banker. I believe, who kind of made this hypothetical, which I kind of thought was a little bit absurd. But he was like, well, if everyone's going to use a personal agent and everyone says, hey, help me with my finances, then what if we have a bank run?
49:44And because everyone's agents tell them, hey, you ought to take all your money out of the bank and put it in X, Y, Z assets. Do you, you know, I guess, how do people feel? I mean, not specifically about that, but this general high level idea, which is the more that the more diffusion that happens with these AI tools within humanity, more people are going to outsource their decision making, perhaps to this homogenous sort of entity that makes decisions in, you know, surprisingly similar ways across different groups. of people and what are the implications of that for the systems that weren't built for this sort of you know mass homogenous behavior to happen at once um whether it's something as simple as everyone now wants to have reservations to the top hotel a top restaurant in san francisco and so they can't accommodate anything um or whether it's a bank run or whether it's one of like a hundred different things that could go wrong because systems weren't really designed to handle that sort of behavior?
50:49Do you think that's a real risk? And what happens? And how do we control for something like that? I think there's a lot of systems in the world that are built on breakage, right? The very sophisticated, that are at scale, that can hire people to help them navigate, whether it's you fighting the airline to get your$600 back, or it's insurance is a very common one, obviously, like where they deny things that they should really do and put hoops, but there's always been like rebates, all sorts of things that make you feel like you're getting a good deal on something, but then it makes it very painful to actually execute and get what you're due.
51:29I think agents will hopefully reduce some of that friction, right? And it'll become more transparent and more rules-based. Banks will have to become more competitive on the interest rates that they pay for being your bank account. Otherwise, you'll move it to somewhere that pays a higher interest because the friction is removed for doing that. And so I think it will force some leveling of the playing field, I think, in certain areas. And there'll certainly be companies that try to fight back against that. But at some point, when you have a relentless agent on your side, I think businesses will have to clean up some of those elements.
52:09We are not going back to a world where we don't know this capability of AI exists or the fact that an agent can take care of the things that I don't want to do. And if businesses on their side don't keep up, like if they can't accommodate an agent calling and taking the order, et cetera, or like doing the number of orders they can do or trust issues or mentioned bank funds, like the businesses will have to solve it at their end. That's more opportunities for startups to go solve that and think differently. Like what could be the right way? What could be the long-term correct way of having this system work.
52:42We're already seeing it in searches, right? Like how Google search was designed for humans and you would go down like one page, two pages, and that's it, like that was a search. Agents don't search like that. Agents can go to 800 of those links and try to surface our best information. This is, you know, high-level Parag's company, Parag Systems is working on, because we have to rethink how the world is going to exist given now that we know AI and agents are not going away. And I think this is going to be opportunities for businesses to get more AI pilled and serve their customers with matching capability.
53:16There's one thing that I can't get away with this podcast not talking about, which is, man, I woke up, I think it was Wednesday last week to the funniest drama that I have seen on Twitter in a while. and this was a factory v cognition and i happened to be on a plane with a bunch of people who happened to be in venture and we were just smiling to ourselves going like i'm not sure why this is such a big deal but i'm glad it happened because it's so entertaining um and um and then we we we happen to we happen to have you know ganu obviously you're at kosla who was at the center of this spat. I'm not really sure what the spat really even was besides, you know, of course, there was somebody, I think, on the board of Factory who went to Cognition, and that's how it started.
54:15And then, which, you know, I mean, I would say this sort of stuff happens, and usually it's a small thing, but it really, really blew up. Matan posted this tweet. Matan is the CEO of founder of Factory. Scott, the founder of Cognition, responded to it. And then at some point, Vinod says, you know, Factory is a second tier competitor, despite being investors in Factory. And I think, you know, kind of like everyone just wanted to know, like, you know, did Vinod forget that he was an investor? Is that was his true opinion? What happened? Honestly, I know what you guys know, right? Because I'm not in boardrooms of any of those companies.
55:03I do like that the MDs in the firm are going to stand for their founders. But that's all I get to say about it. That's all I know, honestly. I was not in boardrooms or the conversations with Chris or any of the founders. I'm a huge fan of both Madan and Scott. So yeah, I know what you guys know. I think there's an interesting question here around should investors be investing in competitive companies? And that, you know, really for a long time, it wasn't a thing. And then in this cycle, it's super accelerated. And most of the multi-stage funds have multiple investments in competing companies. And I think what I've seen is people tend to take favorites and it's the company that's doing well.
55:53I think in this case, both are doing fairly well. And you have just different partners defending their company. I wonder if at some point founders say, hey, this is bullshit. Like, we're not going to talk to you. That was the case like 10 years ago. And then it became not the case more recently. But I wonder if that changes. Look, I think a couple of things. First off, it's one of the greatest strengths of Silicon Valley, but is also a weakness, is that the amount of, you know, basically sharing that happens across the valley, founder to founder on like best practice and what's going on. And it's an incredibly open place from that standpoint, which is wonderful.
56:33So tons of people are investors in all sorts of other companies, you know, in adjacent domains and all these types of things is, I think, really wonderful to pay it forward. Um, the flip side is you also gain a lot of intelligence and like, where are the lines on, you know, crossing something that should be proprietary and whatnot, and where can you use them in what context? And people are going to disagree. And there's obviously going to be, you know, two perspectives on that. And sometimes people cross into the unethical. Like, I'm not saying I have no sense of what happened in this scenario, but, um, you know, just, it's, it's very inherent that like, you know, that when we have this coopetition is, you know, look at Microsoft and OpenAI or all these types of things.
57:15So their partners one minute and then their enemies and then their frenemies. And now they're, you know, you just have a lot of this dynamic or NVIDIA, right, is the, you know, selling to all the big labs. Now they're, you know, competing on the open weight side, just like, you know, many of the cloud guys are now competing on the chip side. And so you've got just sort of this fierce competition and it's going to spill out and get ugly at times when people feel the pressure, right? And so I think overall, it's generally a net positive because it drives competition and innovation and acceleration of this and transference of knowledge across the whole ecosystem.
57:53It's what makes Silicon Valley so unique is that you've got so many people competing and collaborating and learning from each other. But it has these ramifications. And then there's some big personalities. And PR, I guess, is no PR is bad PR, I guess. I don't know. We'll see. But it's certainly a fight for attention as well in there in the mix. We should mention Menlo put some money into factory this week since since the story broke. So was this was this part of the discussion at all behind the scenes or is just like you're talking about the company and the fundamentals and not what's going on on X?
58:30No, you know, look, as you know, these decisions are typically made before they're announced. So they were two separate decisions. And, you know, I think one of the things, it was funny because after that was announced that we were, you know, invested in Factory, Matan gave us a shout out. Sean gave us a shout out. I had my Cognition friends text me saying, hey, like, do you guys, like, do you guys not like Cognition? And I'm like, look, man, firms invest. Like, I think both of them are great companies. I don't know why you're making it seem like I have to pick a side. Like, we invested in one of them.
59:15But, you know, I think both of them are fantastic companies. And I think there is a bunch of room for both. It is interesting to be a part of, you know, a vocal public spat in some way, though. So I wasn't involved personally in the investment. So I don't know. Kind of waded in there just at the last second. So I had to ask.
59:39I think the question about, you know, are funds investing in competing companies and is, you know, how does that work? I don't think we, at the outset, we are trying to find because we obviously don't want them to fight or want to go after the same talent or the same customer. Like those are hard problems. But companies pivot and sometimes these markets are so large and, you know, the starting point and the starting wedge could be very different. How the companies grow can be very different, especially given how quickly AI agents, the approach and how enterprises are going to think about this capability is changing so rapidly.
1:00:14It'll be silly to say like, hey, we have one robotic investment and that's it and we won't make another one or we have like one investment in agents and we won't make another one. I mean, the world's going to move really quickly. And if you find interesting wedges, and yes, you know, if companies do land up going after similar talent or similar customers, then, you know, what we do at Costa Lunch is we have a very crisp, very clear Chinese wall between those two partners that are working with those companies and none of the information gets shared. and we take that very very strictly and you know it's again we don't start like we want to invest in five of these companies but companies pivot very quickly right now especially when they're trying out different business models and different use cases and if it happens that they're weighing up at the same cases they don't share information.
1:00:59Shiel, Chris do you do how do you feel about that would you invest in competing companies or has it happened that you have invested in companies that happen to pivot to being competitive and how do you deal with that? Yeah at the outset we don't Um, we have had it happen where companies sort of pivoted into a closer competitive nature than we had originally planned. And, you know, there's not much you can do about it. Yeah. I mean, you know, especially, I mean, you want to go in intentionally doing that. And, and I know some firms have like a mortal enemies list and then like, you know, some founders are like, yeah, everything's a competitor, you know?
1:01:34Yeah. You know, if I'm open AI, like everything from like Apple to instinct are all competitors. Right. And so, you know, I mean, and I think you're going to you're going to follow a lot of the same trends like you sort of you sort of find your wedge of entry. But then the world evolves the way it does. And everyone is reacting to those same developments in technology or business models or the landscape or whatever it is. And so they may converge at some points. But at the end of the day, most of these companies do not win and lose based on another competitor. Right. Most of it is execution and team and they find different either geographies or segments or verticals or whatever they end up going.
1:02:17So they may compete for a while and then diverge again. And, you know, obviously sometimes they're direct competitors, but even ones that are direct competitors in one element often, you know, don't overlap in others. So, I mean, you just have to do your best to try to create Chinese walls when this happens and, you know, be very transparent with founders as much as you can to try to manage those dynamics. But it's an inevitability, regardless of where you start, you know, everyone's following the same trend. So they often end up overlapping at some point. And didn't like Grubbot announced today that they will use underlying model, whichever is the best model for their use case?
1:02:56I think that's just how quickly people's, what's important, what is important for Grogbot to win is getting access to the best model. And Elon announced, yeah, this could be Claude, Opus, it could be Codex, Image Journey, whichever is the best model, they would use that, right? So I think it's partly in, yes, you need to have the best experience for your customer, figure out what your flywheel, et cetera, is, and just use the best stack that you can create for them. Another example with X is that, you know, they had the capacity on the data center side and they ended up, you know, leasing it a very large lease to the tens of billions of dollars to Anthropic, right?
1:03:35So they're both competing with each other and supplying each other. I mean, it's just the nature of these things. I have a couple of big things that I think are pretty interesting as well that to talk about that are somewhat related to this. Let's maybe talk about a somewhat easier one, which is we've seen what's happening with Bending Spoons. And there was this Miro acquisition. There was the Airtable acquisition. Some of the SaaS darlings from the 2010s are now not looking as hot as they once were. yet you know like I think you know some people like to be very alarmist and say you know maybe maybe SaaS is dead and yes there's some public market multiple compression on some of these SaaS businesses but at the same time like Figma continues to deliver phenomenal numbers quarter after quarter there are a bunch of companies like that that continue to outperform in public markets.
1:04:35That being said, when we look at early stage companies, as you may see and may know, it is really hard to, I think Hamon from GC says this a bunch in a bunch of his interviews, which is you look at a traditional SaaS business, which is a top-down sell, and it has these 3-3, 2-2-2 dynamics where you grow 3x, 3x, 2x, 2x, 2x year over year, typically because that is just the physics of the growth of the business and how fast you can hire sales reps and sell and quota, et cetera. And those are just not quick enough compared to how quickly some of, you know, quote unquote, the AI businesses or any businesses that charge on consumption can grow.
1:05:18So, you know, what tends to happen is, you know, even if you have a very solid modern day enterprise SaaS business, the talent goes away, the salespeople are like, hey, why am I working here? I should be working somewhere that's going to pay me way, way more. How does that affect how you view top-down SaaS businesses? Is it a concern for you at all? Or is it, you know, this is a storm you have to weather and then you come out the back of 28 and everything's fine again? Or, you know, what is your view? Is this dead? Is SaaS just dead forever? What's going to happen? Is it not venture-backable anymore?
1:05:56I don't think that is ever true. All these things go through cycles. Semis was out of favor for a while. Now it's the new hotness. I mean, defense tech would have been like uninvestable for venture literally by LPA at all for most venture funds. And now, you know, it's a hot domain. So I don't think in absolutes in any of these things. I mean, obviously, the SaaS players have been selling sort of systems of record and workflows and tooling and whatnot. And the AI companies are often going after the labor on top of it. So it expands the TAM pretty significantly because the spend on labor tends to be or magnitude bigger than the spend on tech budgets underneath them for worker productivity and whatnot.
1:06:42You know, but as I said, it also changes other dimensions, right? So you move from the biggest enterprise companies were always horizontal because the sales force could sell to every industry, whereas somebody who was a vertical SaaS player could only sell into that market. But when you flip it in AI and say all of a sudden, OK, you're going to the labor layer, so it's an order of magnitude bigger. And now you can build data density and specialization and whatnot to get to better outputs quicker in verticals than you can do horizontally. you know, like Vertically AI became, you know, an area that became much more investable in AI than it was in the SaaS era.
1:07:20So I think these things all sort of migrate through and it's difficult to predict over the long horizon. But, you know, if SaaS enables the upstream workflows and you can monetize that, I think the historical, as you said, physics of the growth rate of SaaS change. If you have to stay at the system of record level and can't move into, you know, the intelligence on top of it and the actual agency and execution, then then you're probably more constrained and it just has to be factored into the price. It doesn't mean you can't invest in it. It's just the capital efficiency and the price that you enter, you know, has to has to be, you know, adjusted accordingly.
1:07:57Yeah, I think it's interesting. I mean, I think like some of these businesses are literally I think of as like cigar butt businesses where like literally probably they're going to disappear over some amount of time because what you can do with these SaaS tools is not different than what you can do with a general tool. And we've seen that in like 1.0, you know, in the AI era, actually, we saw it with like the writing tools and obviously writing was something that was so simple that you don't need a specific tool for it. I think that's probably true of some of these SaaS businesses too. I think, you know, it doesn't happen overnight.
1:08:33I think what happens is seat-based pricing probably shrinks. Software creation gets easier. There are probably more competitors. You can do a lot more with the general models. The interface gets disintermediated. And then I think one that is probably more true now than ever is like switching costs have gone away. It's just so easy to switch from one product to another because AI can do the switching for you. Agents can do the migration. So I think that is a challenge. and you look at a bunch of these companies like the ones that bending spoons has acquired and they have to figure out something and like what they're doing is dramatically taking out costs and increasing prices they're kind of like a private equity firm in that way and i think they're doing a good job with it i don't know that i would want to be customer of those products going forward um but you know i i think what they're doing makes sense at the same time some Some of these companies that are trading down so poorly actually are delivering great results and they continue to deliver great results.
1:09:39But I think part of this is, look, we had a big overhang of COVID era 0 % interest rates where we thought those multiples made sense and held up. And part of it is just like when that went away, those multiples went away, coupled with the AI disintermediation. So it's a bunch of things happening at once that has led to this. Guys, I know we're 70 minutes in, and so we should probably try to wrap up at some point. But the last sort of set of questions I'd love to get from you is, what is top of mind for all of you now? What are you thinking about, worried about, trying to learn more about when it comes to tech and startups today?
1:10:19And, you know, maybe while you answer that, like, where are you spending your time to learn about this kind of stuff? Is it are you call people and talk to them or read stuff online, read reports? I think that's what a lot of people who probably watch this want to know, like, where do we get our information from? So and what sort of questions are you asking? So maybe I'll go around Robin and you just spoke. So, Kanu, why don't we start with you not to put you on the spot? I mentioned I do early stage investing in normally fairly like frontier tech, deep tech. So one of the things that I try to keep up with everything that's advancing in AI.
1:10:56Right. So and it's tough, right? Like even my friends at OpenAid Anthropic tell me like it's hard to keep up with everything that's happening in AI. So, you know, what shot do I have? But I try. I do spend a lot of time with universities, with professors, with some of their staff students to understand, you know, what can be demonstrable in the next two, three years versus what might still be hype. And having a sense of that as I make investment decisions becomes very important. It helps me assess technical talent for my portfolio, helps me assess, you know, what's really novel, what's defensible.
1:11:29So I do spend a lot of time in universities. I also have been, you know, helping leading some of the robotics investments at KV. I do line up spending time with their target customers, right? Like some factories and manufacturing plants. I want to understand what is happening on those end sites. Like what is the challenge in scaling deployment? What needs to exist? How it can be solved? Where is AI helping adoption? And so, yeah, the physical world around us and, you know, what needed to get past the known challenges of adoption in robotics. I've been spending a lot of time on that. And one of the things I'm trying to get smarter about is, you know, new rulings by the FCC on, you know, what needs, what is required for it to be compliant for a robotic sale.
1:12:15You know, today it's like 65 % of the bomb needs to be sourced or licensed from U.S., U.S. friendly. Next year is going to be 75%. And really, what is the intent behind it and what can companies do and what are current companies doing? I'm trying to get smarter about that and talking to a ton of companies, but also people who are coming up with these laws. What is the intent? And what are we trying to achieve and what can be done differently? So, yeah, all of those are areas where I spend time trying to just have firsthand knowledge on what should be investable. You know, very similar. I mean, I think you've got to go to the source, right?
1:12:48I think it's very easy to be in the echo chamber. So that means the technologists at one level that are not just developing the technology but applying it, right? It means the labs and whatnot that are working on the frontier. It means the end customers that are trying to sort of use it in their work case and justify an ROI and the change management and all that type of stuff. And as Kanda said, also the regulators, which is a relatively new phenomenon. I think as you go after the labor layer, it looks a lot different than selling IT into a CTO. And it's going to be very political. right you see it in data center build outs but i think you're going to see it in spades and in on physical ai you know as you start to have humanoid robots people are going to feel it's hard to choke jack gpt but you can definitely throw a beer bottle at a at a robot or or a waymo so i i think that physical incarnation will will have a very visceral reaction for a lot of people and so sort of understanding consumer psychology you know how enterprise is adopting and i just try to do a listening tour of everyone all over the world who is trying to apply this stuff to understand and calibrate, particularly as the head of the firm, all of the people who are doing deep dives in individual domains, how it fits into a broader context and dynamic that you're going to do.
1:14:12Because it's not just technology that's going to impact it. It's the capital markets, it's financing because the capital intensity is so big now. Understanding that side of Wall street and the credit side and things that, you know, have not historically been something that really mattered in venture. But when you've got multi-billion dollar data center build outs for, you know, or chip buys, like it's, it's a, it's a different dynamic and you've got to understand at lots of altitudes and lots of horizontal, you know, so it's, it's a, it's, it's a multidimensional problem. That's only gotten more complex.
1:14:44Yeah. I think your first question was like what what's on my mind and it's really thinking about we're a seed fund and it's like how how does what does a seed fund mean in 2026 and when you know when seed rounds are getting done at 100 million dollar plus valuations at times um where do we invest are there still diamonds in the rough to be found? Or are we okay investing in companies where the outcome is a billion dollars or less than a billion dollars? And so it's like all things I'm thinking about as we think about the future of our firm. And it feels like up until now, it still feels like we've been able to find diamonds in the rough and come in at a reasonable price, own enough of the company that has done really well.
1:15:41And I wonder if that's even changed in the last couple of years. In terms of like, you also asked like, what do we read? How do we do research? I think we at times have had a thesis on a general area. So a few years ago, we had a thesis around accounting and did some homework, spoke to a bunch of execs in the accounting space, decided we want to invest there. We invested in three companies and they've all done quite well, would do more in that category. And so, you know, I think we try to just have a prepared mind going into anything and the focus area for us in fintech helps us do that. Makes sense.
1:16:24I'll just close out by saying my answer to the above and there's elements of what, you know, all of you said, and particularly I like trying to go to the source as much as possible for a lot of the, what I invest in that source happens to be playing with products and software and tools and actually seeing how these things function, where the gaps, where the holes, it helps me do a lot of diligence because man, I, we see so many claw decks and it's so easy to say spew a jargon, um, that a lot of investors sometimes don't understand. And so really understanding how the technology works for me has been very critical to be able to cut through some of that jargon and be like, okay, what are you actually building here?
1:17:05What is actually hard to do here? And so I do spend a lot of time there. And then of course, yes, universities and professors as well, probably less with the regulators than some of you, but yeah. But look, I mean, thank you guys for coming and I think we covered a lot of ground so I'm really happy with how this podcast went I don't think I can take J. Cal's job quite yet but hopefully he you did a great job man you were right there I think J. Cal will be proud of you try to channel my inner J. Cal and hopefully he watches this and says it's up to his bar. I'm sure he will. Thank you to our whole panel, Kanu Gulati of Cozla Ventures.
1:17:55Chris Farmer, he had to step out, but he's the founder and CEO of SignalFire. Of course, Shio Monat from Better Tomorrow Ventures and Didi Das, our incredible fill-in guest host from Menlo Ventures. Amazing job. Couldn't have done it better myself. Legitimately, I could not have. Just as a quick, before we dip out, we have lots of open launch roles here in Austin, in Tokyo, in Saudi Arabia, and now in Calgary, Alberta. If you want to find out more about those, go to careers.launch.co. Finally, we have a new Substack as part of our Substack family of newsletters. This Week in VC, it's going to feature VC roundtables just like the one that you just watched and other informative ones like this.
1:18:34Every week, plus fresh deals, posts, and more stuff. That's it, thisweekinvc.substack.com. Again, thanks to our panel. Thanks to our guest host, Didi Das. We will see you next time on This Week at Startups. Thank you.
This Week In Startups is made possible by:
Plaud: http://Plaud.ai/twist Odoo: http://Odoo.com/twist Northwest Registered Agent: http://www.northwestregisteredagent.com/twistdomain Quo: https://www.quo.com/perks/twist
Today’s show:
A personal AI agent that’s still invite-only just hit a $10 billion valuation. Guest host Deedy Das of Menlo Ventures is joined by Kanu Gulati of Khosla Ventures, Chris Farmer of SignalFire, and Sheel Mohnot of Better Tomorrow Ventures to debate whether Instinct can beat Meta’s Muse, what they actually use their own agents for, and what breaks when everyone’s agent makes the same move.
Plus we’re also talking about OpenAI’s math breakthroughs, the Factory vs. Cognition fight, and whether SaaS is still worth funding. Catch the whole debate on this week’s VC roundtable.
Guests:
Deedy Das on X: https://x.com/deedydas Menlo Ventures: https://menlovc.com
Kanu Gulati on X: https://x.com/KanuGulati Khosla Ventures: https://www.khoslaventures.com
Chris Farmer on X: https://x.com/chriswfarmer SignalFire: https://www.signalfire.com
Sheel Mohnot on X: https://x.com/pitdesi Better Tomorrow Ventures: https://www.btv.vc
Relevant Links:
OpenAI’s 722 math manuscripts from an unreleased internal model → https://github.com/openai/math
OpenAI’s earlier Navier–Stokes Millennium Prize result → https://openai.com/index/navier-stokes-solution/
Pramaana Labs, the Khosla-backed startup turning tax rules into provable logic → https://pramaanalabs.ai
Logos Research, the Imperial College spinout verifying AI-written financial code → https://www.imperial.ac.uk/news/articles/2026/logos-research-plans-to-make-ai-reasoning-trustworthy-in-high-stakes-environments/
EvenUp, SignalFire’s AI company for personal injury law → https://www.evenuplaw.com/
WindBorne Systems, the Khosla-backed weather balloon company Kanu mentioned → https://windbornesystems.com
Gokul Rajaram’s X Post: https://x.com/gokulr/status/2107851356019290569?s=20
micro1, the AI training data company Chris and LAUNCH backs → https://www.micro1.ai
Vercel’s data on companies shifting to open-weight models → https://vercel.com/i/open-weight-models
Instinct raises $1B at a $10B valuation (SiliconANGLE) → https://siliconangle.com/2026/09/28/everyday-personal-ai-assistant-startup-instinct-raises-1b-at-10b-valuation/
Meta introduces Muse, its personal AI agent → https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
Wajo, maker of the Fo agent Kanu uses → https://wajo.ai
Factory’s CEO accuses his board adviser of spying for Cognition (TechCrunch) → https://techcrunch.com/2026/09/30/factory-ceo-just-accused-his-vc-board-advisor-of-spying-for-cognition/
xAI rents its Colossus 1 data center to Anthropic (Data Center Dynamics) → https://www.datacenterdynamics.com/en/news/musk-spacex-xai-is-actively-seeking-more-ai-compute-customers-following-anthropic-deal/
Bending Spoons agrees to buy Miro, weeks after Airtable → https://investors.bendingspoons.com/newsroom/bending-spoons-agrees-to-acquire-miro
The FCC’s new rule on foreign-made robots → https://www.fcc.gov/covered-list-faqs-robots-inverters
Timestamps:
0:00 VC roundtable with guest host Deedy Das
1:23 If your work depends on conversations - interviews, meetings, calls - you need a Plaud NotePin S. You can check it out at https://Plaud.ai/twist and use code TWIST for 10% off!
2:11 OpenAI's math results and what's left to invest in
5:10 Formalizing taxes and code so AI can prove its work
6:37 Why science breakthroughs still need the real world
9:44 Odoo: The all-in-one business platform. Your first app is free! Get started today at https://Odoo.com/twist
10:54 Can startups compete with the labs on compute?
13:02 Private data and vertical AI as the moat
18:54 Would you invest in AI training data companies?
20:34 Got a new business idea? Northwest Registered Agent helps you bring it to life. Get a free domain, email, phone number, and more - with no purchase required! - Learn more at www.northwestregisteredagent.com/twistdomain
26:29 What a great founder looks like in the AI era
29:13 Quo (formerly OpenPhone) gives you a clean, modern way to handle every customer call, text, and thread all in one place. Try it free at https://quo.com/TWiST
34:35 Is Instinct worth $10 billion?
41:31 Can Muse reach a billion daily users?
44:49 Favorite personal agent use cases
49:06 What happens when everyone's agent makes the same call
53:17 Factory vs. Cognition and VCs backing rivals
1:03:49 Is SaaS still worth venture money?
1:09:49 Where the panel gets its information
1:14:45 What a seed fund means in 2026
Subscribe to our newsletters on Substack:
TWiST: https://twistartups.substack.com/
TWiAI: https://www.thisweekinai.ai/
TWiVC: https://thisweekinvc.substack.com/
Subscribe to This Week in Startups on Apple: https://rb.gy/v19fcp
Follow Lon:
X: https://x.com/lons
Follow Jason:
X: https://twitter.com/Jason
LinkedIn: https://www.linkedin.com/in/jasoncalacanis
Thank you to our partners:
Check out all our partner offers: https://partners.launch.co/