What’s Beyond GitHub Copilot? With Copilot's Chief Architect and founder of Minion.AI Alex Graveley

Everyone talks about the future impact of AI, but there’s already an AI product that has revolutionized a profession. Alex Graveley was the principal engineer and Chief Architect behind Github Copilot, a sort of pair-programmer that auto-completes your code as you type. It has rapidly become a product that developers won’t live without, and the most leaned-upon analogy for every new AI startup – Copilot for Finance, Sales, Marketing, Support, Writing, Decision-Making. Alex is a longtime hacker and tinkerer, open source contributor, repeat founder, and creator of products that millions of people use, such as Dropbox Paper. He has a new project in stealth, Minion AI. In this episode, we talk about the uncertain process of shipping Copilot, how code improves chain of thought for LLMs, how they improved product, performance, how people are using it, AI agents that can do work for us, stress testing society's resilience to waves of new technology, and his new startup named Minion. No Priors is now on YouTube! Subscribe to the channel on YouTube and like this episode. Show Links: Alex Graveley - San Francisco, California, United States | Professional Profile | LinkedIn Minion AI Sign up for new podcasts every week. Email feedback to [email protected] Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @alexgraveley | @ai_minion Show Notes: [1:50] - How Alex got started in technology [2:28] - Alex’s earlier projects with Hack Pad and Dropbox Paper [07:32] - Why Alex always wanted to make bots that did stuff for people [11:56] - How Alex started working at Github and Copilot [27:11] - What is Minion AI [30:30] - What’s possible on the horizon of AI

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:05Speaker 00

  The most fun stuff was like these people that would pop up and be like, "I don't program, but I just like learned how to write this hundred line script that like does this thing that I need." It's like, "Oh my God."

• 0:17Sarah Guo?

  This is the No Priors podcast. I'm Sarah Guo.

• 0:20Speaker 02

  I'm Elad Gil.

• 0:21Sarah Guo?

  We invest, invest, and help start technology companies.

• 0:24Speaker 02

  In this podcast, we're talking with the leading founders and researchers in AI about the biggest questions.

• 0:40Sarah Guo?

  Everyone talks about the future impact of AI, but this week on No Priors, we're talking with the architect of arguably the first AI product that has already revolutionized a profession: GitHub Copilot. Alex Gravely was the principal engineer and chief architect behind this product, a sort of pair programmer that auto completes your code as you type. It's rapidly become a product that developers won't live without, and the most leaned upon analogy for every new startup: Copilot for finance, for sales, marketing, support, writing, decision making, everything. Alex is a longtime hacker and tinkerer, open source contributor, repeat founder, and creator of products that millions of people use, such as Dropbox Paper. I have huge respect for the range of work he's done, ranging from

• 1:23Sarah Guo?

  hardware-level security and virtualization to real-time collaborative, web-native, and AI-driven UI. He has a new project in Stealth: Minion. Just as a heads up, we recorded this podcast episode in person, and you'll notice we don't sound as crisp as we usually do. Still, I think you'll enjoy listening to this conversation. Let's get into the interview.

• 1:44Sarah Guo?

  Welcome to the podcast, Alex.

• 1:46Speaker 00

  Hi.

• 1:46Sarah Guo?

  So let's start with some background. How did you end up working in tech or AI?

• 1:50Speaker 00

  Tech was earlier. I started, you know, really young, and I got really into Linux when I was 14 or so. And yeah, it was right around the time when the web was like a new thing and you had to work to kind of get on the web. The idea of helping in the open and making it freely available so other people could learn from things like I was learning from things seemed great. So I went and spent many years just working on open source stuff. Yeah, spent many hours compiling kernels and hacking on stuff. And yeah.

• 2:24Sarah Guo?

  And what was the thought process behind starting like HackPad?

• 2:28Speaker 00

  Oh, HackPad. Yeah, so I had just finished like four or five years at VMware. And I wanted to get into startups. I knew that. And then, so I left VMware and I started working on an education startup like many of us do. Many, many founders start with like the idea of an education startup. It's like a rite of passage. Yeah, it's like a rite of passage. Yeah, so I spent, I don't know, nine months working on that. Turns out education is very hard. So after nine months, I was like, all right, this isn't going anywhere. I know, I don't know if there's a value prop here. I mean, that's, that was the value is that I learned that like you have to make something that is both achievable and that people want to pay for, for spend their time on.

• 3:11Speaker 00

  So yeah, then I was just kind of fishing around. I was living in like a warehouse in San Francisco, a bunch of Burning Man people. And we were having trouble organizing large scale Burning Man projects. And so I forked Etherpad and started hacking on it, recruited a friend in the community to start working on with me and yeah, just grew from there. Before we did YC, we had most, many of the large Burning Man camps using it to organize their, their builds.

• 3:37Sarah Guo?

  Can you describe the product experience?

• 3:39Speaker 00

  Oh, yeah. It was a real time text editor, kind of like Google Docs. Google Docs is the only other one that did that at that time. It was kind of nice because it would highlight who said what. So you could go track down if somebody had a contribution, be like, oh, what did you mean? Or I heard you say something about this. So that's very useful in like large scale anonymous, pseudo anonymous groups where you don't really know who has ownership over what, like Burning Man camps. And let's see, and then we did YC and then we got a bunch, almost many of those people did it. It might like, my hack was to go do YC and try and get all the companies doing YC to use my product. Which many of them did. I also took like very extensive notes of all the YC presentations

• 4:24Speaker 00

  using the product that everyone would then like look at. And we were able to get Stripe, and Stripe used us for many years, actually. They built for the, I think we were their first knowledge base. They used it for a long time. A bunch of other big companies as well.

• 4:37Sarah Guo?

  And post Dropbox acquisition, you worked on what became Paper. How did you think about like what you wanted to go work on next?

• 4:46Speaker 00

  I spent the next few years kind of poking around at stuff. I knew that I wanted to make a robot that does stuff for you. So there's this company called Magic, so I worked at this company called Magic. They were doing this like text-based personal assistants.

• 5:00Sarah Guo?

  You remember this one?

• 5:01Speaker 02

  I think it's like everybody starts an education startup. Magic is one of those names of cycling, and there's a really cool AI company right now called Magic as well. So I feel like there's also these names that kind of persist from generation to generation, which I think is really cool.

• 5:13Sarah Guo?

  I'm sure you know it's CodeGen.

• 5:15Speaker 00

  Yeah, yeah, yeah. No, yeah. I haven't seen a demo yet, but it sounds like they're doing the right thing. There's a few people doing that kind of whole repository changes. So it seems like a great direction. It was all op space. It was super heavy op space. So, you know, they had teams of people, and it was 24 hours, and they would cycle in and, you know, lose context. And like, they're all busy because they're trying to like deal with lots of people with lots of requests all the time. And so really, it was like a crash course in like human behavior, right? Like, what do people do under stress? How do they act? What do they say? What can you train? What can't you train? Like, can you bucket stuff? And the answer is no. Like, humans are complicated, especially in text. The beauty of the web

• 5:56Speaker 00

  and like traditional UIs, right, is that you fill them in. And like, if it doesn't do what you want, then you make the decision to either go forward or not go forward, right? Text has this annoying property of you can be all the way at the 99% mark and then change the goal entirely. Yeah, so it's complicated. It's hard to understand what people want. It's hard to understand the complexity involved, especially when you're dealing with the real world. Flights get delayed, passwords get lost.

• 6:24Speaker 02

  Do you think we're the last generation to deal with that? In other words, it feels like we're about to hit transition point in which agents can actually start doing some of these things for us for real? When before I think all these products really started with these operations heavy approaches. I remember, you know, there was a really early sort of personalized search engine called Art Bark. For similarly, if you kind of look behind the hood, it was a lot of ops people and there was a little bit of algorithm, right?

• 6:46Speaker 00

  That was really would like, you would like sort of describe what you're good at or something and then they would try and send questions to you?

• 6:52Speaker 02

  Yeah, exactly. They kind of route things, but I think there was actually people doing some of the routing at least. Or I can't exactly remember, right? But I think it was, you know, I think a lot of people wanted to build these really complex sort of bots or agents that were doing really rich things and the technology just wasn't there, it feels like for a period of time. There was also some startups. I remember one company that I got involved with that was trying to do like virtual scheduling and assistance, you know, six, seven, eight years ago. And again, it felt like it was a little bit early, maybe.

• 7:19Sarah Guo?

  Clara Labs.

• 7:20Speaker 00

  Clara, that's the one. Yeah, yeah, yeah.

• 7:21Sarah Guo?

  Yeah, I remember this Clara and then we had an operator. Do you remember operator?

• 7:27Sarah Guo?

  And then like on the question answering stuff, we had Jelly. So there's a whole series of.

• 7:32Speaker 00

  Yeah, what we were able to do with magic as like an aside there, you know, I started working there because I wanted to work on the AI part. I think somewhere in there, Facebook M started as well. And it was just, it was a fun place to try and learn everything I could about solving those problems. So yeah, before Transformers, there was a sequence to sequence was kind of the previous iteration. We were able to like take all the, all the histories of the chats between assistants and people and train a little model on it. And a little by today's metrics and run it and it would like, it would show some gray text in the little text bar and people could edit it and hit enter and.

• 8:11Sarah Guo?

  The operators.

• 8:12Speaker 00

  Yeah, the operators. And yeah, we would measure how much time they spent typing with it and without it. And so it would save like half an hour across a hundred people per day, across an eight hour shift. So yeah, that was like my first shipping AI product.

• 8:26Sarah Guo?

  And then how do you end up going to Microsoft?

• 8:29Speaker 00

  Oh, yeah, there's a bunch of other stuff along with it. So after that, I got into crypto. My friend was doing H-Captcha, which was a sort of a captcha marketplace, which is now like something like the number one or number two captcha service in the world, which is crazy. So kind of launched that. That was fun. Annoyed people the world over for many, many man hours in aggregate. And then left that to work with Moxie on, on a cryptocurrency for Signal. So that was really fun, complicated, and it all worked in a few seconds. We were shooting for Venmo quality.

• 9:06Speaker 02

  When you think about crypto in the context of AI, because people talk about it in a few different contexts, right? One is you have programmatic sort of money as code running. And so that could create all sorts of really interesting things from an agent-driven perspective. But then the other piece of it is identity. And some people think, I mean, WorldCoin would be one example, but there's other examples of effectively trying to secure identity cryptographically on the blockchain in an open way. And then using that identity in the future to differentiate between AI-driven agents and people. Do you think that's going to be important or does that stuff not really matter in terms of the identity portion of not only crypto, but just like how we think about the future of agents?

• 9:41Speaker 00

  It's a good question. The honest answer is I think we're going to go through a many-year period of extreme discomfort where AIs pretend to be things or or confuse people or extract money from your grandparents or drain people's life savings in ways that are scary. And, you know, OpenAI is trying to do their best, but for some reason, the focus has been on OpenAI doing everything. And instead of like, we should go build the systems that prevent that. We should go pass the legislation that drops the hammer on people doing that stuff. We should go, all this kind of stuff that is, unfortunately, it seems like we're going to need some really bad things to happen before we align correctly.

• 10:24Speaker 00

  I'm not really scared about AIs killing us, although I'm very grateful that there are people that are thinking about it. I'm more worried about bad people using new technology to hurt us.

• 10:34Speaker 02

  Yeah, Elia from NEAR has some really interesting thoughts on this because he was one of the main author, or he was the last author on the transformer paper before he started NEAR. And he's brought up these concepts of like, how do you stress test society relative to the coming wave of AI, which I think is an interesting concept.

• 10:47Speaker 00

  Yeah, it's a great, it's a great way to look at it. Like, it's not as bad as it could be, right? If you think about it, there, most of the things that you want to spam either have a spam blocker or are somewhat difficult to create an account on. So doing a better job of SOC puppet account filtering is going to be really important going forward. You know, I like what Cloudflare is doing with their kind of fingerprinting instead of visual captures, which are not good enough anymore. One thing that is like kind of a saving grace here is that many of the things that you would want to do cost money. So calling everybody costs money. Texting everyone should be hopefully illegal soon, but also costs money. Maybe not enough money to prevent these things, but.

• 11:30Sarah Guo?

  Probably not enough as agents can make money, right? They can just look at the trade offs of cost.

• 11:36Speaker 00

  Yeah, I think it's interesting. I guess I would say it's not, you know, it's somewhere in the middle. It's like, imagine there's a, you know, North Korea has been trying to do this to us for a long time, right? Now there's a North Korea that has more resources or is more distributed or whatever. It's, you know, we have some mitigations. We need more. We need to be thinking about it a lot more.

• 11:55Speaker 02

  How did you end up at GitHub and how did you end up working on Copilot?

• 11:58Speaker 00

  While I was working on Mobile Coin, my dad's kidneys failed and I tried to donate a kidney. And they found a lump in my chest as part of the scans they do. And I had to have my right, most of my right lung removed in 2018. And so that was a big deal. And so it took some time. It's weird, healing from internal injuries takes a lot longer than you think. Anyway, happy story is that I don't have cancer now for over four years. And my dad's kidneys got, he got a transplant also, so things are good. And so after, oof, I don't know.

• 12:38Speaker 00

  I guess I was recovering for for quite a while and then I went and begged my friend for a job. I figured I should start working again. That was the transition, I guess. So I worked on some random stuff at first. I converted GitHub to using their own product to build GitHub, which was kind of fun. They weren't, yeah. So I think people still use CodeSpace now to build GitHub, which is pretty cool. But yeah, then this kind of opportunity to work with OpenAI came up. And because I had been tracking AI in the past and was pretty aware of what was going on, I jumped on it. Was that proposed by OpenAI or by GitHub or who kind of initiated it all? So I don't know the exact beginnings. I know that OpenAI and Microsoft were working on a deal for supercomputers.

• 13:21Speaker 00

  So they wanted to build a big cluster for training and there was a big deal that was being worked out. And there was some software kind of provisions thrown in. I think Office and Bing probably. And GitHub was like, oh, okay, well, maybe there's something GitHub can do here. I think OpenAI threw a small, threw a small fine tune over and was like, here's a small model trained on, um, on some code, see if this is interesting, you know? So we played around with it.

• 13:44Sarah Guo?

  What's small?

• 13:45Speaker 00

  I don't know. This was, uh, I have to remember now, I think this is before I knew very much. So it was definitely not a Da Vinci sized model, that's for sure. I don't know what size it was. Yeah. And so it was just like, I learned later that this was basically a training artifact. So they had wanted to see what introducing code into their base models would do. I think it had positive effects on chain of thought reasoning. Code is kind of linear, so you can imagine that, you know, you kind of do stuff one after another. And the things before have an impact. And yeah, it was not that good. It was very bad. It was, I think, just, like I said, just an artifact and a small sample of GitHub data that they had crawled. And this is before, actually, I joined.

• 14:27Speaker 00

  Me and this guy, Albert Ziegler, were the first two after UGAY. We got a hold of this model and started playing with it. And he was able to say like, oh, you know, like it doesn't work most of the time, but here it is doing something, you know, here's, here's, and it was only Python at that time. Here it is, you know, generating something useful. We didn't really understand anything. So that was enough to like, okay, well, you know, go fetch a couple of people and start working, see, see, see if there's anything there. We didn't really know what we had. So the first, you know, task was to go test it out, see what it did. We crowdsourced a bunch of Python problems in-house, stuff that we knew wouldn't be in the training set. And then we, we started work on fetching repositories and finding the tests in them so that we can basically generate functions that were being tested and see if the test

• 15:12Speaker 00

  still pass. There had been like a brand new PyTest feature introduced like recently that allowed you to see which functions were called by the PyTest. So you'll find that function, zero the body, ask the model to generate it, and then rerun the test and see if it passed. And I think it was less, I don't know, 10%, something like that of, of those guys. And the dimensions are kind of like, how many chances do you give it to solve something? And then how do you test whether it's worked or not? Right? So for the standalone tests, that was, we had people write test functions and then we would try to generate the body. And if the test passed, then you know it works. And in the wild test harness, we would download a repository, run all the tests, look at the ones that

• 15:57Speaker 00

  passed, find the functions that they called, make sure that they weren't trivial, generate the bodies for them, rerun the test, see what it passed, see what, and you get your percentage. Yeah, I mean, it was something like some very, very low percentage upfront, but we knew that there was kind of a lot more juice to squeeze. So like getting all of GitHub's code into the model, and then a bunch of other tricks that we hadn't, you know, we hadn't even thought of at that time. Yeah. And eventually, you know, it, it went from, you know, less than 10% on, in the wild test to over 60%. So that's like one in two tests it can just generate code for, which is insane. Right? Somewhere along the

• 16:37Speaker 00

  way, you know, there was like 10% to 20% to 35% to 45%. You know, these kind of like improvements along the way. Somewhere along the way, we did more prompting work so that the prompts got better. Somewhere along the way, they used, you know, all the versions of the code as opposed to just the most recent version. They used diffs so that it could understand small changes like.

• 17:00Speaker 00

  Yeah, it got better. And so but when we first started, we were just, we didn't know we had, we were just trying to figure it out. At the time, they were thinking in terms of like, maybe you can replace Stack Overflow or something, you know, do a Stack Overflow competitor.

• 17:13Sarah Guo?

  Was that the first like product idea you guys had for it?

• 17:16Speaker 00

  I don't know that we had that idea. I think that was kind of like a.

• 17:19Sarah Guo?

  Yeah, on high.

• 17:20Speaker 00

  Yeah, yeah, yeah. That was more like, it'd be nice if you made something that competed with Stack Overflow, because we have all this code, wouldn't it be nice to leverage it? Yeah. And so we made some UIs, but like, early on, you know, it was like, early on, it was bad. So it'd be like, you'd watch it, and it would, it would like, it would run, and most of them would be bad. And be like, it'd be like, one success. And be like, oh, sweet, I got a success, but I had to wait, you know, some number of seconds for.

• 17:45Sarah Guo?

  Was the test user group just like six of you, like some larger group?

• 17:50Speaker 00

  The first iteration was just like an internal tool that people would, that helped people write these tests. And then we wanted to see if maybe we could turn that into some UI that people would use. Where if there was some way to cover up the fact that one in 10 things pass, right? So you tried to, we've tried a few UI things there. And then it was actually OpenAI, it was like, it would be nice if we could, we're testing these model fine tunes, it'd be nice if we could like, test them more quickly. Can, what about doing like a VS Code extension? Just do autocomplete. And so, all right, sure, why not? And yeah, so we did autocomplete at first, and that was kind of, that was a big jump, you know, because they were still thinking in terms of like Stack Overflow, you know? But this is, it's like, you know, I didn't have any ideas, basically. I didn't know how to beat Stack Overflow with this thing.

• 18:36Speaker 00

  But we could play with some stuff in in VS Code that was maybe closer to the code, you know? First, we did autocomplete, and that was kind of fine, it was useful. It would show this little pop-up box like autocomplete does, and you could pick some strings. And so that format, actually, you know, the the usage was, you know, it's fine. It wasn't the right metaphor exactly, right? But you've got like this code generated mixed in with the specific terms that are in the code, and it's a little, it's not exactly the same thing. We tried things like adding a little button over top of empty functions. So you would go generate them or you could like hit a control key and it would create a big list on the side that you can choose from, or there's a little pop up thing.

• 19:19Speaker 00

  So basically tried every single UI we could think of in VS Code.

• 19:22Sarah Guo?

  And multiple generations, like the list, that didn't work?

• 19:25Speaker 00

  Yeah, none of them like really worked. I think lists were like, you know, maybe you'd get one generation per person per day. And this was just a small sample. It was just like a few people that were interested in GitHub, language nerds or people that had written tests for us and open app people. Yeah, so very early on I had this idea that should work like a Gmail gray text autocomplete, which was, I was like enamored with that product. It's like, it was the first quote unquote large language model deployment in the wild. It was fast, it was cool. Like, the paper's great. Like, they give you all sorts of details on how they do it, all the, all the workarounds they had to do. So that was always in the back of my head, you know? It was bad also.

• 20:09Speaker 00

  It was like, you know, those completions are not good. But it seemed like the right, the right thing. Anyway, somewhere along the way after, after we tried all the UI, you know, sort of come up with some idea for, VS Code didn't support this, so I tried to hack it in. Finally came up with a way to hack it in and enough to make a demo, like a little demo video.

• 20:28Sarah Guo?

  Was there support to like build real support for it within the organization?

• 20:33Speaker 00

  It's a little complicated. I guess, you know, we're pretty much a Skunkworks project. No one knew about us. So we would go to like the, if you go to VS Code, people would be like, hey, we need you to go implement this very complex feature. We're like, I don't even know who you are. Like, what are you talking about? There was definitely some politicking that happened to, to get the VS Code people to dedicate some resources to that on a short, short timeframe. Like, we were moving really fast. You know, it was less than a year before from beginning to ship, public launch.

• 20:59Sarah Guo?

  Was there a certain metric where you're like, this is good enough? Like, we need to actually put it in the public product?

• 21:08Speaker 00

  Which with the completions or like

• 21:10Sarah Guo?

  Yeah, it was completions. It was.

• 21:13Speaker 00

  Yeah, I say was, I mean, we had a long, we had a nice, long window of public access before GA. So where it was free and you could use it and and we did a bunch of optimizing for different groups of people that would be, you know, okay, well, you know, do we want more experienced people? Do we want more new people who want people from this area or this area? And that gave us a bunch of really good stats. So we were able to learn that, for instance, like, speed is the only thing that matters. Yeah, there's something crazy thing like every 10 milliseconds is 1% fewer completions that people would do. But adds up 10 milliseconds is pretty fast. We learned that because somewhere in our first few months of public release,

• 21:53Speaker 00

  we noticed that Indian completions were really low, like for whatever reason, they were just significantly lower than than Europe.

• 21:59Sarah Guo?

  So network latency to India.

• 22:02Speaker 00

  Yeah, and it turns out because only had one data center, so it was all in Texas. And so it was going, if you can imagine, you're like typing, and so that goes from India through Europe, over the water, down to Texas and back and back and back. And now if you've typed something that doesn't match the thing that you requested, then that's useless, right? So you don't get a completion.

• 22:22Speaker 02

  But by the way, that I know it's obvious, but that's happened on every single product I've ever worked on. You know, like when I was at Google, like I worked on a variety of mobile products, same thing, you know, page load times, obviously, search in general, 100 millisecond difference, like,

• 22:34Speaker 00

  It's a big deal.

• 22:34Speaker 02

  Makes a big shift in market share. So yeah, it's kind of nuts how much speed matters.

• 22:38Speaker 00

  Yeah. And so we, once we figured that out, we knew that we had something awesome, right? Like people that were close to Texas were like, this is freaking great. Like, they were like, you know, we had a Slack channel and people were posting in it all the time. And the most fun stuff was like, these people that would pop up and be like, I don't program, but I just like learned how to write this 100 line script that like does this thing that I need. It's like, oh, my God.

• 22:59Sarah Guo?

  I definitely feel like I now speak different languages that I don't actually know the syntax of. It's very exciting.

• 23:04Speaker 00

  Yeah, it turns out these models are really great at finding patterns. So that, once we had the the UI mechanism that worked. So we knew, we knew around something like that. And then it was just this like, we're using as much performance as we can get out. We like, we basically never found the bottom of like, you know, we hit, we made as fast as we could and it was still improving on completions. And then that perpetuated like, okay, okay, I know there's this plan for like Azure to run OpenAI in six months.

• 23:33Speaker 00

  We need you to do that in the next month. So like, can we, let's figure out how to make this happen. So because we wanted to run a bunch of GPUs in Europe so, so we could hit Asia. You know, there was no, there was no other place that we could, we could run them in Europe. We could run them on West Coast. We could run them in Texas at the time. So that was, and Microsoft stepped up there. We got it running. And then pretty much after that, we launched and yeah.

• 23:58Speaker 02

  Were you surprised by the uptake of PostLaunch?

• 24:00Speaker 00

  No, no, I mean, our retention rate was 50%. Like it never went like months later, it was still above 50% by like weekly cohort, which is like insane. Right. And we didn't, we didn't know if people would pay for it. That was one thing. I lobbied pretty hard for going cheap and capturing the market.

• 24:21Sarah Guo?

  How did you guys think about inference costs for this thing at the beginning?

• 24:24Speaker 00

  Oh yeah, we were, our estimates were wildly off, wildly off. Yeah. So we got estimates who were like, you know, it'll be 30 bucks a month for, for your on average. Right. And then once Microsoft was able to like do some, kind of do their Azure infrastructure, we were able to then like fork off little bits so we could do more accurate projections. And there was a bunch of like moments where like, how much is it going to cost? You'd like wait for these results. The first big one was 10 bucks a month. It'll cost 10 bucks a month. And I was like, oh my God, it's so much cheaper than we expected. And then we could optimize it. And that was like, we hadn't even optimized on price yet. Right. And then we optimized on price a bunch and yeah, now it's less than that.

• 25:09Speaker 00

  So like, it was very fortuitous, right? Like we were thinking like, okay, well, maybe it's, maybe it's enterprise only because that's the only people who are going to be willing to pay for this. Like 30 bucks a month is not, it's a lot. And that's like with no margins, right?

• 25:24Speaker 00

  But yeah.

• 25:24Sarah Guo?

  For 40% of your code, it's not a lot.

• 25:26Speaker 00

  Yeah, that's the thing. We know that. We know that now.

• 25:29Speaker 02

  Where do you extrapolate all this, like three to five years out? Are there basically going to be just like agents writing code for us in certain ways? Is it 95% of code is written by co-pilots and you know, humans are kind of directing it? Like what, what do you kind of view the, the world evolving into in the next couple of years? And next couple, I mean, like three to five, not 20.

• 25:49Speaker 00

  Yeah, it's hard. I don't know.

• 25:52Speaker 00

  I think it's hard for me to, you know, it's hard for me to imagine what that world looks like because it's such a shift from like, I have my hands on something and I know that it's right to where we are now, which is I mostly know it's right, or I have a sense that it's right, but I have to, I have to test it and see it run to know that it's right to then just write this and I'll, I'll trust that it's going to work. Like those are pretty crazy transitions, right?

• 26:17Speaker 02

  Yeah.

• 26:18Speaker 00

  They exist, like.

• 26:20Speaker 02

  Sure. They could also imagine like certain ways to do like a code review post some chunk or some, some other sort of quality check to.

• 26:26Speaker 00

  Yeah, I think every, if that's the goal we want to get to as a people, like I think every barrier to that is achievable. So we can code review only this, the dangerous parts or only the confusing parts, or we can do things like train a model on functions before and after changes to say like, okay, this looks like a more polished version of this function would be this, or yeah, we can do things like, you know, just start the very basic, you know, very basic main loop and then add everything piece by piece with tests so that it's, you know, what, what works and what doesn't. And then just have the, the I keep generating what's the logical next feature, like all these things will be figured out. So if, if that's what we want to do, that's what's going to happen.

• 27:09Sarah Guo?

  So what's the, what's the idea behind Minion?

• 27:12Speaker 00

  Yeah, I think I mentioned making bots that do stuff for you. It's a broad topic and I think that's where we see it going. You know, the next few years are in the AI's taking action, not just answering questions or writing copy, but actually helping us in our daily lives. Things like organizing my schedule or booking flights or finding a trip for me to take or doing my taxes or telling me which contacts I haven't talked to in a long time and should reach out to, you know? There's a lot of stuff that we can do by giving AIs access to information and letting them act on that information

• 27:55Speaker 00

  in a controlled way that checks to make sure they were, they were aligned. And yeah, I think that'll be a really fun future. Almost like you can imagine Copilot applied to everyday activities, right? Like Copilot gives you a little bit of help. So I want Minion to give you a little bit of help outside of your code editor. How did you decide to work on Minion specifically? Minion specifically. So basically, I stopped working at Magic. I quit because I couldn't figure out how to hook up the AI to data. I was like, if in order to improve the quality, I need a PhD in math. Like, I don't know what to do now. And it just sort of, the models got better enough where that specific problem seems solvable. Yeah, so the tech got better.

• 28:40Speaker 00

  And so that specific problem is where interacting with the real world broke down in the past. You know, like I said, floods get delayed, prices change, not just like a little bit, all the time. You know, they might not have the seat you want at the concert that you want to go to. And so, you know, AIs are this kind of compression of everything that you in their training set, but they're not a real time mechanism. Anyway, so that was the idea. I was like, okay, well, I think we can, you can work on this old problem of how to make a bot do stuff for people. That's what we want. Let's go make it. I don't know. It's like, maybe I can use the excitement from Copilot to launch into something which is incredibly hard

• 29:24Speaker 00

  and which I believe the technology is around for, if we can figure it out. So yeah, that's it. That's the only reason I've ever gotten startups, actually, is like, okay, I want to do a startup so that I can do a harder startup. You know, or I want to do a project so I can do a harder project.

• 29:37Sarah Guo?

  Is this one sufficiently hard?

• 29:39Speaker 00

  This one's hard, yeah. It's good and hard, I think. Part of that is, there are these fun things that you kind of learn along the way that keep you, that keep you like engaged, right? Like, the thing with code with Copilot is like, it turns out code is pretty special, right? Like, you can run it. So if an AI generates some code and it runs, then you know something about that code that you wouldn't know necessarily the text. And accomplish, like, doing agent-like work on the internet or on apps is, has a lot of these similar kind of properties where it's like, oh, yeah. You can, you can learn something here. We can learn what, we can learn from what people do, or we can see if it's a success or a failure and then learn from that. Or we can optimize what works based on what doesn't work, or what, figure out what's annoying and try to improve it.

• 30:23Speaker 00

  So I think these things all compound in some really interesting way. You know, I think. The goal, right, is, you know, straight out of sci fi, right? You want to make a thing where you say, hey, computer, file my taxes and it does the right thing. You know, and I think we can get there. It's certainly in the next few years, but it's also fun to like, think about how to break these things down. And it turns out breaking down tasks in the same way that humans do. Take a complex task, you figure out, you know, you write a list of things to go through it. Same kind of thing works for AIs as well. So you break them, you break down complex tasks, figure out if there's any information you need. Maybe you have to write some code, maybe you have to ask some questions, maybe you have to query some data. Same thing a human do.

• 31:07Speaker 00

  Humans don't usually write code to do their taxes, but sometimes it's kind of the same thing. You go through a list of your pay stubs, you know, that's executing a for loop. Also, I like data sets that don't exist. So people clicking on stuff on the web is a gigantic data set, which is currently unowned. And so I think that's pretty exciting. I think there's a lot of stuff to learn from that.

• 31:31Sarah Guo?

  So you're hiring. And one thing we've talked about is like, it's kind of a funny thing to try to hire for building products with this new set of technologies. And that like, working in machine learning for the last decade may not help you that much. How do you, how do you think about the people you need?

• 31:49Speaker 00

  Yeah, it's been strange. It seems very bimodal, like very senior people get it and very young people get it. And so I don't know that the middle has really caught up yet or realizes that.

• 32:03Sarah Guo?

  Are we on the other side? Are we in the middle?

• 32:06Speaker 00

  No, I just mean the middle in terms of like, here and there. I just mean the middle in terms of like, it's like.

• 32:11Sarah Guo?

  I'm very young.

• 32:12Speaker 00

  No, it's not an age thing. It's not an age thing. It's more like you tell someone who's very naive and you tell someone who's very experienced, like, I can make magic. Here's how I can do it. And then the very naive person is like, that's awesome. And then the very experienced person is like, ah, maybe, you know? But the response from almost everybody else is that's not possible or I don't see it.

• 32:33Speaker 02

  Yeah, a number of companies I know are trying to build that intuition internally now. And often, this is the first time in a really long time that I've seen certain founders come back and start coding again. You know, they'll have a multi hundred person company and they'll get so enamored by what's happening that they'll just dive in. A number of them have told me that they feel like the, some of their team members just don't have that intuition. For what this can actually do. And so they don't even know what to build or where to start or how interesting or important it is. And so it's almost like this founder mindset is needed to your point in terms of this is a really interesting new set of capabilities and how to actually learn what these are and then how do I apply them? And I think people lack natural intuition for what this does right now.

• 33:13Speaker 00

  It's definitely not intuitive. Yeah, I mean, I would say I have some intuition now and people that are, again, on an experience spectrum have some intuition, but it's not intuitive. There's many kinds of different learners and there's many kinds of different programmers, right? It takes a certain kind of programmer to be comfortable with this idea of like, I don't know what's going to work. Let me try some stuff. It turns out that some of our attributes to the web are product, product making, being able to test stuff, look at the right metrics, find the right metrics. Yeah, those are, those are useful skills, I think, that are usable. But the intuition and the tenacity that it takes to like, this doesn't work at all. What do I do? That's still rare, I think, in that field because there's, there's so much uncertainty that it's easy to go like, I tried 10 things and it didn't work. So like, this is impossible. Like, okay.

• 33:58Sarah Guo?

  I think the natural reaction of somebody looking at something that has 10% performance at the beginning is like, why even bother? We're not going to get there. Right?

• 34:07Speaker 00

  That's a great point. The crazy thing is that these things work at all, right? Not only that, but they scale and they improve the scale, right? Most things break at scale. Almost everything breaks at scale. And that's what, you know, that's what you hire people to deal with, you know? Every doubling, you usually, something breaks, you gotta fix it. It's a pretty crazy thing to think of what kind of emergent properties might still be out there. If we can get these things 10 times bigger. So I'm always, I'm always in favor of people pushing those limits, you know? Like, it doesn't make sense. Like, even, even the best people can't explain what's going to happen when you, you know, go bigger. We built the large Hadron Collider for that reason, too, you know? We didn't,

• 34:50Speaker 00

  I don't know. We think maybe we'll find some stuff. These endeavors are valuable. And like, yeah, I'm grateful to be a Lactorian during this time.

• 34:57Sarah Guo?

  It's a great note to end on. Thank you for joining us for the podcast.

• 35:00Speaker 02

  Yeah, thanks so much.

• 35:01Speaker 00

  Sure.

• 35:04Sarah Guo?

  Thank you for listening to this week's episode of No Priors.

• 35:07Speaker 02

  Follow No Priors for new guests each week and let us know online what you think and who in AI you want to hear from.

• 35:12Sarah Guo?

  You can keep in touch with me and Conviction by following at Serenormous.

• 35:16Speaker 02

  You can follow me on Twitter at @AladGill. Thanks for listening.

• 35:22Sarah Guo?

  No Priors is produced in partnership with Podpeople. Special thanks to our team, Synthel Galdea and Pranav Ruddy, and the production team at Podpeople, Alex Vikmanis, Matt Saab, Amy Michado, Ashton Carter, Danielle Roth, Carter Wogan, and Billy Libby. Also, our parents, our children, the Academy, and tyranny.ml, just your average friendly AGI world government.