Nat Friedman (Github CEO) — Reading ancient scrolls, open source, & AI

We have 600-plus kind of roughly intact scrolls that we can't open. And I heard about this, and I thought that was incredibly exciting, like the idea that there was information from 2,000 years in the past. We don't know what's in these things. If we could read all of them, then that would give us approximately a doubling of the total text that we have from antiquity. If there are thousands more papyrus scrolls in there, and we now have the techniques to read them, then there's gold in that mud. And, you know, it's got to be dug out. I just fundamentally don't believe the world sufficient. And so if I see an opportunity to do something, I used to, but I no longer 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.

Okay. Today I have the pleasure of speaking with Nat Friedman, who was the CEO of GitHub from 2018 to 2021. Before that, he started and sold 2 companies, uh, companies, Zimian and Xamarin. And he is also the founder of AI Grant and California YIMBY. Most recently, he is the organizer and funder of the Squirrel Prize, which is where we'll start this conversation. Nat, do you want to tell the audience about what the Squirrel Prize is?

We're calling it the Vesuvius Challenge. This is just this crazy and exciting thing I feel incredibly honored to have gotten caught up in. A couple years ago, I was reading, it was the midst of COVID and I think we were in lockdown, And like everybody, I was sort of falling into internet rabbit holes, and I just started reading about the eruption of Mount Vesuvius in Italy about 2,000 years ago. And it turns out that when Vesuvius erupted, it was AD 79, it destroyed all the nearby towns. Everyone knows about Pompeii, but there was another nearby town called Herculaneum, and Herculaneum was sort of like the Beverly Hills to Pompeii, so big villas, big houses, fancy people. And in Herculaneum, there was one villa in particular that was enormous, and it had once been owned by the father-in-law of Julius Caesar, and so, well-connected guy. And it was full of beautiful statues and marbles and art, but it was also the home to a huge library of papyrus scrolls. And so when the villa was buried, the volcano actually spit out enormous quantities of mud and ash, and it buried Herculaneum in particular in something like 20 meters of material. So it wasn't like a thin layer, it was a very thick layer. Those towns were buried and forgotten for hundreds of years. No one even knew exactly where they were until the 1700s. And so in 1750, a farm worker who was digging a well kind of in the outskirts of Herculaneum struck this marble paving stone of a path that had been at this huge villa. And of course, he was pretty far down when he did that. He was 60 feet down. And then subsequently, this Swiss engineer came in and started digging tunnels from that well shaft, and they found all these treasures. And that was sort of the spirit at the time, was looting. They were taking out incredible— if they encountered wall, they would just bust through it. They were taking out these beautiful bronze statues that had survived, and along the way they kept encountering these lumps of what looked like charcoal. They weren't sure what they were, and many were apparently thrown away until someone noticed a little bit of writing on one of them. They realized they were papyrus scrolls, and there were hundreds, there may have been thousands of them. So they had uncovered really the this enormous library was the only library ever to have survived in any form, even though it was badly damaged. They were carbonized, very fragile, deformed. The only one that survived since antiquity. In the open air, these papyrus scrolls in a Mediterranean climate, they rot and they decay quickly. And so they'd have to be recopied by monks every 100 years or so, maybe even less. And so we only have, it's estimated, something less than 1% of all the writing from that period. And so to find underground hundreds of definitely not in good condition but still present papyrus scrolls where on a few of them you can make out the lettering was like this enormous discovery. People immediately in a well-meaning attempt to read them started trying to open them, but they're really fragile. They turn to ash in your hand. And so hundreds were destroyed. People did things like cut them with daggers down the middle and a bunch of little pieces would flake off and they'd try to get a few letters off of a couple of pieces. And then eventually there was a monk named Piaggio, who's an Italian monk, and he devised this machine under the care of the Vatican to unroll these things very, very slowly, like half a centimeter a day, something like that. And a typical scroll I think could be 15 or 20 or 30 feet long and managed to successfully unroll a few of these. They found on them Greek philosophical texts in the Epicurean tradition by this little-known philosopher named Philodemus. But we got kind of new text from antiquity, which is not a thing that happens all the time. Eventually people stopped trying to physically unroll these things because so many were destroyed. In fact, Some attempts to physically unroll the scrolls continued even into the 2000s, like 1990s, 2000s, and they were destroyed. So the current situation is we have 600+ roughly intact scrolls that we can't open. I heard about this and I thought that was incredibly exciting, the idea that there was information from 2,000 years in the past. We don't know what's in these things. Obviously, people are trying to develop new ways and new technologies to open them. I read about a professor at the University of Kentucky, Brent Seales, who had been trying to scan these using increasingly advanced imaging techniques and then use computer vision techniques and machine learning to virtually unroll them without ever opening them. They tried a lot of different things, but their most recent attempt in 2019 was to take the scrolls to a particle accelerator in Oxford, England called the Diamond Light Source and to make essentially an incredibly high-resolution CT scan of them, sort of 3D X-ray scan. They needed really high-energy photons in order to do this and they were able to take scans at 8 microns, so these really quite tiny voxels which they thought would be sufficient. I thought this was the coolest thing ever. We're using technology to read this lost information from the past. I waited for the news that they had been decoded successfully. That was 2020. Then I think COVID hit, everybody got a little bit slowed down by that. Last year I found myself wondering, I wonder what happened to Dr. Seales and his scroll project? I reached out and it turned out they had been making really good progress. They'd gotten some machine learning models to start to identify ink inside of the scrolls, but they hadn't yet extracted words or passages. It's very challenging. I invited him to come out to California and hang out, and to my shock, he did. We got to talking and decided to team up and try to crack this thing. And the approach that we've settled on to do that is to actually launch an open competition. We've done a ton of work with his team to get the data into a shape where— and the tools and techniques and just the broad understanding of the materials into a shape where smart people can kind of approach it and get productive easily. And then I'm putting up together with Daniel Gross, a prize, sort of like an XPRIZE or something like that, for the first person or team who can actually read substantial amounts of real text from one of these scrolls without opening them. And so we're launching that this week. I guess maybe it's when this airs, I don't know. The stakes are kind of big. What gets me excited are the stakes. So the 600 or 800 scrolls that are there, it's estimated that if we could read all of them, Somehow the technique works and it generalizes to all the scrolls, then that would give us approximately a doubling of the total text that we have from antiquity. This is what historians and classicists tell me. So it's not like, oh, we would get a 5% bump or a 10% bump in the total ancient Roman or Greek text. It would be like, no, we get all of the text that we have again. Multiple Shakespeares is sort of one of the units That I've heard. So that would be significant. I mean, we don't know what's in there. We've got a few Philodemus texts. Those are of some interest, but there could be lost epic poems or God knows what. So I'm really excited and I think my bet is there's like a 50% chance that someone will encounter this opportunity and get the data and get nerd sniped by it and we'll solve it this year.

I mean, really, it is something out of a science fiction novel. It's like something you'd read in Neal Stephenson or something. I was talking to Professor Seales before and apparently the shock went both ways because the first few emails he was like, this has got to be spam. No way Nat Friedman is reaching out and has found out about this prize.

That's really funny because he was really pretty hard to get in touch with. I emailed him a couple times, he just didn't respond. And so I asked my admin Emily to call the secretary of his department and say, Mr. Friedman requested me. And then he knew there was something actually going on there. And so he finally got on the phone with me and we got on Zoom and he was like, why are you interested in this? I mean, I love Brent. He's fantastic. And I think we're friends now and I think we found that we think alike about this and I think he's reached the point where he just really wants to crack the— they've taken this right up to the 1-yard line. This is doable at this point. They've demonstrated, I think, every key component, but putting it all together, improving the quality, doing it at the scale of a whole scroll, this is still very hard work and an open competition seems like the most efficient way to get it done.

Before we get into the state of the data and the different possible solutions, I want to make tangible what could be gained if we can unwrap these. You said there's a few more thousand scrolls. Are we talking about the ones in Philodemus's lair or are we talking about the ones in other lairs?

You'd think if you find this crazy villa that was owned by Julius Caesar's father-in-law that we'd just dig the whole thing out. But in fact, most of the exploration occurred in the 1700s through the Swiss engineers' tunnels underground. The villa was never dug out and exposed to the air. You went down 50, 60 feet and then you dig tunnels. Again, they were looking for treasure, not a full archaeological exploration, so they mostly got treasure. In the '90s, some additional excavations were done at the edge of the villa and they discovered a couple things. First, they discovered this seaside villa that faced the ocean. It was right on the water before the volcano erupted. The eruption actually pushed the shoreline out by depositing so much additional mud there, so it's no longer right by the ocean apparently. I've actually never been. And they also found that there were 2 additional floors in the villa that the tunnels apparently had never excavated. And so at most a third of the villa has been excavated. Now they also know when they were discovering these papyrus scrolls that they found basically one little room where most of the scrolls were. And these were mostly these Philodemus texts, at least that's what we know. And they found apparently several revisions sometimes of the same text. And so they think the hypothesis is this was actually Philodemus's working library. He worked here, this sort of Epicurean philosopher. And in the hallways though, they occasionally found other scrolls, including crates of them. And the belief is, at least this is what historians have told me, and I'm no expert, But what they have told me is they think that the main library in this villa has probably not been excavated, and that the main library may be a Latin library and may contain literary texts, historical texts, other things, and that it could be much larger. Now, I don't know how prone these classicists are to wishful thinking. It is a romantic idea, but they have some evidence in the presence of these partly evacuated scrolls that were found in hallways. and that sort of thing. So there are descriptions. I've since gone and read a bunch of the firsthand accounts of the excavations and there are these heartbreaking descriptions of them finding an entire case of scrolls in Latin and accidentally destroying it as they tried to get it out of the mud. There were maybe 30 scrolls or something in there. So there clearly was some other stuff that we just haven't got to.

You made some Yeah, some scrolls.

We do this. Okay, so papyrus is a papyrus and it's a reed. It's a grassy reed that grows on the Nile in Egypt. And for thousands of years, many thousands of years, they've been making paper out of it. And the way they do it is they take the kind of outer rind off of the papyrus and then they cut the inner core into these strips and they lay the strips out kind of parallel to one another and then they put another layer at 90 degrees to that bottom layer and they press it together in a press or under stones and let it dry out. And that's papyrus essentially. And then they'll take some of those sheets and kind of glue them together with paste made out of flour usually and get a long scroll. And you can still buy it. I bought this on Amazon and it's interesting because It's got a lot of texture, those fibrous ridges of the papyrus plant. When you write on it, you really feel the texture. I got it because I wanted to understand what are these artifacts that we're working with. So we made an attempt to simulate carbonizing a few of these. We basically took a Dutch oven because when you carbonize something and you make charcoal, it's not like burning it with oxygen. You remove the oxygen, heat it up, and let it carbonize. So we tried to simulate that with a Dutch oven, which is probably imperfect, and left it in the oven at 500 degrees Fahrenheit for maybe, I don't know how long these were, but our biggest attempt was like 5 or 6 hours. These things are incredibly light and if you try to unfold them, they just fall apart in your hand very readily. I assume these are in somewhat better shape than the ones that were found because these were not in a volcanic eruption and covered in mud. I think the volcano was probably— maybe that mud was hotter than my oven can go. And it just flakes. You just squeeze it, it's just dust in your hand.

Yeah.

And so we actually tried to replicate many of the heartbreaking 1700s, 18th century unrolling techniques. Like they used rosewater, for example, or they tried to use different oils to soften it and unroll it. And most of them are just very destructive. They poured mercury into it because they thought mercury would slip between the layers potentially. So yeah, this is sort of what they look like. They shrink and they turn to ash.

For those listening, by the way, it kind of looks like a black— I mean, just imagine sort of the ash of a cigar but blacker, and it crumbles the same way. It's just a blistered black piece of rolled-up papyrus.

Yeah, and they blister. The layers can separate, they can fuse.

And so this happened in 79 AD, right? So we know that anything before that could be in here. Which I guess could include—

Yeah, so what could be in there? I don't know. You and I have speculated about this. Well, I think it would be extremely exciting not to just get more Epicurean philosophy, although that's fine too, but almost anything would be interesting and additive. Dreams are— I think it would maybe have a big impact to find something about early Christianity, like a contemporaneous mention of early Christianity. Maybe there'd be something that the church wouldn't want. That would be exciting to me. Maybe there'd be something, some color or detail from someone commenting on Christianity or Jesus. I think that would be a very big deal. We have no such things as far as I know. Other things that would be cool would be old stuff, like even older stuff. So there were several scrolls already found in there that they know were hundreds of years old when the villa was buried. So the villa was probably constructed about 100 years prior, is my understanding. And they can tell from the style of writing, they can date some of these scrolls. And so there was some old stuff in there. And the Library of Alexandria was burned 80 or 90 years prior. And so again, maybe wishful thinking, but there's some rumors that some of those scrolls were evacuated And maybe some of them would've ended up at this substantial, prominent Mediterranean villa. God knows what would be in there. That would be really cool. I think it'd be great to find literature. Personally, I think that would be exciting. Beautiful new poems or stories. We just don't have a ton because so little survived. So I think that would be fun. I think you had the best crazy idea for what could be in there. Which was text which was GPT watermarked. That would be a creepy feeling.

I still can't get over just how much of a plot of a sci-fi novel this is like. Potentially the biggest intact library from the ancient world that has been sort of stopped like a debugger because of this volcano. And I mean, just the philosophers of antiquity forgotten, The earliest gospels. There's so much interesting stuff there. But let's talk about what the data looks like. You mentioned that they've been CT scanned and that they've developed these machine learning techniques to do segmentation and the unrolling. What would it take to get from there to understand the actual content of what is within?

Dr. Seales actually pioneered this field of what he calls, and is now widely called, virtual unwrapping. And he did it actually not with these Herculaneum scrolls. These things are like expert mode. They're so difficult. I'll tell you why soon. But he did it initially with a scroll that was found in the Dead Sea in Israel. It's called the En-Gedi Scroll, and it was carbonized actually under slightly similar circumstances. I think there was a temple that was burned. The papyrus scroll was in a box, so it's like a Dutch oven. It kind of carbonized in the same way. And so it was not openable. It would just fall apart. So the question was, could you non-destructively read the contents of it? He did this 3D X-ray, the CT scan of the scroll, and then was able to do 2 things. First, the ink gave a great X-ray signature, and so it looked very different from the papyrus. There was a high contrast. Second, he was able to segment the winds of the scroll throughout the entire body of the scroll and identify each layer and then just geometrically unroll it using fairly normal flattening computer vision sort of techniques and then read the contents of it. And it turned out to be, I think, an early part of the Book of Leviticus, something of the Old Testament or the Torah. And that was like a landmark achievement. And so then the next idea was to apply those same techniques to this case. And so, okay, this has proven hard. I think there's a couple things that make it difficult. One is that— the primary one is that the ink used on the Herculaneum papyri, it is not very absorbent of X-ray. It basically seems to be equally absorbent of X-ray as the papyrus, or very close, certainly not perfectly. And so you don't have this nice bright lettering that shows up on your tomographic 3D X-ray. So you have to somehow develop new techniques for finding the ink in there. That's sort of problem one and it's been a major challenge. And then the second problem is the scrolls are just real messed up. They were long and tightly wound, highly distorted by the volcanic mud, which not only heated them but partly deformed them. And so just the segmentation problem of identifying each of these layers throughout the scroll is doable, but it's hard. Those are a couple of challenges. And then the other challenge, of course, is just getting access to scrolls and taking them to a particle accelerator. So you have to have scroll access and particle accelerator access and time on those. It's expensive and difficult. Dr. Seales did the hard work of making all that happen. The good news is, very recently, just in the last couple of months, his lab has demonstrated with a convolutional neural network the ability to actually recognize ink inside these X-rays. I look at the X-ray scans and I cannot, at least in any of the renderings that we've seen, I can't see the ink, but the machine learning model can pick up on sort of very subtle patterns in the X-ray absorption at high resolution inside these volumes in order to identify ink. And we've seen that. So you might ask, okay, how do you train a model to do that? Because you need some kind of ground truth data to train the model. The big insight that they had was to train on broken-off fragments of the papyrus. As people tried to open these over the years in Italy, they destroyed many of them, but they saved some of the pieces that broke off. And on some of those pieces, you can kind of see lettering. And if you take an infrared image of the fragment, then you can really see the lettering pretty well in some cases. And so they think it's 930 nanometers. They take this little infrared image. Now you've got some ground truth. Then you do a CT scan of that broken-off fragment and you try to align it, register it with the image, and then you have data that you can use potentially to train a model. And that turned out to work in the case of the fragments. Okay, so now I think this is sort of the why now. This is why I think launching this challenge now is the right time because we have a lot of reasons to believe it can work. The core techniques, the core pieces have been demonstrated. It just all has to be put together at the scale of these really complicated scrolls. So yeah, I think if you can do the segmentation, which is probably a lot of work, maybe there's some way to automate it, and then you can figure out how to apply these models inside the body of a scroll and not just to these fragments, then it seems like you could probably read lots of text.

Why did you decide to do it in the form of a prize rather than just giving a grant to the team that was already pursuing it or maybe some other team that wants to take it up?

We talked about that. But I think what we basically concluded was 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.

Yeah.

So having a contest means lots of people are going to try lots of things and someone's going to figure it out quickly. Many eyes may make it shallow as a task. And so I think that's the main thing. Probably someone could do it, but I think this will just be a lot more efficient and it's fun too. I think this is fun. I think it's interesting to do a contest and who knows who will solve it or how. People may not even use machine learning. We think that's the most likely approach for recognizing the ink, but they may find some other approach that we haven't thought of.

One question people might have is that you have these visible fragments mapped out. Do we expect them to correspond to the burned-off or the ashen carbonized scrolls that you can do machine learning on. The ground truth of one can correspond to the other.

I think this is a very legitimate concern. They're different. When you have a broken-off fragment, there's air above the ink. So when you CT scan it, you have ink next to air. Inside of a wrapped scroll, the ink might be next to papyrus because it's pushing up against the next layer and your model may not know what to do with that. I think this is one of the challenges and sort of how you take these models that were trained on fragments and translate them to this slightly different environment. But maybe there's parts of the scroll where there is air on the inside and we know that to be true. You can sort of see that here. I think it should at least partly work and clever people can probably figure out how to make it completely work.

You said the odds are about 50/50. What makes you think that it can be done?

I think it can be done because we've recognized ink from a CT scan on the fragments. I think everything else is probably geometry and computer vision. The scans are very high resolution. They're 8 microns, 8 micrometers. If you stood a scroll on end like this, they're taken in these slices through it like this. It's like this in the Z-axis from bottom to top, there are these slices. And the way they're represented on disk is each slice is a TIFF file. And for the full scrolls, each slice is like 100-something megabytes. So they're quite high resolution. And then if you stack, for example, 100 of these, they're 8 microns, right? So 100 of these is 0.8 millimeters. So a millimeter is pretty small. So they're fairly— We think the resolution is good enough, or at least right on the edge of good enough. that it should be possible. There seem to be 6 or 8 pixels or voxels, I guess, across an entire layer of papyrus. That's probably enough. And we've also seen with the machine learning models, Dr. Seales has got some PhD students who have actually demonstrated this at 8 microns. So I think that the ink recognition will work. I think the data is clearly physically in the scrolls. The ink was carbonized, the papyrus was carbonized, A lot of data actually physically survived. Then the question is, did the data make it into the scans? I think that's very likely based on the results that we've seen so far. I think it's just about a smart person solving this or a smart group of people or just a dogged group of people who do a lot of manual work. That could also be true. You may have to be smart and dogged. I think that's where most of my uncertainty is, just whether somebody does it.

If a quarter of a million dollars doesn't motivate you.

Yeah, I think money's good. There's a lot of money in machine learning these days.

That's true as well. Do we have enough data in the form of scrolls that have been mapped out to be able to train a model if that's the best way to go? I guess because one question somebody might have is, listen, if we already have this ground truth, Why hasn't Dr. Seale's team already been able to just train a model?

I think they will. I think if we just let them do it, they'll get it solved. It might take a little bit longer because it's not a huge number of people and there is a big search space here. But if we didn't launch this contest, I'd still think this would get solved, but it might take several years. I think this way it's likely to happen this year.

Okay. And then what happens, let's say the prize is solved, somebody figures out how to do this and we can read the first scroll. You mentioned that these other layers haven't been excavated. How is the world going to react? Let's say we get one of these mapped out. How's the world going to react?

That's my personal hope for this. I always like to look for these cheap leverage hacks, these moments where you can do a relatively small thing and it creates— you kick a pebble and you get an avalanche. The theory is, and Brent shares this theory, the theory is that if you can read one scroll, just one scroll, and we only have 2 scanned scrolls, there's hundreds of surviving scrolls. It's relatively expensive to book a particle accelerator. So if you can scan one scroll and you know it works and you can generalize the technique out and it's going to work on these other scrolls, then the money, which is probably low millions, maybe only $1 million to scan the remaining scrolls, will just arrive. It's too sweet of a prize for that not to happen. And the urgency and return on excavating the rest of the villa will be incredibly obvious too. Because if there are thousands more papyrus scrolls in there and we now have the techniques to read them, then there's gold in that mud and it's got to be dug out. It's amazing how little money there is for archaeology. Literally for decades, no one's been digging there. So that's my hope is that this is the catalyst that it works. Somebody reads it, they get a lot of glory, we all get to feel great, and then the diggers arrive in Herculaneum and they dig out the rest.

I wonder if the budget for archaeological movies and games like Uncharted or Indiana Jones is bigger than the actual budget to do real-world archaeology. But I was talking to some of the people before this interview and that's one thing they emphasized is your ability to find these leverage points. For example, with California YIMBY, I don't know the exact amount you seeded it with, but for that amount of money and for an institution that is that new, it is one of the very few institutions that has had a significant amount of political influence. If you look at the state of YIMBY in California and nationally today, I guess, how do you identify these things? There's plenty of people who have money who get into history or get into whatever subject. Very few do something about it. How do you figure out where to—

I'm a little bit mystified by why people don't do more things too. I think first of all, I don't know, maybe you can tell me, why aren't more people doing things? I think most rich people are boring and they should do more cool things. So I'm hoping that they do that now. I think part of it is I just fundamentally don't believe the world is efficient. And so If I see an opportunity to do something, I no longer 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. Someone must be taking care of housing policy in California, right? Or somebody must be taking care of this or that. So I think first, I don't have that filter that says the world's efficient, don't bother, someone's probably got it covered. Then the second thing is I kind of have learned to trust my enthusiasm. This gets me in trouble too, but if I get really enthusiastic about something and that enthusiasm kind of persists, I just indulge it and just think, oh yeah, I'm gonna go. I like doing the things I'm enthusiastic about. And so I just kind of let myself be impulsive. Frequently what you do— there's this great image that I found and tweeted which said, we do these things not because they are easy, but because we thought they would be easy. That's frequently what happens is the commitment to do it is impulsive and it's done out of enthusiasm and then you get into it and you're like, oh my God, this is really much harder than we expected. But then you're sort of committed and you're stuck and you're going to have to get it done. I thought this project would be relatively straightforward. We're just going to take the data and put it up. But of course everything is— and truly 99% of the work has already been done by Dr. Seales and his team at the University of Kentucky. I am a kind of carpetbagger. I've shown up at the end here to try to do a new piece of it.

The last mile is often the hardest.

Well, I mean, it turned out to be fractal anyway. Just all the little bits that you have to get right to do a thing and have it work. I hope we got all of them. So I think that's part of it is just not believing the world's efficient, then just allowing your enthusiasm to cause you to commit to something that turns out to be a lot of work and really hard and then you just are stubborn and don't want to fail and so you keep at it. I don't know. I think that's it.

Yeah, I don't know. I feel like the efficiency point, do you think that's particularly true just of things like California maybe or this where there isn't a direct monetary incentive?

No. I mean, look, certainly parts of the world are more efficient than others and you can't assume equal levels of inefficiency everywhere, but I'm constantly surprised by how even in areas you expect to be very efficient, there are things that are sort of in plain sight. It's not that I see them and others don't. There's lots of stuff I don't see too. I was talking to some traders at a hedge fund recently and I asked them— I was trying to understand the role secrets play in the success of a hedge fund. The reason I was interested in that is because I think the AI labs are going to enter a new similar dynamic where their secrets are very valuable. If you have a 50% training efficiency improvement in your training runs cost $100 million, that is a $50 million secret that you have that you want to keep. Hedge funds do that kind of thing routinely. I asked some traders at a very successful hedge fund, if you had maybe your smartest trader get on Twitch for 10 minutes once a month and on that Twitch stream describe their 30-day-old trading strategies, right? So not your current ones, but the ones that are a month old. How would that affect your business after 12 months of doing that? So 12 months, 10 minutes a month, 30-day lookback, so it's 2 hours in a year. And to my shock, they told me 80% reduction in their profits. Like, it would have a huge impact. And then I asked, okay, so how long would the lookback window have to be before it would have a relatively small effect on your business? And they said 10 years. That I think is just quite strong evidence that the world's not perfectly efficient because these folks make billions of dollars using secrets that could be related in an hour or something like that. And yet others don't have them or their secrets wouldn't work. I think there are Different levels of efficiency in the world, but on the whole, our default estimate of how efficient the world is is far too charitable.

On the particular point, by the way, of AI labs potentially sharing secrets, you have this sort of strange norm of different people from different AI labs not only being friends but often living together. So it would be like Oppenheimer living with somebody working on the Russian atomic bomb, or something like that. Do you think those norms will persist once the value of these secrets is realized?

Yeah, I was just wondering about that some more today. I mean, it seems to be sort of slowing. They seem to be trying to close the valves, but I think there's a lot of things working against them in this regard. One is, again, that the secrets are relatively simple. Two is that you're coming off this academic norm of publishing and really open— the entire culture is based on sharing and publishing. 3 is, as you said, they all live in group houses, some are in polycules. There's just a lot of intermixing. And then it's all in California and California is a non-compete state. We don't have non-competes. So you'd have to change the culture, get everybody their own house and move to Connecticut and then maybe it would work. I think ML engineer salaries and compensation packages will probably be adjusted to try to address this because you don't want your secrets walking out the door. There are engineers, Igor Babuschkin for example, who has just, I believe, joined Twitter. Is that right? Elon hired him to train— I think that's public. Is that right? I think it is.

It will be now.

Igor's a really, really great guy and brilliant, but he also happens to have trained state-of-the-art models at DeepMind and OpenAI. That's the set of people who have that set of— I don't know whether that's a consideration or how big of an effect that is, but it's the kind of thing that it would make sense to value if you think there are valuable secrets that have not yet proliferated. I think they're going to try to slow it down. Publishing has certainly slowed down dramatically already, but I think there's just a long way to go before you're anywhere in hedge fund or Manhattan Project territory. Probably secrets will still have a relatively short half-life.

As somebody who has been involved in open source your entire life, are you happy that this is the way that AI has turned out or do you think that this is less than optimal?

Well, I don't know. My opinion's been changing. I have increasing worries about safety issues, not the hijacked version of safety, but some industrial accident type situations or misuse. We're not in that world and I'm not particularly concerned about it in the short term, but in the long term, I do think there are worlds that we should be a little bit concerned about, although I don't know what to do about it, where bad things happen. The probability mass of my belief though is that it's probably better on the whole for more people to get to tinker with and use these models, at least in their current state, For example, when Georgi Gerganov this weekend did a 4-bit quantization of the LLaMA model and got it inferencing on an M1 or M2, I was very excited and I got that running and it's fun to play with. Now I've got a model that is very good. It's almost GPT-3 quality, runs on my laptop. I've sort of grown up in this world of the tinkerers and open source folks and the more access you have, The more things you can try. So I think I do find myself very attracted to that.

I guess that is the scientist and the ideas part of what is being shared, but there's also another part about the actual substance, right? So like the uranium and the sort of atom bomb analogy. As different sources of data realize how valuable their data is for training newer models, do you think that these things will become harder to scrape? LibGen archive, Are these going to become rate limited in some way or what are you expecting there?

Well, first there's so much data on the internet. The 2 primitives that you need to build models are you need lots of data. We have that in the form of the internet. We digitized the whole world into the internet. And then you need these GPUs, which we have because of video games. You take the internet and video game hardware and you smash them together and you get machine learning models. And they're both commodities. I don't think anyone in the open source world is really going to be data limited. for a long time. There's so much that's out there. Probably people who have proprietary datasets that are readily scrapable have been shutting those down. So get your scraping in now if you need to do it. But that's just on the margin. I still think there's quite a lot that's out there to work with. I think this is the year of proliferation. This is a week of proliferation. We're going to see 4 or 5 major AI announcements week. New models, new APIs, new platforms, new tools from all the different vendors. In a way, they're all looking forward. My Herculaneum project is looking backwards. I think it's extremely exciting and cool, but it is sort of a funny contrast.

Okay, so I guess before I delve deeper into AI, I do want to talk about GitHub. I think we should start with, you were at Microsoft and at some point you realized that GitHub is very valuable and worth acquiring. How did you realize that and how did you convince Microsoft to purchase GitHub?

I had started a company called Xamarin together with Miguel de Icaza and Joseph Hill and we had built mobile tools and platforms. Microsoft acquired the company in 2016 and I was excited about that and I thought it was great. But to be honest, I didn't actually expect or plan to spend more than a year or so there. But when I got in there, I got exposed to what Satya was doing and just the quality of his leadership team. I was really impressed. And actually, I think I saw him in the first week or so I was there and he asked me, what do you think we should do at Microsoft? And I said, well, I think we should buy GitHub.

When would this have been?

This was my first week. It was March or April of 2016. And then he said, Yeah, it's a good idea. We thought about it. I'm not sure we can get away with it or something like that. And then it was about a year later, a little more than a year later, I wrote him an email, just a memo. I sort of said, I think it's time to do this. There was some noise that Google was sniffing around. I think that may have been manufactured by the GitHub team, but it was a good catalyst because it was something I thought made a lot of sense for Microsoft to do anyway. And so I wrote an email to Satya sort of a little memo saying, hey, I think we should buy GitHub. Here's why. Here's what we should do with it. And the basic argument was developers are making IT purchasing decisions now. It used to be this sort of IT thing, and now developers are leading that purchase. And it's this sort of major shift in how software products are acquired. And Microsoft really was an IT company. It was not a developer company in the way most of its purchases were made, but it was founded as a developer company. Microsoft's first product was a programming language. Yeah, I said, look, the challenge that we have is there's an entire new generation of developers who have no affinity with Microsoft and the largest collection of them is at GitHub. And if we acquire this and we do a merely competent job of running it, we can earn the right to be considered by these developers for all the other products that we do. And to my surprise, Satya replied in like 6 or 7 minutes and said, I think this is very good thinking. Let's meet next week or so and talk about it. And I ended up at this conference room with him and Amy Hood and Scott Guthrie and Kevin Scott and several other people. And they said, okay, tell us what you're thinking. And I kind of did a little 20-minute ramble on it. And Satya said, I think we should do it and why don't we run it independently like LinkedIn? Nat, you'll be the CEO. And he said, do you think we can get it for $2 billion? And I said, we could try. And 3 weeks later, he said, okay, go do this. Scott will support you on this. 3 weeks later, we had a signed term sheet and an announced deal. And then it was an amazing experience for me. I'd been there less than 2 years and Microsoft was made up of and run by a lot of people who'd been there for many years. and they trusted me with this really big project and it made me feel really good to be trusted and empowered. I had grown up in the open source world and so for me to get an opportunity to run GitHub, it's like getting appointed mayor of your hometown or something like that. It felt cool and I really wanted to do a good job for developers. And so that's how it happened.

That's actually one of the things I want to ask you about because often when something succeeds, we kind of think it was inevitable that it would succeed. But at the time, I remember, I mean, it was a while back, but I remember that there was a huge amount of skepticism. I would go on Hacker News and the top thing would be a blog post about how Microsoft's going to mess up GitHub. I guess those concerns have been alleviated throughout the years, but how did you deal with that skepticism and deal with that distrust?

Well, I was really paranoid about it and I really cared about what developers thought. I think there's always this question of who are you performing for? Who do you actually really care about? Who's the audience that's in your head that you're trying to do a good job for, impress, earn the respect of, whatever it is? And though I love Microsoft and care a lot about Satya and everyone there, I really cared about the developers. I'd grown up in this open source world, and so for me to do a bad job with this central institution in open source would've been a devastating feeling for me. It was very important to me not to. So that was sort of the first thing, is just that I cared. Then the second thing is that the deal leaked. It was going to be announced, I think, on a Monday. It leaked on a Friday that Microsoft's buying GitHub. The whole weekend there were terrible posts online, people saying we've got to evacuate GitHub as quickly as possible. We're like, oh my God, this is terrible. Then Monday we put the announcement out and we said, we're acquiring GitHub. It's going to run as an independent company. And then it said Nat Friedman's going to be CEO. I don't want to overstate it or whatever, but I think a couple people were like, oh, Nat comes from open source. He spent some time in open source and it's going to be run independently. So I don't think they were really that calmed down, but at least a few people thought, oh, maybe I'll give this a few months and just see what happens before I migrate off. And then my first day as CEO after we got the deal closed, 9:00 AM the first day, I was in this room and we got on Zoom and all the heads of engineering and product. I don't know what people were expecting, but I think maybe they were expecting some kind of longer-term strategy or something. But I came in and I said GitHub had no official feedback mechanism that was publicly available, but there were several GitHub repos that community members had started. Isaac from npm had started one. where he'd just been allowing people to give GitHub feedback and people had been voting on this stuff for years. I kind of shared my screen and put that up sorted by votes and said, 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. I think people were like, this is the new CEO strategy? And they were like, I don't know, we can't. You have to do database migrations. Can't do that in a day. And then someone's like, well, maybe we can do this. We had this sort We actually have a half implementation of this. And we eventually found something that we could fix by the end of the day. And what I'm thinking is, what I hope I said was, what we need to show the world is that GitHub cares about developers, not that it cares about Microsoft. If the first thing we did after the acquisition was to add Skype integration, developers would've said, oh, we're not your priority. You have new priorities now. And so the idea was just to find ways to make it better for the people who use it and have them see that we cared about that immediately. And so I said, we're going to do this today and then we're going to do it every day for the next 100 days. And it was cool because I think it created some really good feedback loops, at least for me. One was, you ship things and then people are like, oh hey, I've been wanting to see this fixed for years and now it's fixed. It's a relatively simple thing. So you get this sort of nice dopaminergic feedback loop going there. And then people in the team feel the excitement of shipping stuff. I think GitHub was a company that had a little bit of stage fright about shipping previously and sort of break that static friction and ship a little bit more, I think felt good. And then the other one is just the learning loop. By trying to do lots of small things, I got exposed to like, okay, this team is really good. Or this part of the code has a lot of tech debt. Or, hey, we shipped that and it was actually kind of bad. How come that design got out? Whereas if the project had been some 6-month thing, I'm not sure my learning would've been quite as quick about the company. There's still things I missed and mistakes I made for sure, but that was part of how I think. No one knows counterfactually whether that made a big difference or not, but I do think that earned some trust.

Most acquisitions don't go well. Not only do they not go as well, but they don't go well at all, right? As we're seeing in the last few months with a certain one. Why do most acquisitions fail or fail to go well?

Yeah, it is true. Most acquisitions are destructive of value. What is the value of a company? In an innovative industry, the value of the company, a lot of it boils down to its ability culturally to produce new innovations and is some sensitive harmonic of cultural elements that sets that up, that makes that possible. And it's quite fragile, I think. And so 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 of the company. So I think that maybe one way to think about it is companies are a little bit fragile. And so when you acquire them, it's relatively easy to break them. They're also more durable than people think in many cases too. I would say another version of it is the people who really care leave. And so the people who really care about building great products and serving the customers, maybe they don't want to work for the acquirer. And the set of people that are really load-bearing around the long-term success is small. When they leave or get disempowered, you get very different behaviors.

I want to go into the story of Copilot because until ChatGPT, I guess it was the most widely used application of the modern AI models. Whatever part of the story you're willing to share in public.

I've talked about this a little bit. GPT-3 came out in May, I think, of 2020. and I saw it and it really blew my mind. I thought it was amazing. I was CEO of GitHub at that time and I thought, I don't know what, but we've got to build some product with this. We've got to build something. So Satya had, at I think Kevin Scott's urging, already invested in OpenAI like a year before GPT-3 came out. This is quite amazing. He invested like a billion dollars.

By the way, do you know why he knew that OpenAI would be worth investing in at that point?

I don't know, actually. I've never asked him. But yeah, I'm not sure. That's a good question. I think OpenAI had already had some successes that were noticeable. And I think if you're Satya and you're running this multi-trillion-dollar company, you're trying to execute well and serve your customers, but you're always looking for the next gigantic wave that is going to upend the technology industry, it's not just about trying to win cloud. It's like, okay, what comes after cloud? And so you have to make some big bets and I think he thought AI could be one. I think Kevin Scott deserves a lot of credit for really advocating for that aggressively. And I think Sam Altman did a good job of building that partnership because he knew that he needed access to the resources of a company like Microsoft to build large-scale AI and eventually AGI. I think it was some combination of those 3 people coming together to make it happen, but I still think it was a very prescient bet. I've said that to people and they've said, well, $1 billion is not a lot for Microsoft. Yeah, but there were a lot of other companies that could have spent $1 billion to do that and did not. So I still think I think that deserves a lot of credit. Okay, so GPT-3 comes out. I pinged Sam and Greg Brockman at OpenAI and they were like, yeah, we've already been experimenting with GPT-3 and derivative models encoding context. Let's definitely work on something. And to me at least and a few other people, it was not incredibly obvious what the product would be. Now I think it's trivially obvious. Autocomplete, my gosh, isn't that what the models do? But at the time, actually my first thought was that it was probably going to be like a Q&A chatbot Stack Overflow type of thing. That was actually the first thing we prototyped. We grabbed a couple of engineers, this guy Uge who had come in from an acquisition that we'd done and Alex Gravely, and started prototyping. The first prototype was a chatbot. What we discovered first was that demos were fabulous. Every AI product has a fantastic demo. You get this sort of wow moment. That turns out to be maybe not a sufficient condition for a product to be good because at the time the models were just not reliable enough. They were not good enough. I ask you a question, 25% of the time you give me an incredible answer that I love, 75% of the time your answer is useless or wrong. It's not a great product experience. Then we started thinking about code synthesis and our first attempts at this were actually large chunks of code synthesis, like synthesizing whole function bodies. We built some tools to do that and put them in the editor and that also was not really that satisfying. The next thing that we tried was to just do simple small-scale autocomplete with the large models and we used the IntelliSense dropdown UI to do that. That was better, definitely pretty good. But the UI was not quite right and we lost the ability to do this large-scale synthesis. We still had that, but the UI for that wasn't good. We had it, I think, so that to get a function body synthesized, you would hit a key. And then I don't know why this was the idea everyone had at the time, but several people had this idea that it should display multiple options for the function body and then the user would read them and pick the right one. I think the idea was that we would use that human feedback to improve the model. But that turned out to be a bad experience because first you had to hit a key and explicitly request it, then you had to wait for it, and then you had to read 3 different versions of a block of code. Reading one version of a block of code takes some cognitive effort. Doing it 3 times takes more cognitive effort. And then most often the result of that was none of them were good or you didn't know which one to pick. That was also like you're putting a lot of energy in, you're not getting a lot out. It's sort of frustrating. So once we had that sort of single line completion working, I think Alex had the idea of saying we can use the cursor position in the AST to figure out heuristically whether you're at the beginning of a block in the code or not. And if it's not the beginning of a block, just complete a line. If it's the beginning of a block, show inline a full block completion so that the sort of number of tokens you request and when you stop gets altered automatically with no user interaction. And then the idea of using this sort of gray text like Gmail had done in the editor. So we got that implemented and it was really only 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 the median new user loved Copilot and wouldn't stop using it. That took 4 months, 5 months of just tinkering and exploring. There were other dead ends that we hit along the way. Then it became quite obvious that it was good because we had hundreds of internal users who were GitHub engineers. I remember the first time I looked at the retention numbers, they were extremely high. It was like 60+% after 30 days from first install. If you installed it, the chance that you were still using it after 30 days was over 60%. And it's a very intrusive product. I mean, it's sort of always popping UI up. And so if you don't like it, you will disable it. Indeed, 40-something percent of people did disable it. But those are very high retention numbers for an alpha first version of a product that you're using all day. And so then I was just incredibly excited to launch it and now it's improved dramatically since Sounds very similar to the Gmail story, right?

It's an incredibly valuable insight and then it becomes obvious that it needs to go outside. Okay, we'll go back to the AI stuff in a second, but some more GitHub questions. By what point, if ever, will GitHub profiles replace resumes for programmers?

That's a good question. I think they're a contributing element to how people try to understand a person now, but I don't think they're a definitive resume. We introduced READMEs on profiles when I was And I was excited about that because I thought it gave people some degree of personalization. I think some people have done that. Many thousands of people have done that. There's forces that push in the other direction too on that one where people don't want their activity and skills to be as legible. And there may be some adverse selection as well where the people with the most elite skills, it's rather gauche for them to signal their competence on their profile. So there's some weird social dynamics that feed into it too. But I will say, I think it effectively has this role for people who are breaking through today. One of the best ways to break through— I know many people who are in this situation. You were born in Argentina. You're a very sharp person, but you didn't grow up in a highly connected or prosperous network family. etc. And yet you know you're really capable and you just want to get connected to kind of the most elite communities in the world. And so if you're good at programming, you can join open source communities and contribute to them and you can very quickly accrete a global reputation for your talent, which is legible to many companies and individuals around the world. And suddenly you find yourself getting a job and moving maybe to the US or maybe not moving you end up at a great startup. I know a lot of people who've deliberately pursued the strategy of building reputation in open source and then you've got the sail up and the wind catches you and you've got a career. I think it plays that role in that sense, but in other communities like in machine learning research, this is not how you— there's 1,000 people, the reputation is more on Archive than it is on GitHub. So I don't know that it'll ever be comprehensive.

Are there any other industries for which proof of work of this kind will eat more into the way in which people are hired?

Well, I think there's a labor market dynamic in software where 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. And if the labor market dynamics weren't like that, if programming well were less economically valuable, then they would not— Companies wouldn't let them do that. They wouldn't let them publish a bunch of stuff publicly. They'd say that's a rule. And that used to be the case, in fact. As software has become more valuable, the leverage of a single super talented developer has gone up and they've been able to demand over the last several decades the ability to work in public. And I think that's not going away.

Other than that, we talked about this a little bit, but what has been the impact of developers being more empowered in organizations, even ones that are not traditionally IT organizations?

Software is kind of magic. You can write a for loop and do something a lot of times. When you build large organizations at scale, one of the things that does surprise you is the degree to which you need to systematize the behavior of the people who are working. When I first was starting companies and building sales teams, I had this wrong idea coming from the world as a programmer that salespeople were hyper-aggressive, hyper-entrepreneurial, making promises to the customer that the product wouldn't do, and that the main challenge you had with salespeople was restraining them from going out and aggressively cutting deals that shouldn't be cut. And what I discovered is that does exist sometimes, but the much more common case is that you need to build a systematic sales playbook, which is almost a script that you run on your sales team where your sales reps know the process they need to follow to exercise this repeatable sales motion and get a deal closed. And so I just had bad ideas there. I didn't know that that was how the world worked. But software is a way to systematize and scale out a valuable process extremely efficiently. I think the more digitized the world has become, the more valuable software becomes and the more valuable become the developers who can create it essentially.

Would 25-year-old Nat be surprised with how well open source worked? And how pervasive it is.

Yeah, I think that's true. I think we all have this image when we're young that these institutions are these implacable edifices that are evil and all-powerful and are able to, with master plans, substantially orchestrate the world. That is sometimes a little bit true, but they're very vulnerable to these new ideas and new forces and new communications media and stuff like that. Right now I think our institutions overall look relatively weak and certainly they're weaker than I thought they were back then. I thought Microsoft could stop open source. I thought that was a possibility. They can do some patent move and there's a master plan to ring-fence open source in. That didn't end up being the case. In fact, Microsoft, when we bought GitHub, we pledged all of our patent portfolio to open source. That was one of the things that we did as part of it. And so that was a kind of poetic moment for me, having been on the other side of patent discussions in the past, to be a part of and instrumental in Microsoft making that pledge. That was quite crazy.

Oh, that's really interesting. It wasn't that there was some business strategic reason, more so it was just an idea whose time had come.

Well, GitHub had made such a pledge and so I think in part in acquiring GitHub, we had to either try to annul that pledge or sign up to it ourselves. And so there was sort of a moment of a forced choice, but everyone at Microsoft thought it was a good idea too. So I think in many senses it was a moment whose time had come and the GitHub acquisition was a forcing function.

What do you make of critics of modern open source like Richard Or people who advocate for free software saying that corporations might advocate for open source because of practical reasons for getting good code. The real value of software and the real way that software should be made, it should be free in that you can replicate it, you can change it, you can modify it, and you can completely view it. And that the ethical values about that should be more important than the practical values. What do you make of that critique?

I think those are the things that he wants. I think the thing that maybe he hasn't updated is that maybe not everyone else wants that. He has this idea that people want freedom from the tyranny of a proprietary intellectual property license, but 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 kind of work.

Right.

There are places where freedom is really valuable, but There's always this thing of, 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, whether I want them to want it or not. And I think Richard is strongly in the former camp.

What is the most underrated license, by the way?

I mean, I don't know. Maybe the MIT license is still underrated because it's just so simple and bare.

Nadia Iqbal had a book recently where she argued that the key constraint on open source software and on the time of the people who maintain it is the community aspect of software. They had to deal with feature requests and discussions and maintaining for different platforms and things like that. It wasn't the actual code itself, but rather this extracurricular aspect that was the main constraint. Do you think that is the constraint for open source software? How do you see what is What is holding back more open source software?

I think by and large I would say that there is not a problem. Open source software continues to be developed, continues to be broadly used, and there's areas where it works better and areas where it works less well, but 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, standalone components and very horizontal things like operating systems. And it tends to be worse at user experiences and things where you need a sort of dictatorial aesthetic or an editorial control. I've had debates with Dylan Field at Figma as to why it is that we don't have lots of good open source applications. And I've always thought it had something to do with this governance dynamic of, gosh, it's such a pain to coordinate with tons of people who all sort of feel like they have a right to to push the project in one way or another. Whereas in a hierarchical corporation, there can be a head of this product or a CEO or founder or designer who just says, we're doing it this way, and you can really align things in one direction very, very easily. Dylan has argued to me that it might be because there's just fewer designers, people with good design sense in open source. I think that might be a contributing factor too, but I think it's still mostly the governance thing. And I think that's what Nadia's pointing at also. You're running a project, you gave it to people for free. For some reason, giving people something for free creates the sense of entitlement and then they feel like they have the right to demand your time and push things around and give you input and you want to be polite and it's very draining. I think that where that coordination burden is lower is where open source tends to succeed more. And probably software and other new forms of governance can improve that. expand the territory that open source can succeed in.

Theoretically, those 2 things are consistent. You could have very tight control over governance while the code itself is open source.

This happens in programming languages. Languages can't be designed— I mean, they often are eventually set in stone and then advanced by committee. But certainly you have these benign dictators of languages who enforce a strong set of ideas. They have a vision master plan. So that would be the argument that's most on Dylan's side. It's like, hey, it works for languages, why can't it work for end-user applications? I think the thing you need to do though to build a good end-user application is not only have a good aesthetic and idea, but somehow establish a tight feedback loop with a set of users where you can give it to Torcash, try this. Oh my gosh, okay, that's not what you need. Doing that is so hard even in a company where you have total hierarchical control of the team in theory. and everyone really wants the same thing and everyone's salary and stock options depend on the product being accepted by these users, it still fails many times in that scenario. Then additionally doing that in the context of open source, I think it's just slightly too hard.

The reason you acquired GitHub, as you said, is that there seemed to be this complementarity between Microsoft and GitHub's missions. And I guess that's been proven out over the last few years. Should there be more of these collaborations and acquisitions? Should there be more tech conglomerates? Would that be good for the system?

I don't know if it's good, but I think there are— yes, it is certainly efficient in many ways. I think we are seeing agglomeration occur because the math is sort of pretty simple. If you are a large company and you have a lot of customers, then 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. So if you can then go and just buy the product for a small amount of money and make it available to all of your customers, then there's often an immediate, really obvious gain from doing that. So in that sense, acquisitions make a ton of sense. And I've been surprised that the large companies haven't done many more acquisitions in the past until I got into a big company and started trying to do acquisitions and I saw that there are strong elements of the internal dynamics that make it hard. It's easier to spend $100 million on employees internally to do a project than to spend $100 million to buy a company. The dollars are treated differently, the approval processes are different, the cultural buy-in processes are different. And then To the point of the discussion we had earlier, many acquisitions do fail. And when an acquisition fails, it's somehow louder and more embarrassing than when some new product effort you've spun up doesn't quite work out as well. So I think there's lots of internal reasons, some somewhat justified and some less so, that they haven't been doing it. But just from an economic point of view, it seemed like it makes sense to see more acquisitions than we've seen.

Why did you leave?

I think as much as I loved Microsoft and certainly as much as I love GitHub, I really truly still feel tremendous love for GitHub and everything that it means to the people who use it. I didn't really want to be a part of a giant company anymore. I think building Copilot was an example of this. It wouldn't have been possible without OpenAI and Microsoft and GitHub, but building it also required navigating this really large group of people between Microsoft and OpenAI and GitHub. And you reach a point where you're spending a ton of time on just navigating and coordinating lots of people. And I just find that less energizing. Back to enthusiasm, just my enthusiasm for that was not as high. And I was torn about it because I truly love GitHub, the product, and there was so much more still knew we could do, but I was proud of what we'd done and I miss the team and I miss working on GitHub. It was really an honor for me. But yeah, it was time for me to go do something. I was always a startup guy. I always liked small teams and I wanted to go back to a smaller, more nimble environment.

Okay, so we'll get to it in a second, but first I want to ask about nat.org and the 300 words, the list there, which is I think one of the most interesting I guess it's the very Strawsonian list of 300 words I've seen anywhere. I'm just going to mention some of these and get some of your commentary. You should probably work on raising the ceiling, not the floor. Why?

Well, first I say probably, but what does it mean to raise the ceiling or the floor? I've just observed a lot of projects that set out to raise floor, meaning, gosh, we are fine, but they are not, and we need to go help them with our superior prosperity and understanding of their situation. And many of those projects fail. So for example, there were a lot of attempts to bring internet to Africa by large and wealthy tech companies and American And I won't say they all had no effect. That's not true. But many of them were far short of successful. There were satellites, there were balloons, there were high-altitude drones, there were mesh network laptops that were pursued by all these companies. And by the way, by perfectly well-meaning, incredibly talented people who I think did in some cases see some success, but overall probably much less than they ever hoped. But if you go to Africa, there is internet now. And the way 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. I mean, in the movie Wall Street from the '80s, you know, he's got that gigantic brick cell phone. That thing cost like 10 grand at the time. That was a ceiling-raising technology. It eventually went down the learning curve and became cheap. and the cell towers and cell phones. Eventually, we've got now hundreds of millions or billions of them in Africa. It was that initially ceiling-raising technology and then the force of capitalism that made it work in the end. It was not any deus ex machina technology solution that was intended to raise the floor. I think there's something about that that's not just an incidental example. But on my website, I say probably because there are some examples where I think people set out to kind of raise the floor and say no one should ever die of smallpox again. No one should ever die of Guinea worm again. And they succeed. I wouldn't want to discourage that from happening, but I think on balance, we have too many attempts to do that that look good, feel good, sound good, and don't matter and in some cases have the opposite of the effect they intend to.

Here's another one and this is under the EMH section. In many cases, it's more accurate to model the world as 500 people than 8 billion. Now here's my question. What are the 8 billion minus 500 people doing? Why are there only 500 people?

I don't know exactly. It's a good question. I ask people that a lot. The more I've done in life, the more I've been mystified by this. Like, oh, somebody must be doing X. And then you hear there's a few people doing X and then you look into it, they're not actually doing X. They're doing some version of it that's not that. And so all the best moments in life occur when you find something that to you is totally obvious that clearly somebody must be doing, but no one is doing. Mark Zuckerberg says this about founding Facebook. Surely the big companies will eventually do this and create this social and identity layer on the internet. Microsoft will do this. But no, none of them were and he did it. So what are they doing? I think the first thing is many people throughout the world are optimizing local conditions. They're working in their town, their community, they're doing something there. So the set of people that are thinking about global conditions is just naturally narrowed by the structure of the economy. That's number one. I think number two is most people really are quite memetic. I think we all are, including me. We get a lot of ideas from other people and so our ideas are not our own. We kind of got them from somebody else. It's kind of copy-paste. So you have to work really hard not to do that and to be decorrelated. I think this is even more true today because of the internet. I don't know if Albert Einstein as a patent clerk wouldn't he have just been on Twitter just getting the same ideas as everybody else? What do you have as decorrelated ideas? So I think the internet's correlated us more. The exception would be really disagreeable people who are just naturally disagreeable. And so I think the future belongs to the autists in some sense because they don't care what other people think as much. Those of us on the spectrum in any sense, I think, are in that category. Then yeah, I think we have this belief that the world's efficient and it isn't. I think that's part of it. The other thing is that the world is so fractal and so interesting. I mean, Herculaneum Papyri is this corner of the world that I find totally fascinating, but I don't have any anticipation that 8 billion people should be thinking about that or that that should be a priority for everyone.

Okay, here's another one. Large-scale engineering projects are more soluble in IQ than they appear. Here's my question. Does that make you think that the impact of AI tools like Copilot will be bigger or smaller because of engineering? Because one way to look at Copilot is actually its IQ is probably less than the average engineer, so maybe it'll have less impact, right?

Yeah, but I think it increases the productivity of the average engineer to bring them higher up. I think it increases the productivity of of the best engineers as well. Certainly a lot of the people I consider to be the best engineers tell me that they find it increases their productivity a lot. So yeah, I think AI is going to completely change. It's really interesting how so much of what's happened in AI has been sort of soft fictional work. You have Midjourney, you have copywriting, you have Claude from Anthropic is so literary it writes poetry so well, except for Copilot, which is this real hard area where the code has to compile, it has to be syntactically correct, it has to work and pass the tests. We see this steady improvement curve where now already on average more than half of the code is written by Copilot. I think when it shipped it was low 20s. So it's really improved a lot as the models have gotten better and the prompting has gotten better. But I don't see any reason why that won't be 95 5%. It seems very likely to me. I don't know what that world looks like. It seems like we might have more special purpose and less general purpose software. 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. Once you have much cheaper software, do you create more special purpose software? That's a possibility. Every company just a custom piece of code in a sense. Maybe that's the kind of future we're headed towards. So yeah, I think we're going to see enormous amounts of change in software development.

Another one. The cultural prohibition on micromanagement is harmful. Great individuals should be fully empowered to exercise their judgment. The rebuttal to this is that if you micromanage, you're preventing people from learning and to develop their own judgment.

Imagine you go into some company, they hire Dwarkesh, And you do a great job with the first project that they give you.

Right.

And so everyone's really impressed. Man, Dwarkesh, he made the right decisions. He worked really hard. He figured out exactly what needed to be done and he did it extremely well. And so over time you get promoted into positions of greater authority. And the reason the company's doing this is they want you to do that again, but at a bigger scale. Do it again, but 10 times bigger. The whole product instead of part of the product or 10 products instead of one. And so the company is telling you, you have great judgment and we want you to exercise that at a greater scale. Meanwhile, 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. And I think maybe in the— I don't know if this is true or not, but maybe in the '80s I would've felt the other side of this, that we have too much micromanagement. I think the other problem that people have is that they don't like micromanagement because they don't want bad managers to micromanage. So you have some bad managers, they have no expertise in the area, they're just kind of people managers, and they're starting to micromanage something that they don't understand where their judgment is And my answer to that is stop empowering bad managers. Don't have them. Just don't have bad managers. Promote and empower people who have great judgment and do understand the subject matter that they're working on. If I work for you and I just know you have better judgment and you come in and you say, Nat, you're launching the Scroll thing and I think you've got the file format wrong. Here's how you should do it. I would welcome that even though it's micromanagement because it's going to make us more successful and I'm going to learn something from that. And I know your judgment is better than mine. in this case, or at least we're going to have a conversation about it and we're both going to get smarter. So I think on balance, yeah, there are cases where people have excellent judgment and we should encourage them to exercise it. Sometimes things will go wrong when you do that, but on balance, you'll get far more excellence out of it. We should empower individuals who have great judgment.

There's a quote about Napoleon that if he could have been in every single theater of every single battle he was part of, that he would've never lost a battle. I was talking to somebody who worked with you at GitHub and she emphasized to me, and this is really remarkable to me, that even the applications that were being shipped out to engineers, how much of the actual suggestions and the actual design came from you directly, which is kind of remarkable to me that as CEO you would—

Yeah, you can probably find people you can talk to who think that was terrible. The question is always, does that scale? And the answer is it does not scale. It doesn't. But the set of people who really do have great judgment— the experience that I had as CEO was I was terrified all the time that there was someone in the company who really knew exactly what to do and had excellent judgment, but because of cultural forces, that person wasn't empowered. That person was not allowed to exercise their judgment. and make decisions. When I would think and talk about this, that was the fear that it was coming from. They were in some consensus environment where their good ideas were getting whittled down by lots of conversations with other people and a politeness and a desire not to micromanage. So we were ending up with some kind of average thing. 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 just a disaster and it didn't work. So now you know it didn't work and you can start over. I would rather have those more high-variance outcomes and I think it's a worthy trade.

Let's talk about AI. What percentage of the economy is basically text-to-text?

It's a good question. We've done the Bureau of Labor Statistics analysis of this and Yeah, it's not the majority of the economy or anything like that. We're in the low double-digit percentages. The thing that I think is hard to predict is what happens over time as the cost of text-to-text goes down. I don't know what that's going to do, but there's plenty of revenue to be got now. One way you can think about it is, okay, we have all these benchmarks for machine learning models. There's LaMDA and there's this and there's that. Those are really only useful and only exist because we haven't deployed the models really at scale. So we don't have a sense of what they're actually good at. The best metric would probably be something like what percentage of economic tasks can they do? Or on a gig marketplace like Upwork, for example, What fraction of Upwork jobs can GPT-4 do? I think it's sort of an interesting question. My guess is extremely low right now autonomously, but over time it will grow. Then the question is, what does that do for Upwork? I'm guessing it's probably a $5 billion GMV marketplace, something like that. Does it grow? Does it become $15 billion or $50 billion? Does it shrink because the cost of text-to-text tasks goes down? I don't know. My bet would be that we find more and more ways to use text-to-text to advance progress. And so overall, there's a lot more demand for it. But yeah, I guess we'll see.

And at what point does it happen? GPT-3 has been a rounding error in terms of overall economic impact. Does it happen with GPT-4 or GPT-5 where we see billions of dollars of usage?

I've got early access to GPT-4 and I've gotten to use it a lot. I honestly can't tell you the answer to that because it's so hard to discover what these things can do that the prior ones couldn't do. I was just talking to someone last night who told me GPT-4 is actually really good at Korean and Japanese and GPT-3 is much worse at those. So it's actually a real step change for those languages. I think people didn't know how good GPT-3 was until it got instruction tuned for ChatGPT and was put out in that format. You can imagine the pre-trained models as kind of unrefined crude oil and then once they've been RLHF'd and trained and then put out into the world, people can find the value.

What part of the AI narrative is wrong in the overoptimistic direction?

The probably overoptimistic case from both the people who are fearful of what will happen and from people who are expecting great economic benefits is that we're definitely in this realm of diminishing returns from scale. For example, I think GPT-4 is, my guess is, 2 orders of magnitude more expensive to train than GPT-3, but clearly not 2 orders of magnitude more capable. Now, is it 2 orders of magnitude more economically valuable? That would also surprise me. And so I think it's possible when you're in these sigmoids, where you are going up this exponential and then you start to asymptote, it can be difficult to tell if that's going to happen. So I think the idea that we might not run into hard problems or that scaling will continue to be worth it on a dollars basis, I think, are reasons to be a little bit more pessimistic than the people who have high certainty of GDP increasing by 50% per month or something like like that, which I think some people are predicting. But on the whole, I'm very optimistic. So you're asking me to make the bear case for something I'm very bullish about.

That's why I asked you to make the bear case because I know on the whole you are. I want to ask you about these foundation models. What is the stable equilibrium you think of how many of them will there be? Will it be a sort of oligopoly like Uber and Lyft?

I think there will probably be wide-scale proliferation. You sort of ask, what are the structural forces that are pro-proliferation and the structural forces pro-concentration. So I think the pro-proliferation case is a bit stronger. The pro-proliferation case is they're actually not that hard to train. The best practices will promulgate and you can kind of write them down on a couple sheets of paper. And to the extent that secrets are developed that improve training, those are relatively simple and they get copied around easily. Number one. Number two, the data is mostly public. It's mostly kind of data from the internet. Number 3, the hardware is mostly commodity and the hardware's improving quickly and getting much more efficient. And then I think there's a lot of techniques that overcome— I think some of these labs potentially have 50%, 100%, 200% training efficiency improvement techniques. And so there's just a lot of low-hanging fruit on the technique side of things. And so we're seeing it happen. I mean, it's happening this week and it's happening this year is that we're getting, I a lot of proliferation. The only case against proliferation is that you'll get concentration because of training costs. I don't know that that's true. I don't have confidence that the trillion-dollar model will be much more valuable than the $100 billion model and that it will even be necessary to spend a trillion dollars training it. Maybe there will be So many techniques available for improving efficiency that how much are you willing to spend on researchers to find techniques if you're willing to spend a trillion on training? That's a lot of bounties for new techniques and some smart people are going to take those bounties.

How different will these models be? Will it just be everybody chasing the same exact marginal improvement leading to the same marginal capabilities or will they have entirely different repertoires of skills and abilities?

Right now, back to the memetic point, they're all pretty similar, I would say. Basically the same rough techniques.

Right.

What's happened is an alien substance has sort of landed on Earth and we are trying to figure out what we can build with it. I think we're in this multiple overhangs here. We have sort of a compute overhang where there's much more compute in the world than is currently being used to train models. Much, much more. I think the biggest models are trained on maybe 10-ish thousand GPUs, but there's millions of GPUs. There's the compute overhang. And then we have, I think, 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, but that's just a matter of time until people do that. And then those capabilities haven't reached— because many of them are in the hands of the labs, they haven't reached the tinkerers of the world. I think that is where the new— what can this thing actually do? Until you get your hands on it, you don't really know. I think OpenAI were themselves surprised by how explosively ChatGPT has grown. I don't think they put ChatGPT out expecting that to be the big announcement. I think they thought GPT-4 was going to be their big announcement. I think it still probably is and will be big, but ChatGPT really surprised them. I think that's It's hard to predict what people will do with it and what they'll find valuable and what works. So you need tinkerers. It basically goes from hardware to researchers to tinkerers to products. That's the pipe. That's the cascade.

When I was scheduling my interview with Ilya, it was originally supposed to be around the time that ChatGPT came out. Their comms person tells me, listen, just so you know, this interview will be scheduled around the time we're going to make a minor announcement. It's not the thing you're thinking. It's not GPT-4, but it's just a minor thing. They didn't expect what it ended up being. Have incumbents gotten smarter than before? It seems like Microsoft was able to integrate this new technology really well.

There's been 2 really big shifts in the way incumbents behave in the last 20 years that I've seen. The first is that it used to be incumbents got disrupted by startups all the time. You had example after example of this in the mini computer, microcomputer era, et cetera. And then Clay Christensen wrote The Innovator's Dilemma. I think what happened was that everyone read it and they said, oh, disruption is this thing that occurs and we have this innovator's dilemma where we get disrupted because the new thing is cheaper and we can't let that happen. And they became determined not to let that happen and they mostly learned how to avoid it. They learned that you have to be willing to do some cannibalization and you have to be willing to set up separate sales channels for the new thing and so forth. And so we've had a lot of stability in incumbents for the last 15 years or so. And I think that's maybe why. That's my theory. So that's the first major step change. And then the second one is, man, they are paying a ton of attention to AI. If you look at the prior platform revolutions like cloud, mobile, internet, web, PC, all the incumbents derided the new platform and said, gosh, no one's going to use web apps. Everyone will use full desktop apps, rich applications. So there was always this sort of laughing at the new thing. The iPhone was laughed at by incumbents and that is not happening at all with AI. Now we may be at peak hype cycle and we're going to enter the trough of despair. I kind of don't think so though. I think people are taking it seriously and every live player CEO is adopting it aggressively in their company. So yeah, I think incumbents have gotten smarter.

All right, so let me ask you some questions that we got from Twitter. This is a former guest and I guess mutual friend Austin Vernon. Nat is one of those people that seems unreasonably effective. What parts of that are innate and what did he have to learn?

Well, that's very nice of Austin to say. I don't know. I think we talked a little bit about this before, but I think I just have a high willingness to try things. And get caught up in new projects and then I don't want to stop doing it. And so I think I just have a relatively low activation energy to try something and I'm willing to sort of impulsively jump into stuff. And many of those things don't work, but enough of them do that I guess I've been able to accomplish a few things. The other thing I would say, to be honest with you, is that I do not consider myself accomplished or successful.

Oh, really?

My self-image is that I haven't really done anything of tremendous consequence. I don't feel like I have this giant bed of achievements that I can go to sleep on every night. I think that's truly how I feel. I'm kind of an insecure overachiever. I don't really feel good about myself unless I'm doing good work. But I also have kind of cultivated a tried to cultivate a forward-looking view where I try not to be incredibly nostalgic about the past. I don't keep lots of trophies or anything like that. You go into some people's offices and there's things on the wall and trophies of all the things they've accomplished and that always seemed really icky to me. So I don't know, I just had a sort of revulsion to that.

Is that why you took down your blog?

Yeah, I just wanted to move forward.

Simeon asks, for your takes on alignment. He seems to invest both in capabilities and alignment, which is the best move under a very small set of beliefs. So he's curious to hear your reasoning there.

I guess we'll see. I'm not sure capabilities and alignment end up being these opposing forces. It may be that capabilities are very important for alignment. It may be that alignment is very important for capabilities. I think a lot of people believe, and I think I'm included in this, that AI can have tremendous benefits but that there's a small chance of really bad outcomes. Maybe some people think it's a large chance. I think the solutions, if they exist, are likely to be technical. They're probably some combination of technical and prescriptive. So there's probably a piece of code and a README file that says if you want to build aligned AIs, use this code and don't do this, or something like that. I think that's really important and more people should try to actually build technical solutions. I think one of the big things that's missing that sort of perplexes me is 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. I think we're going to end up, because the alignment people have this belief that they shouldn't release their work, in a world where there's a lot of open source pure capabilities work and no open source alignment work for a little while. And then hopefully that'll change. So yeah, I wanted to on the margin invest in people doing alignment. It seems like that's important. I thought Sydney was kind of an example of this. You had Microsoft essentially release an unaligned AI, And I think the world sort of said, hmm, sort of threatening its users? That seems a little bit strange. If Microsoft can't put a leash on this thing, who can? And so I think there'll be more interest in it and I hope there's open communities.

That was so endearing for some reason. It threatening you just made it so much more lovable for some reason.

Yeah, I think 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 post jobs or something like that, then I don't know, on a gig worker site or something. I think it could have been scary.

All right, before we go, where can people learn more about the Vesuvius Challenge?

Yeah, so Vesuvius Challenge is at scrollprize.org, S-C-R-O-L-L-P-R-I-Z-E dot org. Yeah, check it out. And I think it's very likely that somebody listening to this could be the one who wins the grand prize and decodes the scrolls.

Okay, excellent. Awesome. Okay, well, Nat, this is a true pleasure. Thanks for coming on the podcast.

Thanks for having me.

Hey everybody, I hope you enjoyed that episode. Just wanted to let you know that in order to help pay for the bills associated with this podcast, I'm turning on paid subscriptions on my Substack at warkeshpatel.com. No important content on this podcast will ever be paywalled, so please don't donate if you have to think twice before buying a cup of coffee. But if you have the means and you've enjoyed this podcast or gotten some kind of value out of it, I would really appreciate your support. As always, the most helpful thing you can do is just share the podcast, send it to people you think might enjoy it, put it in Twitter, your group chats, et cetera. Just blitz the world. Appreciate you listening. I'll see you next time. Cheers.