[State of AI Startups] Memory/Learning, RL Envs & DBT-Fivetran — Sarah Catanzaro, Amplify

Okay. We're here with Sarah from Amplify. Welcome.

Thank you. First time being here. I took too long.

I know, I know. We've known each other for so long, and you never made an appearance.

You also made the transition from data to AI. I don't know if you were always as deep on AI, but obviously there's a lot of symbiosis between the two.

Yeah. I've always actually kind of oscillated between data and AI. Arguably, I started my career in so-called AI. It was more like symbolic systems back then, but as you said, they're so symbiotic that it's almost hard to divorce them.

That's actually what brought me into data. I wanted to better understand what happens when I write a SQL query.

Let's briefly touch on data, because that's a lot of where you and I first met. The dbt-Fivetran merger was so cool. How do you think about the end of the modern data stack?

A lot of people look at the dbt-Fivetran merger and talk about the end of the modern data stack, and I think that's a fundamentally wrong take. Both of these companies were growing very healthily.

You funded dbt?

We funded dbt. Both companies were beating their revenue targets. I think what you're seeing is an IPO environment in which companies are expected to have far more than $100 million in revenue.

What would you say the bar is now? $300 million?

No, above $600 million.

$600 million. Yeah. And the combined company is at $400 million.

I believe they'll actually be close to $600 million. I don't have the exact number, but they're clearly just getting ready for an IPO. Basically, the merger was a way to accelerate that path to liquidity, as you might remember.

They were the presumptive winners in their categories anyway.

Exactly. One of the things that has pleasantly surprised me—and this speaks again to the symbiotic relationship between data and AI—is that many of the big frontier labs are actually using both dbt and Fiverr.

I recall talking to folks at Thinking Machines Lab within weeks of the company's formation, and dbt was already an important part of their stack. Certainly, training datasets need to be managed. We need insight into what users are doing on these platforms, and the way in which you would analyze interactions with an agent or an LLM is even more complicated.

While I think the demand for analytics engineers and data scientists didn't explode in the way that some people thought—analytics engineers are not one-third of personnel—that doesn't mean the demand for those tools isn't still very prevalent.

Well, you got what you wanted. You wanted to democratize things. You got it.

Yeah. I guess we democratized things by perhaps reducing the need for people. I don't know whether or not that is a good thing, but I do think that the fact that it is easier than ever, from a tooling standpoint, for people to make data-driven decisions is probably a step in the right direction.

I've become convinced that while every company does need analytics engineers and data scientists, they probably don't need armies of them. Having a moderately sized data and analytics team is probably a good thing.

You touched on something interesting that I wasn't planning to ask about. I come from the data field, where data was synonymous with analytics.

You're now saying that dbt and Fivetran are being used for training data. Are there any notable differences in the workloads or the requirements?

Undoubtedly. One of the things we saw with analytics workloads that was surprising to some people in the data infrastructure space was that they were actually quite predictable. They were predictable because many of them were not being generated by humans, but rather by deterministic systems.

A lot of it was BI dashboards—Tableau hitting your database, or maybe not Tableau, but Looker, or Hex, or something like that. With analyzing, curating, and preparing datasets, it's a bit more ad hoc, so undoubtedly it will be less predictable.

I don't know if that really changes the way we approach developing data infrastructure. Some people are still quite interested in things like learned indexes and learned optimizers, and it's a bit easier to build a learned optimizer if you have more predictable workloads. It could change the way we approach things like that.

Data catalogs: do they become more important? Are they transferable?

Oh man, straight to the gut. That was something I got wrong.

I'm sorry. I don't know the background. What did you get wrong?

I really believed that data catalogs were going to become an important part of the modern data stack. I spent so much time working on data catalogs as a data scientist, and I felt like this was the thing I wanted. I didn't want to have to build the—

The main players are Atlan. I know they're Singaporean, so—

Yeah. There was data.world and Metaphor within our portfolio.

They've all struggled as a category.

They've all struggled a bit as a category. Many of them have subsequently been acquired, which suggests that this was perhaps not a standalone category.

With pretraining data, you have a lot more heterogeneous data all over the place. You need to keep on top of it, and you need to make it discoverable and accessible.

So why didn't it work?

I think there were a couple of things. We have seen some consolidation in the modern data stack, particularly around some of the key components, whether it was Fivetran, dbt, Hex, or Snowflake. Many of these products offered data-cataloging capabilities as a feature.

For humans, that was good enough. The data catalog available in Snowflake was good enough. The data-cataloging capabilities available in dbt were good enough.

With dbt, obviously, since they didn't build the cloud, they were going to build it.

Yeah, what else do you do?

My colleague Bar at Amplify was focused on these kinds of metadata services. I think it's still not obvious to me, but one opportunity that might have existed—or could have been realized—was building data catalogs not for humans, but for machines. That would look a little bit more like metadata services.

I don't just mean for agents, although I think that opportunity is arising more, but also for microservices and things like that. I do wonder at times if we built data catalogs for the wrong people and potentially even for the wrong use cases.

A lot of data-cataloging companies ended up focusing on discoverability, when perhaps the real market opportunity was in governance.

Governance is very important. Any other comments about what you know so far about the data stacks of the large labs? A lot of data people who are listening would probably want to sell into them.

A couple of observations. One is that they're paying careful attention to their data stacks. They're thinking about problems ranging from data discoverability to data preparation to the efficiency of data loading.

If you're unable to load data to a GPU efficiently, then the GPU is going to sit idle, and that's going to be a cost.

Exactly. What solutions are they using?

I get to talk about—yes, exactly—a portfolio company. We have a portfolio company called Spiral that has developed a file format called Vortex, and they make data loading super efficient, specifically to GPUs.

Okay.

Yeah. Good to know.

One of the things that has surprised me, though, is that so much data infrastructure has actually scaled quite elegantly to meet the AI use case.

You would hope.

You would, but the scale of these AI companies is incredible.

It's not as big as ads.

Maybe, maybe. I think that could change as agents become more prevalent and interface with each other, and perhaps the number of transactions explodes.

I have a friend who works on transactional databases at OpenAI, and I said, “You must be building databases. This is a paradigm shift in terms of the scale that databases are going to need to handle.”

He's like, “No, we use Rockset.” This one, right?

Yes.

Exactly.

Yeah. Very cool. Okay, let's talk about the funding environment, because obviously that's a big theme this year.

What comes to mind in terms of looking back at 2025? What stands out?

It was crazy.

You can give anonymized examples of what crazy looks like.

Yeah. I think crazy looks like raising upwards of $100 million in a seed round, where you have a long-term vision but not a near-term roadmap.

Seed?

Like, upwards of $100 million in a seed round.

Yeah.

This is something that I'm seeing happening not just occasionally but quite frequently.

Yes.

It definitely makes me anxious because, firstly, when founders are asking me how much they should raise, I'm typically saying,

Three to five?

Well, what do you need to do? What are your milestones for the next, let's call it, 12 to 24 months? What resources do you need in terms of headcount, compute, and equipment to unlock those milestones? Then maybe add a 20% buffer or something like that.

But doing that analysis requires you to understand what you're going to build in the next 0 to 24 months. I've talked to some companies and they're like, "We're building a frontier lab for X." I'm like, "Okay, cool. I get the long-term vision. There is an opportunity to make AI more secure, make AI more humane, make AI more data-efficient, whatever it might be."

So I'm bought into the long-term vision, and for me as an investor, that's super important. But let's talk about what your team's going to work on in the next 6 months. They're like, "Maybe we might build a consumer app."

I feel like I know exactly the company you're talking about.

I wish I was talking about one specific company. I'm actually talking about several companies. And look, I'd be a hypocrite to say that I've never done investments like that. But I've done investments like that when I really know the people and I'm like, "They're going to figure it out."

What's frightening about this funding environment is that you meet a founder who's like, "I'm raising $100 million." I'm raising like a billion dollars, maybe, at times. And you need to make a decision in 7 days, and I can't tell you what I'm going to do for the next 6 months. Conviction.

I think what some of the founders are missing is that you only have 7 days to get to know me. If you haven't figured it out, you probably want a partner who's going to be working closely with you to help you figure it out.

I mean, they're absolutely viewing it as transactional, right? They don't care.

No, they care about the most money at the highest valuation. The crazy thing is that they don't even seem to care about dilution. It's just the most money at the highest valuation.

Yeah. But it does send a signal that helps.

Yes, I think it does right now send a signal.

Okay, I'll tell you how it affects me, and I hate it. I hate it. Antithesis came out of stealth this week, right? The only thing I know about them is that they do something in AI testing, and James Street led a $100 million seed round.

We invested in it, too. I can tell you what they do. They do deterministic simulation. The thing that leads is the money.

Yeah.

And then, like, who else uses it other than Jean Street? What do you do that's innovative?

Palanteer.

Okay.

WarpStream.

Yeah. Anyway, maybe Antithesis is a bad example because they're actually legitimate. But there are a lot of similar examples where they just lead with the money, and there's not much substantiation behind it.

Maybe it's just bad storytelling, and that's why I, as a podcaster, get to talk to General Intuition. Once you spend some time with them, you're like, "Oh, okay, this is why they raised $100 million." But without that context, it's really hard to understand anything.

Well, I think there are some companies that are raising $100 million or more because they need it. A good example might be Periodic, in addition to—

Yeah, they need to build out a wet lab. Designing a wet lab that can support high-throughput biology, which is absolutely critical to their goals, is costly.

So I understand why they need that funding. But again, there are others where they don't have these near-term milestones.

I think the thing that is a little bit perturbing to me is that many of them are doing it because it makes it easier for them to hire. There are all of these candidates who want to work at a company that is a unicorn or a near-unicorn. They're pitching—

Because the alternative is working at a big lab, where the prestige and the money are there.

Yeah. Or the alternative is working at an early-stage startup. But there's something about the big valuation that becomes enticing.

They're also kind of pitching candidates. They have a compelling equity pitch where they're like, "Okay, maybe you're getting less than 0.1% of the company, but given the valuation, the value of your equity is already $10 million or something like that." They also guarantee a dollar value—

Of the equity.

Yeah. You mean that they'll offer them a loan to pay—

A buyback—

Yeah.

If you want to sell it.

Yeah. But—

Because they have so much cash.

But the thing, though, is that the valuation is a made-up number. Valuation, until a company exits, is an entirely made-up number. I could just say, "You know what? The Latent Space podcast is worth $5 billion," and we could agree. I, as an investor, could say that is the price, and now the company is worth $5 billion.

Yeah, it's not real. It's not actually traded in any volume.

Given the funding amounts that they're raising, if they spend that and then get acquired for less than that amount, their teams are getting nothing. I wish people were more sensitive to this dynamic and thought more about what the upside associated with the company is. More fundamentally, do I deeply believe in this vision? Joining companies because they have a billion-dollar valuation is just not the right way to choose a job.

I hear you. Okay, so obviously we can go on about that forever. There's also some stuff with cyclical funding and all that, but I do want to be more relevant to engineers and researchers.

What are the themes that are really strong? One thing I'll point out is that world models, just in general, are a really strong bet.

I would say that, every year, I go to this group of researchers and we take a vote on the top themes of the year. Everyone's extremely skeptical about world models. I think it's a trailing indicator because LLMs have been so enormously successful. You're like, "I don't need anything else." I don't know if you ever take on world models or any other top theme of the year.

My take on world models is that we have not yet defined what a world model is.

Oh, yeah. There are about 3 definitions right now.

Yeah, I think there's a lot of confusion about what a world model is and, therefore, what it should be used for. We're already seeing plenty of market potential for video models, including for things like perhaps Bayol's video editing.

I think we're already seeing some applications of world models to things like autonomous driving and potentially even coding. But again, it really hinges upon how you're defining world models.

I think one challenge that people have seen is that world models designed for one specific use case might not generalize to others. As an example, world models for video game generation might not generalize to factory settings or robotics. I use the word "might" strategically because I think it is potentially a research problem that might be figured out.

Yeah. That's part of the General Intuition podcast that we did. They had some evidence.

Yeah. I think it is possible. It's just that we're not there yet today.

Yeah.

A theme that I've been spending a lot of time thinking about is memory management and continual learning. I work with a lot of—

The same startup, I think.

Okay. I think I know what startup you're thinking about as well. But I actually see a lot of market potential for memory management and continual learning. My interest in this is more driven by conversations with practitioners.

Personalization is so important right now. I think what we're seeing is that a lot of AI application companies are growing really quickly, but they suffer from relatively low retention and relatively high churn.

So, you know, if you're developing an app like Cursor, how do you ensure that your users don't switch over to Windsurf, cloud code, Cognition, or whatever else when they release new features?

Yeah. Cursor Rules isn't enough, right? It's the shittiest form of memory.

Yeah. You know, and it's great, but I agree with that. I've publicly mused about this before: memory is very poorly implemented today in a lot of surfaces. Even ChatGPT—I wouldn't say people are particularly excited about it.

Okay, all right. You feel stronger about it than I do.

Yeah. Yeah. I wish ChatGPT had much better memory.

Yeah. Has this been the leading one?

I don't know. So then I think, in general, it makes product management harder, because what is the product? It's a combination of you plus memory. When you have a bug, is it the memory or is it something core? As a user, especially if it's consumer, there's going to be zero patience for any of this.

I agree. But that said, consumers seem to be tolerating products with no implementation of memory today. So I think early is still probably better than what exists now. Better is better than nothing, I guess.

Would you agree with the statement that, basically, let's say a key theme of 2026 is this personalization? I would call it kind of the consumerization of AI, in the same way that the consumerization of enterprise was a trend like 10 years ago.

Yeah. I mean, I think that is a good way to put it, too. For what it's worth, I don't think this is just a consumer or prosumer phenomenon. If you're in an enterprise that is adopting, again, Devon or Augment or something like that, you probably also want your models to learn.

Like, you start to track K-factor. I had to explain what that is to so many founders.

And, you know, if you're in normal SaaS, this is what you obsess over. To AI founders, they're like, “What do you mean? Growth just doesn't show up?”

Yeah. Yeah. I mean, it has, though. But I think it has because, for a while, AI has just felt magical. Now we're getting more accustomed to the magic, and it's no longer enough. I think we need to revert to some of the old tips and tricks for retaining people and bringing them in. Personalization is one of them.

I intermingle memory and continual learning, because I think one interesting element of personalization is not just learning facts about you or your preferences, but actually learning new skills from interactions with you and learning as the world changes. There are new versions of languages and frameworks, and other repositories are coming out all the time. The world is changing all the time. Human intelligence is incredibly dynamic, and yet artificial intelligence is just so static today.

So it must update weights for you?

But that also means that it's an interesting systems problem, because if you must update weights, then weights become stateful, and today inference is not stateful. I think there are going to be a lot of fun, gnarly problems to figure out as we figure out things like personalization and continual learning. That's also a fascinating infrastructure problem, because you have to load and unload and cache and all the good stuff.

Yeah. One more thing—I think we have time for one more take on RL environments.

Huge topic. Is it just a Docker container with some custom software loaded and logging stuff out? What are the good ones like, and what are the average ones like? I know I'm going on record on this, and I'm actually okay to be wrong, but I think RL environments are just a fad.

Oh, God. Oh, no.

They're all fake. I mean, the thing that makes me take it seriously is that the labs I know are paying 7 or 8 figures for RL environments, even though they could build them in-house. They're not, and I don't understand why. They were paying 7 or 8 figures for piss-poor data annotation, too.

Yeah.

And then, before that, data labeling. The labs have a lot of money. I think perhaps RL environments could create some value in the short term, but to the point about what makes a good RL environment and what makes a bad RL environment, I think the best RL environment is the real world. Why would I want to buy a DoorDash clone when I can just use logs and traces from DoorDash itself? It doesn't mean that we don't need to—

In parallel. Yeah. I mean, I think using the real world, using real apps as an RL environment is in fact the best thing. This is what Cursor does: they actually use real user activity on their platform to significantly improve both their coding agents as well as Tab. I think it's one of the approaches that has made the platform so compelling. You still need to figure out the right rubrics, and you still need to figure out the right set of tasks. So there are some aspects of RL environment design, at least as we're talking about it today, that I think are going to remain incredibly relevant, but just building a clone of an app, I think, is not that useful.

Yeah.

Yeah. Okay. That's all I'll take. We have maybe 3 minutes for any other stuff that you think about—just the state of startups in general, state of funding.

Yeah. So maybe I can talk about just the archetype startup that is most exciting to me.

Yes.

Yeah. I love investing in infrastructure tools, platforms, et cetera. As we talked about with continual learning, I think there will be opportunities for new tools, platforms, and infrastructure in the future. I've spent a lot of time thinking about applications today, and specifically the relationship between research and applications.

An example of this is that I think there were a lot of advances in RAG, and the biggest beneficiaries of these advances were the application companies for whom retrieval was a critical unlock. As an example of this, Harvey Haba—

I knew you were going to say Harvey.

Yeah, I mean, they have really interesting RAG implementations. They have hired really good researchers to advance the state of the art, and that enables them to build a better product. I feel this way very much about rule following and customer support. Rule following is a hard research problem, but if you solve rule following, then you unlock better customer support. I think a lot of Sierra's success can be attributed to their focus on this.

I've been thinking about, even for something like continual learning or memory, what the killer use case is where you can either offer a dramatically better experience by having a good memory implementation, or do something that just wasn't possible before. You can also think about this in the inverse, and often the best companies emerge in this way: “I'm trying to do this thing, but in order to actually do it, I need to solve this hard technical problem.” That's kind of the story of Runway. I don't think they would have built models if they didn't have to.

But I love that combination of delivering something that is better for consumers, better for users, while solving these really gnarly research and engineering problems.

Yeah. I don't want to—God, there's so much that I want to dig into there, but we're short on time. Thank you, just thank you in general. I don't know if you have a general call to startups, like a page somewhere that you can point people to.

Twitter, whatever it's called. Yeah, you can find me. You can find me there or in South Park. With the one-eyed dog. I'm easy to spot.

Okay.

Okay. Well, thank you so much for your time. I know you have to go, but I appreciate it.

Of course. It was great seeing you, and thanks for having me.

Yeah. Thanks.