MoneyNeverSleeps

Christian Catalini argues AI verification is the bottleneck and value accrues to complementary assets, not labs

16 min Listen to the original Episode page
Artificial IntelligenceVenture CapitalLabor EconomicsOpen Source SoftwareRegulatory RiskComplementary AssetsTacit KnowledgeHistorical Technology

Summary

Christian Catalini argues that while AI excels at automating measurable tasks, verification of intent and context remains the scarce resource because models lack human implicit knowledge. He contends that fundraising is difficult to automate due to long feedback loops and that AI-generated pitch decks create a 'Sybil attack' on effort. Catalini rejects the notion that AI will reduce labor, citing historical general-purpose technologies that expanded economic activity. He predicts value will accrue to holders of complementary assets like chips and distribution rather than AI labs themselves. Finally, he outlines defensible strategies for founders, including securing ground truth and hiring top talent for tacit knowledge.

Key takeaways

  1. 1
    Training AI models on rejected deals (errors and corrections) is more valuable for learning than training on accepted deals, as it codifies the expert's judgment and override capabilities.

    Unidentified speaker (speaker_0) · 6:40

Episode sections

  1. 0:52 – 3:57

    Introduction and the Verification Bottleneck

    Pete Townsend introduces Christian Catalini, noting his background with LightSpark, MIT, and Meta's Libra. Catalini argues that while AI excels at automating measurable tasks, verification of intent and context remains the scarce resource because models lack the 'weights' of human expertise and implicit knowledge.

  2. 3:57 – 7:12

    Fundraising, Sybil Attacks, and Learning Loops

    Catalini explains that fundraising is hard to automate due to long feedback loops and confounded signals. He describes AI-generated pitch decks as a 'Sybil attack' on effort, making it hard to distinguish real work from polished fake output. He suggests recording rejected deals to train models, as errors teach more than successes.

  3. 7:12 – 8:29

    Labor Expansion and Historical GPT Patterns

    Catalini rejects the idea that AI will reduce labor, citing historical general-purpose technologies (GPTs) like railways and electricity that expanded economic activity. He predicts AI will allow investors and engineers to take on more complex projects rather than reducing their workload.

  4. 8:29 – 12:42

    Complementary Assets and Open Models

    Catalini argues that value in AI will accrue to holders of complementary assets (chips, data centers, distribution) rather than the AI labs themselves. He predicts companies like Microsoft and Meta will use open models to build niche specialists, preventing labs from absorbing the entire economy.

  5. 12:42 – 15:55

    Defensible Strategies for Founders

    Catalini outlines three defensible strategies for founders: securing 'ground truth' (friction with the real world like regulation or R&D), hiring top talent to provide tacit knowledge, and acquiring complementary assets like distribution and customer interaction data.

Claims worth verifying

Christian Catalini co-created Libra at Meta.

Unidentified speaker (UNKNOWN) ·0:26

How to verify: Meta's official announcements or press releases regarding the Libra/Diem project from 2026-2026; Christian Catalini's LinkedIn profile. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Amtropic is about to launch its IPO.

Pete Townsend ·0:46

How to verify: SEC EDGAR filings for Amtropic's S-1 registration statement or press releases from Amtropic around the release date of 2026-11-20. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Primary source

Meta has launched an AI assistant scalable to three billion users.

Pete Townsend ·0:46

How to verify: Meta's official product announcements or press releases regarding their AI assistant (e.g., Meta AI) and user metrics. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Christian Catalini is the co-founder of LightSpark.

Unidentified speaker (UNKNOWN) ·0:22

How to verify: LightSpark's official website, press releases, or Christian Catalini's LinkedIn profile. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Christian Catalini is the founder of the MIT Cryptoeconomics Lab.

Unidentified speaker (UNKNOWN) ·0:22

How to verify: MIT's official directory or the MIT Cryptoeconomics Lab's website. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the historical return data for inventors vs. complementary asset holders in the railway and electricity sectors using academic economic history sources.
  • Check Microsoft's and Meta's current AI integration strategies and open-source model adoption rates via their latest earnings calls and technical blogs.
  • Analyze the correlation between pitch deck polish and subsequent fundraising success in current VC databases like Crunchbase or PitchBook to assess the 'Sybil attack' claim.
  • Review regulatory filings and legislative proposals regarding AI utility regulation in the US, EU, and China to evaluate the 'regulatory moat risk' hypothesis.
  • Identify companies holding 'ground truth' data in highly regulated industries (e.g., healthcare, finance) and assess their moat durability against AI acceleration.
  • Examine the cost structure and margin trends of AI labs versus infrastructure providers (chips, data centers) to test the complementary assets value accrual thesis.
  • Look for case studies of failed AI startups that lacked tacit knowledge or real-world friction to validate the importance of human expertise in verification.
  • Compare the learning curves of models trained on successful vs. rejected deal data in existing AI research papers to assess the value of error-based training.
Full transcriptexpand

0:00Unidentified speaker (speaker_0): AI will take all the jobs. Or there's no point in launching a startup, right? The labs will do it all.

0:04Unidentified speaker (UNKNOWN): I

0:04Unidentified speaker (speaker_0): couldn't disagree more. I think every time we had a technology that really drastically expanded what we can do, we did more. We got more busy. So sorry, the four-hour work week not happening.

0:15Pete Townsend: This is Money Never Sleeps. Sharp rifts, big ideas, and real insights from smart people. I'm P Townsend, GP at Norio Ventures. Let's

0:22Unidentified speaker (UNKNOWN): go.

0:24Unidentified speaker (speaker_2): Money never sleeps, pal.

0:26Unidentified speaker (UNKNOWN): My

0:26Pete Townsend: guest today is Christian Catalini, co-founder of LightSpark and founder of the MIT Crypto Economics Lab. Christian co-created Libra at Meta. He put Bitcoin in the hands of every MIT undergrad back in 2014. And these days he's working on what happens to the economy when intelligence gets cheap and then checking the work stays expensive. Christian, welcome to Money Never Sleeps.

0:51Unidentified speaker (speaker_0): Great to be here.

0:52Pete Townsend: Listen, you've got this line that I keep coming back to. Anything that can be measured will be automated. If that's right, why does it then make verification of what's being measured the real bottleneck rather than intelligence itself?

1:10Unidentified speaker (UNKNOWN): Look,

1:10Unidentified speaker (speaker_0): over the last few months, I feel like people keep getting surprised by just how capable the models are becoming and just by how rapidly they're moving on in new domains. But if you zoom out, most of the action has been around code, mathematics, and any sort of domain where we do happen to have a lot of measurement. These machines are truly magical. Given enough data, they will figure out how to opt. optimize any gold that we give them within what they can see and what they can understand. But there's a lot of important context that they still don't have. Some of that context is strapping people's minds, you know, the weight that every expert in in their domain carries when they think about a problem, when they try to project the future and understand where where this may be going. There's a lot of excitement about word models these days. in the word of AI essentially models that can simulate lots and lots of counterfactuals. And we're all running one, right? So we're all a word model of one. So if you believe that the machines are capable of essentially solving any problem we give them, if they have the context, and you realize that some of these context is It's still elsewhere, then the concept of verification comes out of that. Why? Because verification is really simply making sure that when the machine

2:30has created something for you, whether it's a text, a document, a forecast, a piece of music, an image, a video, does that reflect your original intent? And sometimes we make the mistake of even not spelling out every dimension of the intent that we may have. I think we've all experienced it. You've read run a prompt, and then as the model is starting to think through, it's like, oh no, wait, I forgot to touch. tell you something. Right? There's just so much context that we give for granted. So context and intent require the human to be present and ensure that the machine is executing on something that we actually want. That's why verification is so important, even as models get increasingly capable.

3:08Unidentified speaker (UNKNOWN): I

3:08Pete Townsend: hear so many people talking about this human in the loop model, which is that let's just have a human verify. verifying what machines do. But realistically, it is a weak construct on how we'll actually use agents. And you've said, listen, it's the weights in your own mind. That's what separates a useful AI from, say, rogue agents. So talking about that real life example, that say I did take everything that I learned from my career of helping startups with fundraising. and I put that into an open weight model. Basically, this reference library of context for the AI agents to refer to.

3:50Unidentified speaker (UNKNOWN): Do

3:51Pete Townsend: you think that would be defensible? Or is fundraising one of those things that might escape measurement?

3:57Unidentified speaker (speaker_0): Fundraising is a really interesting one because when you think about fundraising from a verification perspective, uh there's a few issues. First of all, the decisions that you may make today will reveal themselves only often many years later. And so that feedback lag makes the direct causal link between okay, this type of founder, this type of idea, this market, and the conclusion way more murky. And most interesting problems in society don't really. lend themselves to the easy type of verification that you can do on a mathematical proof or on a piece of cone. Whenever you have a design where there's many potential paths and you cannot explore them all, the signal that you're getting back from the universe, whether it's an investment decision or a policy decision or a change in, you know, an educational approach, gets really confounded. That's where, you know, we still have one engine, which is the human brain that you know has has been selected over many, many um millennia to cope with with that kind of environment that still outshines the machines. I think there's this friendly term of a meat proxy, right? The colleague or friend that essentially sends you eye

5:04generated slob and you're the verifier of their work, right? And someone's like, here's my product, what do you vape? VIP Or as an investor, I'm sure you get these pitch tags. decks.

5:11Unidentified speaker (UNKNOWN): Oh yeah.

5:12Unidentified speaker (speaker_0): Essentially fully eye I generated and they impose a message massive cost on the investor to figure out, okay, is this actually the real deal or not? In the past you were able to separate form and effort really quickly, right? So a very polished DAC with a lot of substance, typically what was behind a good team. Today that signal is completely lost because generating output that looks like the real thing. But once you pause and think to, you know, the logic. the the substance often fails. It's gotten much harder. For those of you that are familiar with crypto, there's a massive Sybil attack on all these systems that society has invented for it's actually measuring effort. Did someone spend time and thinking on something or not?

5:54Pete Townsend: Yeah, it's crazy. Like I could take everything that I've learned, create this open weight model. I could use it not only to coach startups on fundraising, but I could also use it to due diligence startups and which side wins, right? So that when I pass on a deal that looks perfect on paper, is that the kind of expert override that you've been talking about that we should be recording and we should be looking for and feeding that back into the model? And how does that work?

6:23Unidentified speaker (speaker_0): That is essentially the key to the learning loop, right? I think. we will want more and more automation. Economic forces drive towards that, and it will happen. But for safe and reliable automation, you do need, at the beginning, still that expert in the loop for a number of decisions.

6:40Unidentified speaker (UNKNOWN): If

6:40Unidentified speaker (speaker_0): you codify those decisions, those reversals, the moment where you decide to decline a deal or go to the next stage, the errors and corrections are even more important, probably than the acceptances, right? So if you just train a model on your closed deals, you would not learn a lot about you as an investor. But if I see every decision, every moment where you decide to go forward or not, now I can start automating more of what you do and you can dedicate your attention to a narrower and narrower slice of the pipeline and potentially but that should go a lot deeper or do more deals. If you zoom out, I feel like right now, what one of the biggest fallacies is that many people believe

7:18Unidentified speaker (UNKNOWN): that

7:18Unidentified speaker (speaker_0): there's a fixed amount of labor, there's a fixed amount of projects that can be launched, there's a fixed amount of like ambition.

7:24Unidentified speaker (UNKNOWN): And

7:24Unidentified speaker (speaker_0): then they come with very negative conclusions like AI will take all the jobs, or there's no point in launching a startup, right? The labs will do it all. I couldn't disagree more. I think every time we had a technology that really drastically expanded what we can do, we did more. We got more. busy. So sorry, the four hour work week not happening, right?

7:42Pete Townsend: I know.

7:43Unidentified speaker (speaker_0): And on the work side, the way it would come is like, yes, a lot of what you're doing today as an investor, you won't do in in two years. You won't need to. The models will do that for you,

7:54Unidentified speaker (UNKNOWN): but

7:54Unidentified speaker (speaker_0): you will move up the value chain and expand kind of the ambition of that investment portfolio. Or if you're an engineer, you build way more complex fakes. That's That that's why I'm quite optimistic.

8:04Pete Townsend: So am I, you've written that in crypto that the value went to the boring intermediaries, as we call them, holding a banking charter, right? And that you're seeing that players with a regulatory moat are the ones that are out acquiring others right

8:21Unidentified speaker (UNKNOWN): now.

8:22Pete Townsend: Is that about to happen again with AI labs and what they're about to go through?

8:29Unidentified speaker (speaker_0): Not investment advice. We're actually recording this uh in a week where you know Amtropic is about to launch their IPOs. So very timely question. If you look at every general purpose technology, the original GPTs, right? Before the internet, railways, combustion engine, electricity, it always plays out the same way. And now I hear that people keep saying this time is different, intelligence is special. Well, look, imagine a world without electricity. It was a very different universe. And a lot of what we do every day, it's powered by electricity. So sure, intelligence is special, but so was electricity. Every time we had a GPT,

9:09Unidentified speaker (UNKNOWN): first

9:09Unidentified speaker (speaker_0): of all, the inventors appropriated almost none of it. Railways, massive investment. The return was 3.5% on average. Social return like 48%. Massive social returns to railways. And so my bet is that in AI, and this is where open models also play a role, most of the value will go to the holders of complementary assets. Think about who can produce the chips, who can assemble the massive data center. Players like Meta that have launched an AI assistant that they can scale to three billion users. They don't need frontier technology to make that useful to millions and millions of people at the billions. And so the lags are in this awkward position where if they're too successful,

9:48Unidentified speaker (UNKNOWN): if

9:49Unidentified speaker (speaker_0): the crazy fast takeoff scenarios materialize, well then they get nationalized and they become like electricity, right? So utility metered, overregulated, some something that becomes an input to production that the government will have heavy ads on. If they're also successful on the commercial side and start peeling off against all these other businesses, those businesses will fight back before it's too late. They're not gonna wait to be disrupted. It would be interesting to see what happens when Microsoft gets more challenged on the Office suite. Office is such a core product for Microsoft. Are they gonna just gonna go let that the business go? I don't think so.

10:23Unidentified speaker (UNKNOWN): And

10:23Unidentified speaker (speaker_0): so what all of these companies will do is they will take open source and open waste models, they will bring their own context to your point at the beginning, right? All your knowledge about investment, they will codify that context, and they're gonna start releasing frontier specialist models. They don't need generalists. They need to just deliver the best model for their industry. And so when that happens, I think the labs can be very successful companies, but they're definitely not gonna absorb all of the economy inside of them.

10:51Pete Townsend: The big opportunity is there for those who can make these open models really accessible and to make them these niche specialists that

11:02Unidentified speaker (UNKNOWN): you

11:02Pete Townsend: pair alongside a generalist, or perhaps you don't even use a generalist because you don't need to. You have one for a specific task.

11:09Unidentified speaker (speaker_0): What's interesting is that the labs are racing towards automating AI research. RSI is so recursive self-improvement, is all about getting AI that can write AI. Their hypothesis is if they can do that, they believe the economy is code, and so we will be able to absorb all of the economy.

11:28Unidentified speaker (UNKNOWN): But

11:28Unidentified speaker (speaker_0): this is the classic problem of Silicon Valley thinking. Jobs are messy, jobs are complex, industries are messy messing and complex. And so if all they end up doing is automating AI. research they're essentially gonna turn into a commodity the product that they're trying to sell and then who benefits is everyone else because inevitably even if you reach ASI even if you have a super intelligence unless you can contain it and put it in a black box which today it's been impossible right knowledge gets shared talent keeps moving there will be more than one lab prices will come down.

12:04Unidentified speaker (UNKNOWN): And

12:04Unidentified speaker (speaker_0): in that scenario, intelligence becomes this background ingredient, and everyone else starts building wonderful products and services on top of it, which I think is the best outcome for the economy and for society at large.

12:16Unidentified speaker (UNKNOWN): But

12:16Unidentified speaker (speaker_0): the irony is that because AI research might be verifiable itself,

12:22Unidentified speaker (UNKNOWN): it

12:22Unidentified speaker (speaker_0): might be one of these domains to fall in and that then everything that's hard to verify is where all of the value will accept. accrue.

12:27Pete Townsend: And on that note, that you know, you said the question you keep asking yourself, Christian, is what stays defensible? So what do you think you tell a founder now that is worth building that'll still be standing in five years with the pace of how everything is moving?

12:42Unidentified speaker (UNKNOWN): I

12:42Unidentified speaker (speaker_0): would say three things. The first one is ground truth.

12:46Unidentified speaker (UNKNOWN): If

12:46Unidentified speaker (speaker_0): you believe that everything that can be measured will be automated,

12:50Unidentified speaker (UNKNOWN): then

12:50Unidentified speaker (speaker_0): if you can't can find friction with the universe, whether it's on technology, RD, market, understanding customers, regulatory in many domains, right? Understanding your regulatory environment better than everybody else. Well, if you look at crypto, turned out that was a key asset, right? If you can find that ground truth, no matter what you're doing, and you're worried word class at that, you will be able to use intelligence on top of it and detail deter will just accelerate. you. So the lux come back with a better model. Guess what? You're still the best entity that can take that context, that ground fruit, and make it better and make a better product or service on the other end. So that's one ground fruit. The second one is top talent. We're not in the scenario where the the models will do it all, they're not super Superintelligence intelligent yet. And so there's still so much tacit knowledge in people's brains. If you have the top verifiers for the domain you care about, whether it's fintech or legal or cyber, you will build a better experience. And eventually maybe they'll have to acquire. But I do think that hiring the best talent globally in a domain is still a very defensible strategy. And third, I would say look for complementary assets. What are the things that combined combine with AI? It's the same with crypto, right? The complementary asset was the last mile. We build wonderful networks, very scalable, crypto primitives. And then in the end,

14:09for them to be useful in the real world, you needed identity, KYC, licenses, all this boring stuff that the crypto people thought could be abstracted away.

14:19Unidentified speaker (UNKNOWN): I

14:19Unidentified speaker (speaker_0): think the labs are making the same mistake. They think that they can just build intelligence and well, will just diffuse into society magically.

14:25Unidentified speaker (UNKNOWN): The

14:26Unidentified speaker (speaker_0): last mile. it's it's again where value will be created. Distribution.

14:30Unidentified speaker (UNKNOWN): I

14:30Unidentified speaker (speaker_0): think people think of distribution just narrowly as scale. But like if you think about it, distribution today in the world with AI is super viable because

14:39Unidentified speaker (UNKNOWN): if

14:39Unidentified speaker (speaker_0): I have distribution, I will learn a lot more about the universe way faster than you. That's why Muse by Meta, it's a faster moving product than a startup building the same AI agent. If you can get more interaction with customers, you know where all the things break, you build a better frontier model that serves shopping and other needs.

14:57Unidentified speaker (UNKNOWN): I'm

14:58Unidentified speaker (speaker_0): sure there's the same on the institutional side, right? So if I'm a top law firm and I can take an open weights model and I get friction every day in some of the most interesting legal cases, guess what? I'm gonna build a better AI for legal cases. So a very repeatable recipe on that.

15:12Pete Townsend: Tons of insights here in only a few minutes, Christian. What's the best place for people to go to follow your thinking?

15:20Unidentified speaker (UNKNOWN): Uh

15:20Unidentified speaker (speaker_0): usual X, C Catalini, also on Substack, but definitely the latest thoughts are always on X.

15:26Pete Townsend: Awesome. Very good. Well, listen, thank you so much for joining. Thank

15:30Unidentified speaker (speaker_0): you so much, Pete.

15:31Unidentified speaker (UNKNOWN): And

15:32Pete Townsend: to all of you out there, thanks for watching and listening. And don't forget to follow or subscribe wherever you get your podcast. If you're enjoying the show, leave us a review while you're there. It helps others to find the show, and it means a heck of a lot to me. Till next time. See ya.

This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights