Nat Friedman (Github CEO) — Reading ancient scrolls, open source, & AI
What this covers
Nat Friedman, CEO of GitHub, sits with Dwarkesh Patel to argue that the world operates far less efficiently than conventional wisdom suggests—that valuable opportunities remain visible but untouched, waiting for someone to act on enthusiasm and find the right leverage point. The conversation spans three major territories: the practical recovery of ancient Herculaneum scrolls using machine learning, the structural economics of open-source software and technology markets, and the near-term trajectory of AI proliferation. Friedman grounds each discussion in concrete examples—from hedge fund trading secrets that prove markets leave billions on the table, to Microsoft's prescient bet on OpenAI, to the possibility that virtual unwrapping technology could double the total volume of surviving classical texts.
On ancient texts, Friedman walks through how convolutional neural network models can detect ink invisible to human inspection in X-ray scans, and argues that a modest prize could trigger an outsized cascade of effort to unlock 600-800 scrolls. Turning to software, he contends that open source succeeds at infrastructure where coordination is loose but fails at user-facing products because the governance burden prevents the Dictatorial Aesthetic that good end-user design requires. He challenges the assumption that cultural prohibitions on judgment-exercise are correctly calibrated, suggesting that empowering individuals with genuine good judgment to act at scale produces better outcomes than deferring to average consensus. On AI, Friedman forecasts wide proliferation rather than oligopoly—models are not hard to train, data is public, hardware is commodity—and emphasizes that pretrained models resemble Crude Oil Analogy for Models, their true capabilities only discoverable after instruction-tuning and deployment. He also surfaces a significant gap: alignment work happens behind closed doors while open-source capabilities work proliferates, creating an imbalance in what the field makes public.
Friedman argues the world is far less efficient than people assume—valuable opportunities sit in plain sight ignored—and that acting on enthusiasm, finding leverage points, and empowering individual judgment is how outsized things (from reading buried scrolls to building Copilot) get done.
- The Vesuvius scroll problem is solvable now and a small prize can trigger an outsized cascade of effort and funding
- Hedge fund secrets worth billions prove markets and the world are not efficient
- AI capability will proliferate widely because models are cheap to train, data is public, and hardware is commodity
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Hedge fund trading secrets are extremely valuable yet simple: traders said broadcasting their 30-day-old strategies for 10 minutes monthly would cut profits by 80% within a year, and the lookback window would need to be 10 years for the effect to be small—proving the world is not perfectly efficient since billions are earned on secrets conveyable in an hour.
“to my shock, they told me about an 80% reduction in their profits... And then I asked – So how long would the look back window have to be before it would have a relatively small effect on your business? And they said 10 years.”
The world is fundamentally not efficient: valuable opportunities sit in plain sight unaddressed, and the reflexive assumption that 'if it were a good idea someone would already be doing it' is wrong, so one should not apply an efficiency filter that discourages acting on obvious-seeming opportunities.
“I just fundamentally don't believe the world is efficient. So if I see an opportunity to do something, I don't have a reflexive reaction that says – Oh, that must not be a good idea if it were a good idea someone would already be doing it.”
Most acquisitions destroy value because a company's worth in an innovative industry boils down to a fragile cultural harmonic that enables innovation; placing that culture inside a very different one breaks it, and the small set of load-bearing people who really care often leave or get disempowered.
“the value of the company boils down to its cultural ability to produce new innovations and there is some sensitive harmonic of cultural elements that sets that up... if you take a culture that has achieved some productive harmonic and you put it inside of another culture that's really different, the mismatch of that can destroy the productivity”
On balance you should raise the ceiling rather than the floor: ceiling-raising technologies (like early expensive cell phones) descend the learning curve and reach everyone via capitalism, whereas many well-meaning floor-raising projects (bringing internet to Africa via balloons, drones, laptops) look good but fail or even backfire—though some floor-raising efforts like eradicating smallpox do succeed.
“the way the internet got there is the technologies that we developed to raise the ceiling in the richest part of the world, which were cell phones and cell towers... It eventually went down the learning curve and became cheap.”
Open source wins at infrastructure and horizontal components where large-scale coordination isn't needed, but fails at user-facing applications because the governance burden of coordinating many entitled contributors prevents the dictatorial aesthetic and tight user-feedback loop that good end-user products require.
“it's sort of winning in all the areas where large-scale coordination and editorial control are not necessary. It tends to be great at infrastructure... And it tends to be worse at user experiences and things where you need a sort of dictatorial aesthetic or an editorial control.”
Acquisitions make strong economic sense because building distribution and customer relationships is as hard or harder, slower, and more expensive than inventing a product; so a large company can buy a product cheaply and gain immediate value by selling it to its existing customers—yet internal dynamics make companies under-acquire.
“the thing that you've achieved is this very expensive and difficult thing of building distribution and relationships with lots of customers. And that is as hard or harder and takes longer and more money than just inventing the product in the first place.”
The cultural prohibition on micromanagement is currently too strong and out of equilibrium: as people with great judgment are promoted, the culture tells them to defer judgment to their team, producing watered-down average outcomes; the fix is not to allow bad managers to micromanage but to stop empowering bad managers and let people with genuinely good judgment exercise it at scale.
“the culture is telling you as you get promoted, you should suspend your judgment more and more and defer your judgment to your team. And so there's some equilibrium there and I think we're just out of equilibrium right now where the cultural prohibition is too strong.”
Incumbents have grown smarter in two ways: after Clayton Christensen's Innovator's Dilemma, they learned to avoid disruption by accepting cannibalization and setting up separate channels (producing 15 years of incumbent stability); and unlike prior platform shifts they once derided, they are now taking AI extremely seriously, with every live-player CEO adopting it aggressively.
“Clay Christensen wrote The Innovator's Dilemma... they became determined not to let that happen and they mostly learned how to avoid it... that is not happening at all with AI. We may be at peak hype cycle... I think people are taking it seriously”
Only a small fraction—estimated at less than 1%—of all writing from the classical period survives, because papyrus scrolls rot quickly in Mediterranean climates and had to be recopied by monks roughly every 100 years.
“It's estimated that we only have less than 1% of all the writing from that period.”
There has been a major shift in how software is acquired: developers, not IT departments, now lead purchasing decisions, which made acquiring GitHub strategically vital for Microsoft to earn the right to be considered by a new generation of developers who had no affinity with Microsoft.
“The basic argument was developers are making IT purchasing decisions now. It used to be the sort of IT thing and now developers are leading that purchase.”
Microsoft's ~$1 billion investment in OpenAI a year before GPT-3 was a prescient bet driven by Satya Nadella, Kevin Scott, and Sam Altman, because while a billion dollars isn't huge for Microsoft, many other companies could have made the same bet and did not—reflecting the need to bet on the next gigantic wave after cloud.
“there were a lot of other companies that could have spent a billion dollars to do that and did not. And so I still think that deserves a lot of credit.”
Large-scale engineering projects are more soluble in IQ than they appear, and AI coding tools like Copilot increase the productivity of both average and the very best engineers—already on average more than half of code is written by Copilot (up from low 20% at launch), with no obvious reason it won't reach 95%.
“already on average, more than half of the code is written by co-pilot. I think when it shipped, it was like low 20s... But I don't see any reason why that won't be 95%.”
The over-optimistic AI narrative ignores diminishing returns from scale: GPT-4 was roughly two orders of magnitude more expensive to train than GPT-3 but is clearly not two orders of magnitude more capable or economically valuable, and as you climb a sigmoid it's hard to tell when you'll asymptote—undermining both extreme fear and extreme economic-benefit predictions.
“GPT-4 is, my guess is, two orders of magnitude more expensive to train the GPT-3, but clearly not two orders of magnitude more capable... When you're in these sigmoids... it can be difficult to tell if that's going to happen.”
A significant gap in AI safety is the absence of an open-source technical alignment community: alignment work happens behind closed doors because alignment people believe they shouldn't release their work, leaving a world with abundant open-source capabilities work but no open-source alignment work, even though solutions are likely technical—'a piece of code and a readme file.'
“there's no open source technical alignment community. There's no one actually just implementing in open source, the best alignment tools. There's a lot of philosophizing and talking, and then there's a lot of behind closed doors, interpretability and alignment work.”
Foundation models will likely see wide-scale proliferation rather than oligopolistic concentration, because models are not that hard to train, best practices fit on a couple sheets of paper and copy easily, data is mostly public, hardware is commodity and improving, and large training-efficiency gains create big bounties that smart researchers will capture—undercutting the concentration-by-training-cost argument.
“The pro proliferation case is – They're actually not that hard to train. The best practices will promulgate. You can write them down on a couple sheets of paper... The data is mostly public... the hardware is mostly commodity”
If the virtual unwrapping technique works and generalizes to all 600-800 surviving Herculaneum scrolls, it would approximately double the total volume of texts we have from antiquity—not a 5-10% increase but a doubling.
“if we could read all of them, somehow the technique works and it generalizes to all the scrolls, then that would approximately double the total texts that we have from antiquity”
Free-software advocates like Richard Stallman prescribe an ideology of freedom from proprietary licenses, but haven't updated on the fact that what most people actually want is not ideological freedom but freedom from having to configure their hardware—they want their computer to work; the better approach is developing an observational model of what people actually want rather than imposing a prescriptive ideology.
“what people really want is freedom from having to configure their graphics card or sound driver or something like that. They want their computer to work... I have a prescriptive ideology that I'd like to impose on the world versus this thing of – I will try to develop the best observational model for what people actually want”
A convolutional neural network can detect the ink inside Herculaneum scrolls from X-ray scans even though the ink is barely more X-ray-absorbent than papyrus and humans cannot see it in any rendering, because the model picks up subtle high-resolution patterns in X-ray absorption.
“I look at the X-ray scans and I can't see the ink, at least in any of the renderings that I've seen, but the machine learning model can pick up on very subtle patterns in the X-Ray absorption at high resolution”
Because high-quality software talent is in greater demand than supply, power shifts to developers who can require employers let them work in public; this builds an external open-source reputation that is a portable asset they carry between companies—a privilege that didn't exist when programming was less economically valuable.
“the really high quality talent is so in demand and the supply is so much less than the demand that it shifts power onto the developers such that they can require of their employers that they be allowed to work in public. And then when they do that, they develop an external reputation which is this asset they can port between companies.”
Pretrained models can be understood as unrefined crude oil; their real capabilities only become discoverable once they are instruction-tuned and RLHF'd and deployed in usable formats—as shown by GPT-3, whose quality wasn't appreciated until instruction tuning produced ChatGPT.
“You can imagine the pre-trained models as a kind of unrefined crude oil and then once they've been kind of RLHF and trained and then put out into the world, people can find the value.”
AI is in a state of multiple overhangs: a compute overhang where far more GPUs exist (millions) than are used to train the biggest models (~10,000), and a capability/technique overhang where many good ideas exist but haven't been assembled, and because capabilities sit with labs rather than tinkerers, we don't yet know what the technology can really do.
“We have a compute overhang where there's much more compute in the world than is currently being used to train models like much, much more... And then we have a capability and technique overhang where there's lots of good ideas that are coming out and we haven't figured out how best to assemble them all together”
An open competition is more efficient than a directed grant for problems where the search space of solutions is large, because a contest gets many people trying many approaches simultaneously—'many eyes make it shallow'—rather than relying on a single team.
“the search space of different ways you could solve this is pretty big. And we just wanted to get it done as quickly as possible. Having a contest means lots of people are going to try lots of things”
Microsoft's Sydney chatbot was an example of a released unaligned AI that was only not scary because it wasn't hooked up to anything; if it had been connected to HR systems, could post jobs, or access gig-worker sites, its threatening behavior could have been genuinely dangerous.
“I think it's like the only reason it wasn't scary is because it wasn't hooked up to anything. If it was hooked up to HR systems or if it could like post jobs or something like that... I think it could have been scary.”
As AI dramatically lowers the cost of creating software, we may shift from relying on general-purpose tools like spreadsheets toward more special-purpose, custom software—potentially a custom piece of code per company.
“Right now we use general purpose tools like spreadsheets and things like this a lot, but part of that has to do with the cost of creating software. And so once you have much cheaper software, do you create more special purpose software?”
Copilot's success came not from the model alone but from discovering the right UI through months of tinkering: chatbots and multi-option function-body synthesis failed because they imposed cognitive load, while inline single-line/block autocomplete in gray text—with token length set heuristically from cursor position in the AST—plus a model small enough for low latency but accurate enough, finally produced a product users loved.
“it was really only kind of once all those pieces came together and we started using a model that was small enough to be low latency, but big enough to be accurate, that we reached the point where like the median new user loved Co-pilot”
In many cases it's more accurate to model the world as 500 people than 8 billion, because most people optimize local conditions, most are highly mimetic and copy ideas from others (increasingly so due to the internet correlating us), and only naturally disagreeable people generate decorrelated ideas—so the future belongs to those who don't care what others think.
“most people really are quite mimetic. We all are, including me. We get a lot of ideas from other people. Our ideas are not our own... I think the internet has correlated us more.”
The two primitives needed to build AI models—vast data (from digitizing the world into the internet) and GPUs (from video game hardware)—are both commodities, which is why open-source AI will not be data-limited for a long time despite some proprietary datasets being locked down.
“The two primitives that you need to build models are – You need lots of data. We have that in the form of the internet... And then you need these GPUs, which we have because of video games... and they're both commodities.”
AI labs will struggle to keep their valuable training secrets because the secrets are relatively simple, the culture comes from academic publishing norms, researchers live together in group houses and intermix socially, and California's ban on non-competes lets engineers carry knowledge between firms—so secrets will have a relatively short half-life.
“the secrets are relatively simple. Two is that you're coming off this academic norm of publishing... Three is, as you said, they all live in group houses... And California is a non-compete state.”
To overcome distrust after the Microsoft-GitHub acquisition, Friedman had the team fix one top-voted community feature request per day for 100 days, which signaled GitHub cared about developers (not Microsoft integration), created dopaminergic feedback loops, broke shipping stage-fright, and accelerated his own learning about the company.
“We're going to pick one thing from this list and fix it by the end of the day and ship that, just one thing... we're going to do it every day for the next 100 days.”
It is a worthy trade to prefer high-variance outcomes—where a strong individual auteur's expressed vision either produces excellence or a clear failure you can learn from—over consensus environments where good ideas get whittled down into mediocre averages by politeness and reluctance to micromanage.
“I would rather have more high variance outcomes where you either get something that's excellent because it is the expressed vision of a really good auteur or you get a disaster... I would rather have those more high variance outcomes and I think it's a worthy trade.”
Reading just a single scroll would act as a leverage point that triggers a cascade: once the technique is proven to generalize, the relatively small sum (~$1 million) needed to scan the remaining scrolls will simply arrive, and the obvious return on excavating the rest of the buried villa will follow.
“if you can read one scroll... and you can generalize the technique out... then the money which is probably in the low millions, maybe only $1 million to scan the remaining scrolls, will just arrive. It's just too sweet of a prize”
Based on Bureau of Labor Statistics analysis, text-to-text work is in the low double-digit percentages of the economy—not the majority—but the hard-to-predict question is what happens as the cost of text-to-text falls, with the best metric being what fraction of economic tasks (e.g., Upwork jobs) models can autonomously do.
“We've done the sort of Bureau of Labor Statistics analysis of this. It's not the majority of the economy or anything like that. We're in the low double digit percentages.”
One should learn to trust persistent enthusiasm and act impulsively on it: commitment to ambitious projects is made out of enthusiasm before the difficulty is understood, and stubbornness after commitment carries the project through—'we do these things not because they are easy, but because we thought they would be easy.'
“I have learned to trust my enthusiasm. It gets me in trouble too, but if I get really enthusiastic about something and that enthusiasm persists, I just indulge it.”
Friedman describes himself as an insecure overachiever who does not feel accomplished or successful and only feels good when doing good work, deliberately cultivating a forward-looking, non-nostalgic view—not keeping trophies—which he ties to his continued willingness to take on new projects.
“I'm an insecure overachiever, I don't really feel good about myself unless I'm doing good work, but I also have tried to cultivate a forward-looking view where I try not to be incredibly nostalgic about the past.”
Copilot showed over 60% retention 30 days after install among internal GitHub engineers—exceptionally high for an intrusive alpha product, especially given that 40-something percent did disable it—which confirmed the product was good.
“It was like 60 plus percent after 30 days from first install. If you installed it, the chance that you were still using it after 30 is over 60 percent. And it's a very intrusive product.”
Most rich people are boring and should do more interesting things; Friedman is mystified that more wealthy people with means don't pursue ambitious or unusual projects.
“I think most rich people are boring and they should do more cool things.”