The Exchanges

Q How did you think about the fundraising processes with Warp? Was it always just preempted? How do you reflect on them?

A So I have done, I have done the, um, the formal process thing before. I didn't do it with Warp. I did it with the company before Warp, where when we went out to raise a Series A, we like, you know, flew out to Silicon Valley and pitched like, 20 firms in like, a week or whatever. Kind of took the best term sheet and then like, got in bed with investor. And I did not think that process led to an amazing result. And so with Warp, I really tried not to replicate that. Like, The way I did it with Warp was as I was developing the idea, I was talking to a few investors I knew who I'd known for a really long time who just, when I was like, I'm ready to do this idea, we're like, we love the idea. We want to invest. So not every founder is going to have that opportunity, but if you have that opportunity, I think that's way preferable. And then for the subsequent rounds of Warp, which were Dylan Field led the A and And Sequoia led the B. That was all preempted. And so those were...

AI assessment note: “for the subsequent rounds of Warp... That was all preempted.”

Q How did you think about the fundraising processes with Warp? Was it always just preempted? How do you reflect on them?

A So I have done, I have done the, um, the formal process thing before. I didn't do it with Warp. I did it with the company before Warp, where when we went out to raise a Series A, we like, you know, flew out to Silicon Valley and pitched like, 20 firms in like, a week or whatever. Kind of took the best term sheet and then like, got in bed with investor. And I did not think that process led to an amazing result. And so with Warp, I really tried not to replicate that. Like, The way I did it with Warp was as I was developing the idea, I was talking to a few investors I knew who I'd known for a really long time who just, when I was like, I'm ready to do this idea, we're like, we love the idea. We want to invest. So not every founder is going to have that opportunity, but if you have that opportunity, I think that's way preferable. And then for the subsequent rounds of Warp, which were Dylan Field led the A and And Sequoia led the B. That was all preempted. And so those were...

AI assessment note: “for the subsequent rounds of Warp... That was all preempted.”

Q Does it change your business having Sequoia? And I, I don't mean that like, oh, I want VCs to just like, you know, jerk each other off. Um, but like, but like when you have Sequoia money, does life get easier?

A A little. Here's what, here's what Sequoia, like, I'm a huge Sequoia fan. Here's what Sequoia has done well for us. So there's like a halo around the brand, right? So like from a recruiting perspective, it's good. From a like customer conversation perspective, it's sometimes good. I'll give you like a really tangible example. We were getting user reports that were being blocked by CrowdStrike. So CrowdStrike is like this security software, right? Which kind of makes sense. Like we're at that layer of the stack. Like we do all this crazy stuff to make work work on people's computers. So we're like, oh fuck, we're getting blocked by CrowdStrike. That's like really, really bad. And it's like, we're, we saw all of our uses start to go down at certain companies. And so I reached out to Andrew. I was like, can you help me here? And then that day I was on the phone with the president of CrowdStrike. And he's like, I'm looking into it. I don't think that we've actually done anything. I think this is a configuration issue. Cause we, we ended up being Blocked at Salesforce of all places. Uh, you know, that, I don't think you get that if you, if you have like your kind of random VC firm, but the connections there are real. It's freaking cool, right? That's like, that's a cheat code. And I, I was, I was telling my, I told my wife that and she was like, oh, that, that is like pretty cool. S…

AI assessment note: “A little. Here's what, here's what Sequoia, like, I'm a huge Sequoia fan.”

Q Would you say agentic coding tools have product market fit? I know that sounds utterly ridiculous when you look at Cognition's ten billion dollar price and Cursor's twenty-eight billion dollar price, and you'd say, of course they do, you moron. But if you have customers that are genuinely questioning productivity gains and kind of going, ah, do they?

A So great question. In like the prosumer market, there's like strong product market fit, but it's not, I would say it's not a great market. Meaning like they're very high churn, price sensitive customers who are building things that don't have a ton of economic value. In the enterprise market, I actually think the thing that has most product market fit is like autocomplete. Which is not the sexiest type of AI. So it's not agentic AI. It's like cursor's original product or even co-pilot. If you're doing like hand editing of code, it's like clearly helpful to have the thing like show you the gray text and kind of read your mind. When you get into like warp or cloud code or cursor's agent or yeah, even I think cognition's its own beast. Honestly, it's like a, it's more of like an automation thing. I think it's like really early in that market and people, there's like a lot of like Intuitive signal that it's helpful, and there's a ton of willingness to pay, but in terms of, like, will companies get measurable impact from it? It's gonna happen, like, it's gonna be transformational, and it's sometimes is transformational now, but not with every coding test, and maybe not in the aggregate.

AI assessment note: “I think it's like really early in that market”

Q Do you think it's open really? I can see the AWS, Azure, Google Cloud trio. I can see three or four players going and taking dominant share. I don't think open's it.

A I don't think open either, if I'm being honest. The reason I don't think open, like, is that it's just so damn expensive to build these things, and so what's the economic incentive? It's not like open source software. People, like, make the analogy, like, oh, open. I'll take the word open and apply it to open models. Open source software works because It's a bunch of, like, hobbyist developers who are giving their time to build something really hard. Open models, somebody's got to spend all the money to, like, make the thing competitive. But here's the flip side. Here's the way it could work is, like, it could get to a point where just, like, for coding, let's say our domain, it's basically solved. Meaning, like, models that are good enough and you don't need the frontier model And what actually starts to matter is like, how good are you getting context in from a company? How good is your interface? How good is your automation, like an orchestration stuff? And that's not crazy to me. Like, it's crazy to me that the frontier model would be open, but it's not crazy to me that there's like a good enough model that is open.

AI assessment note: “I don't think open either, if I'm being honest.”

Q What does that mean, learn how to use them right?

A Yeah, so if you're like, um, someone building in a sort of naive way, and I kind of blame vibe coding for this, like, what you'll do is, uh, you'll be like, build me this thing that looks like X, and the build me this thing that looks like X doesn't tell the agent how the thing should work. At least from like the engineering side. And so if you do this on like a big complicated code base, it fails, or like it produces something that just like wastes a bunch of your time. It might like iterate and go in circles for hours and finally get you something that like seems to work from a user perspective, but it won't be something that you will actually want to release. So to make it work in a production setting, it's like you need to tell it how to build it. And that means you need to understand the code. And so it's just, it's very liable to, like, kind of lazy usage, which doesn't work well.

AI assessment note: “you need to tell it how to build it. And that means you need to understand”

Q Do you like it? Like, help me understand, because normally, I'm on the other end of this, ok. Do you like it, or are you like, oh, fucking another VC, yeah, yeah, you've got a thesis on the space, sure you do.

A So I think it would be like kind of arrogant to be like, oh, I hate this. Like, it's like, no, it's, it's cool. People, you're running a business and people are interested in giving you money. I think that is not the default case for most people in the world. And so like for me to be like, oh, I don't, that's so annoying. No, it's great. It's, it's awesome to have the interest from a, like how to manage it as a founder. I just basically don't take the calls. Right now. I'll do them once in a while to just like gauge like the market and like where we're at. But I have a strong feeling that like when we do it, I'm not going to do it based on an inbound email. I'm going to do it based on like either existing investors or they're going to introduce me to someone who I, like I'm optimizing for the relationship. Right. And so it's hard to get that from just like a sales pitch over an email.

AI assessment note: “No, it's great. It's, it's awesome to have the interest”

Q Does it make one X devs or engineers 10 X, or does it make 10 X a hundred X? Like, who does it favor? The lower quality or the higher quality?

A Favor is a higher quality. And this is counterintuitive. Like it actually causes a lot of problems with people who don't know what they're doing in a professional environment. So if you are a newer engineer or more junior engineer, or maybe just like not that sophisticated of an engineer, it's a problem to me because you want to use these tools the most, but they end up producing code that you don't understand or that can't be shipped or they won't get through code review or that contains security bugs. Whereas To successfully use these tools, you need to be, like, more sophisticated, smarter, more experienced, and if you can do that, you can get a crazy productivity gain, but you already kind of need to know what you're doing. So I think it's a little bit counterintuitive. Like, I would much rather have senior developers using these tools. There's a little bit of a paradox in that they don't necessarily want to use these tools. Like, they feel like they can do the thing themselves, but if you get a senior developer who's really good at wielding these tools, it's very powerful.

AI assessment note: “Favor is a higher quality. And this is counterintuitive.”

Q When you compare Gemini, Claude, and GPT, how do they compare for you internally today?

A The two leaders are GPT-V and the Claude models, and so we measure them. Like, first of all, it's like, it's not so much gut. It's like we literally We, we have a whole bunch of benchmarks. We run sweet bench. We run terminal bench. Claude and GPT-V are the two best. Claude, they have like different personalities, which is kind of interesting. Like when you're using them for coding, like Claude's a little friendlier. GPT-V is like a little bit more like, let me do a whole bunch of thinking. Take, it takes a long time is my, my biggest issue with it, but then it tends to come back with something that's really good. And at the moment, Gemini, which is on two five, and I'm sure they're working on Gemini three is not as good.

AI assessment note: “Claude and GPT-V are the two best... Gemini... is not as good.”

Q about the biased nature of your opinion. Most people, Especially people like me, VCs aren't. We just state these views. They're hoping no one knows that we have massive positions in these companies. Um, but when we actually think about what that market looks like in five years time from the enterprise developer market, can you help me understand what you think it will look like? Maybe putting biases aside.

A Yeah. So just in the development space, I think the market is going to be, there's going to be two pieces to it. So there's going to be like a set of interactive productivity tools. So think of like things that developers literally are using. You're sitting there, you're telling an agent what to do. You're like, Hey, write this code for me, help me understand this production problem. And then there's going to be a set of automation tools. That's going to be more like, okay, crash came in. What was the root cause of it? Is there a PR that can fix it? And it's going to do it automatically. And you'll have some developer who's reviewing it. Those are going to be the two, the two classes of things. I think over time, the automation is clearly like a better market, the more valuable thing. Companies are going to want to like just pay for stuff that automatically creates software. It's a huge market. It goes into the labor spend a little bit, but I think the productivity piece is also, I don't see it being, I'm just guessing at timeframes here. Five years is a long time with these AI models, just to be clear. But in the next few years is what I see happening. So those are the two big The two big buckets, and I think more and more will go into automation.

AI assessment note: “I think the market is going to be, there's going to be two pieces to it.”

Q Okay, it's not a productivity tool, but we do look at productivity gains as a measure of effectiveness. We are x more percent productive. We create x percent more code in less time, whatever, whatever. Are we actually seeing productivity increases in terms of output that are meaningful and measurable, do you think?

A In terms of whether we're seeing productivity gains, I think no one knows. Like, there's a bunch of studies that show not really. Like, again, I would, I would be like a little more nuanced. So if you are using lovable, then yeah, of course, like you're going from like, I can't do anything to I've built an app. That's like insane. And so the cases where it's like zero to one, yes. I think when you go into the professional development environment, which is like, Where warp is focused. There's a bunch of studies that show, not really. It's sort of like the noise that's created by people who are attempting to use like a vibe coding technique on a production code base kind of outweighs or can slow you down. And so I think that's the big problem with it right now. And this is what we see when we talk to companies. It's like, they all kind of rush to deploy some sort of a genetic coding solution. No one really knows if it's working. And whether it's working or not depends a ton on, like, how you use it, and so you can't just vibe code with it. I'm not a huge vibe coding person to begin with, and so there's, like, a way you have to use it in order to get productivity gains, and then it needs to be figured out how you're gonna measure those.

AI assessment note: “In terms of whether we're seeing productivity gains, I think no one knows.”

Q Does it make one X devs or engineers 10 X, or does it make 10 X a hundred X? Like, who does it favor? The lower quality or the higher quality?

A Favor is a higher quality. And this is counterintuitive. Like it actually causes a lot of problems with people who don't know what they're doing in a professional environment. So if you are a newer engineer or more junior engineer, or maybe just like not that sophisticated of an engineer, it's a problem to me because you want to use these tools the most, but they end up producing code that you don't understand or that can't be shipped or they won't get through code review or that contains security bugs. Whereas To successfully use these tools, you need to be, like, more sophisticated, smarter, more experienced, and if you can do that, you can get a crazy productivity gain, but you already kind of need to know what you're doing. So I think it's a little bit counterintuitive. Like, I would much rather have senior developers using these tools. There's a little bit of a paradox in that they don't necessarily want to use these tools. Like, they feel like they can do the thing themselves, but if you get a senior developer who's really good at wielding these tools, it's very powerful.

AI assessment note: “Favor is a higher quality. And this is counterintuitive.”

Q Do we see the collapsing of roles? When we look across the dev spectrum now, you know, you've got products, you've got ends, you've got design. Do we see the collapsing entities like superhuman product builders? Or do functions remain intact?

A In terms of collapsing of all the roles, if I had to, like, design a role for the Uber person who builds with AI, it would be, like, a product-oriented senior engineer. We see a lot of, like, product managers and designers who will use the tools to make prototypes, and that's actually very cool, but to get stuff to production, I think what you want is, like, someone who, like I said, more sophisticated than the agent, and I think there's a real cap on what A product manager or designer is able to build with this stuff right now. So I guess short answer is I don't really think so. And if there was like a role I would design, it would be a product minded engineer.

AI assessment note: “So I guess short answer is I don't really think so.”

Q For SaaS businesses, it changes everything. What do you mean by that?

A I think what I mean is like, if you're working at, there's a few different angles. There's the business angle, there's the worker angle. I think the cost to start, A sort of software driven business is going to go way, way, way down. That's one thing. I think from a worker perspective, if you are working at a place like, I'm going to pick a random company, Salesforce, and there's no real regulatory blocker for the deployment of AI and you're in like a competitive market, AI is going to be deployed really, really rapidly and change what daily work looks like. I think if you're in like the healthcare industry or the government, and it's like, I'm still filling out paper forms when I go see my doctor right now, it's less a question of like, is the technology transformational than like, are the incentives there to deploy the technology? And like, I worry that it's just gonna take a long time.

AI assessment note: “There's the business angle, there's the worker angle.”

Q If you don't rewrite, do you not just do sticky tapes with technical debt though? Like I'd say rewrite and rewrite early, like make a decision, make it fast and go.

A So for the startup people out there, I think try to pick the right thing to start is like a little bit better advice. So one lesson I did take for this is for warp. For instance, we built it the hard way to start. Meaning like, you know, I could kind of see around the corner, like if we built it on sort of web tech or whatever, we're going to run into a bunch of performance problems, which is a lot of what we were trying to fix with the, the sheets rewrite, like good engineering matters, but I think it's highly dependent on like where you're at in the stage of building a thing. And so, yeah, rewrite, if you have like crazy product market fit and, you know, a hundred million plus users, it needs to be perfect. If you're really early, I think rewrite is a little scary.

AI assessment note: “I think it's highly dependent on like where you're at in the stage”

Q When you look at your devs today internally, what are they using? Are they using code code? Are they using cursor? Are they using cognition?

A They use warp. We're now fully into like the coding agent space, like history of the company. Obviously we started as a, as a terminal, but the terminal form factor is the preferred form factor for doing agentic work at this point, whether you're using like a terminal app, like cloud code, Uh, in warp, you just literally use the GUI app to do it. And so every single coding task we do starts with a prompt, which is the big shift in productivity that's happening right now. We don't necessarily finish every task with a prompt. There's certain things that are too hard. It doesn't matter which, which tool you're using, but everything starts with the prompt. Then we try to get all the way to completion if we can with prompt.

AI assessment note: “They use warp. We're now fully into like the coding agent space”

Q Would you say agentic coding tools have product market fit? I know that sounds utterly ridiculous when you look at Cognition's ten billion dollar price and Cursor's twenty-eight billion dollar price, and you'd say, of course they do, you moron. But if you have customers that are genuinely questioning productivity gains and kind of going, ah, do they?

A So great question. In like the prosumer market, there's like strong product market fit, but it's not, I would say it's not a great market. Meaning like they're very high churn, price sensitive customers who are building things that don't have a ton of economic value. In the enterprise market, I actually think the thing that has most product market fit is like autocomplete. Which is not the sexiest type of AI. So it's not agentic AI. It's like cursor's original product or even co-pilot. If you're doing like hand editing of code, it's like clearly helpful to have the thing like show you the gray text and kind of read your mind. When you get into like warp or cloud code or cursor's agent or yeah, even I think cognition's its own beast. Honestly, it's like a, it's more of like an automation thing. I think it's like really early in that market and people, there's like a lot of like Intuitive signal that it's helpful, and there's a ton of willingness to pay, but in terms of, like, will companies get measurable impact from it? It's gonna happen, like, it's gonna be transformational, and it's sometimes is transformational now, but not with every coding test, and maybe not in the aggregate.

AI assessment note: “In like the prosumer market, there's like strong product market fit”

Q What does that mean, learn how to use them right?

A Yeah, so if you're like, um, someone building in a sort of naive way, and I kind of blame vibe coding for this, like, what you'll do is, uh, you'll be like, build me this thing that looks like X, and the build me this thing that looks like X doesn't tell the agent how the thing should work. At least from like the engineering side. And so if you do this on like a big complicated code base, it fails, or like it produces something that just like wastes a bunch of your time. It might like iterate and go in circles for hours and finally get you something that like seems to work from a user perspective, but it won't be something that you will actually want to release. So to make it work in a production setting, it's like you need to tell it how to build it. And that means you need to understand the code. And so it's just, it's very liable to, like, kind of lazy usage, which doesn't work well.

AI assessment note: “you need to tell it how to build it. And that means you need to”

Q Do you think it's open really? I can see the AWS, Azure, Google Cloud trio. I can see three or four players going and taking dominant share. I don't think open's it.

A I don't think open either, if I'm being honest. The reason I don't think open, like, is that it's just so damn expensive to build these things, and so what's the economic incentive? It's not like open source software. People, like, make the analogy, like, oh, open. I'll take the word open and apply it to open models. Open source software works because It's a bunch of, like, hobbyist developers who are giving their time to build something really hard. Open models, somebody's got to spend all the money to, like, make the thing competitive. But here's the flip side. Here's the way it could work is, like, it could get to a point where just, like, for coding, let's say our domain, it's basically solved. Meaning, like, models that are good enough and you don't need the frontier model And what actually starts to matter is like, how good are you getting context in from a company? How good is your interface? How good is your automation, like an orchestration stuff? And that's not crazy to me. Like, it's crazy to me that the frontier model would be open, but it's not crazy to me that there's like a good enough model that is open.

AI assessment note: “I don't think open either, if I'm being honest.”

Q OpenAI at five hundred billion or Anthropic at one 83. Which do you invest in?

A I've said it before in this interview. I'll say it again. I'm very, very bullish on open AI because I think they have a winning consumer product. Anthropic is maybe the best AI lab, like from a research perspective and from like the model, like their models are incredible. I think it's a big problem not to have the consumer product there. And so I don't know is an API business, a gigantic business. So they're going to have to really win on these like enterprise vertical businesses and maybe they will. I hope not. Like We're competing with them as a frustrating thing as someone in the coding space and are a customer of them, but I, I think OpenAI with ChatGPT is going to be gigantic.

AI assessment note: “I'm very, very bullish on open AI because I think they have a winning consumer product.”

Q When you look at your devs today internally, what are they using? Are they using code code? Are they using cursor? Are they using cognition?

A They use warp. We're now fully into like the coding agent space, like history of the company. Obviously we started as a, as a terminal, but the terminal form factor is the preferred form factor for doing agentic work at this point, whether you're using like a terminal app, like cloud code, Uh, in warp, you just literally use the GUI app to do it. And so every single coding task we do starts with a prompt, which is the big shift in productivity that's happening right now. We don't necessarily finish every task with a prompt. There's certain things that are too hard. It doesn't matter which, which tool you're using, but everything starts with the prompt. Then we try to get all the way to completion if we can with prompt.

AI assessment note: “They use warp.”

Q For SaaS businesses, it changes everything. What do you mean by that?

A I think what I mean is like, if you're working at, there's a few different angles. There's the business angle, there's the worker angle. I think the cost to start, A sort of software driven business is going to go way, way, way down. That's one thing. I think from a worker perspective, if you are working at a place like, I'm going to pick a random company, Salesforce, and there's no real regulatory blocker for the deployment of AI and you're in like a competitive market, AI is going to be deployed really, really rapidly and change what daily work looks like. I think if you're in like the healthcare industry or the government, and it's like, I'm still filling out paper forms when I go see my doctor right now, it's less a question of like, is the technology transformational than like, are the incentives there to deploy the technology? And like, I worry that it's just gonna take a long time.

AI assessment note: “cost to start, A sort of software driven business is going to go way, way, way down.”

Q When you compare Gemini, Claude, and GPT, how do they compare for you internally today?

A The two leaders are GPT-V and the Claude models, and so we measure them. Like, first of all, it's like, it's not so much gut. It's like we literally We, we have a whole bunch of benchmarks. We run sweet bench. We run terminal bench. Claude and GPT-V are the two best. Claude, they have like different personalities, which is kind of interesting. Like when you're using them for coding, like Claude's a little friendlier. GPT-V is like a little bit more like, let me do a whole bunch of thinking. Take, it takes a long time is my, my biggest issue with it, but then it tends to come back with something that's really good. And at the moment, Gemini, which is on two five, and I'm sure they're working on Gemini three is not as good.

AI assessment note: “Claude and GPT-V are the two best... Gemini... is not as good.”

Q If you don't rewrite, do you not just do sticky tapes with technical debt though? Like I'd say rewrite and rewrite early, like make a decision, make it fast and go.

A So for the startup people out there, I think try to pick the right thing to start is like a little bit better advice. So one lesson I did take for this is for warp. For instance, we built it the hard way to start. Meaning like, you know, I could kind of see around the corner, like if we built it on sort of web tech or whatever, we're going to run into a bunch of performance problems, which is a lot of what we were trying to fix with the, the sheets rewrite, like good engineering matters, but I think it's highly dependent on like where you're at in the stage of building a thing. And so, yeah, rewrite, if you have like crazy product market fit and, you know, a hundred million plus users, it needs to be perfect. If you're really early, I think rewrite is a little scary.

AI assessment note: “If you're really early, I think rewrite is a little scary.”

Q And can I ask you, end state, what do the margin profiles of these businesses look like? Are these, like, 15 to 20% margin businesses end state? Are they, like, 60, 65?

A End state's not obvious because it totally depends what happens in the model market. Somewhere there's gonna be a margin, right? Like, Anthropic has a big margin on their API business right now. What I've heard is 60% margin on Anthropic's API. I've heard 50% on OpenAI. Again, I've heard this. I don't know if it's exactly right or not, but if there were no margin anywhere in this stack, I think it's really bad, but there's margin. The question then is like, well, who's gonna, who's capturing that? And I think it depends. If the model providers are really competitive, Even better would be like some open source model that's like good enough. I think the app layer is going to get a lot of value. If you think that it's going to be like one model provider runs away with it, then that's like horrible for us. I don't think that's going to happen. I think that's, there's too many smart people who are working on building these models. So our bet is like, there's competition. It might look like G cloud, AWS, Azure level competition, or it might look like actually like open source, like database competition. And then we're going to have an awesome business, but I don't know.

AI assessment note: “End state's not obvious because it totally depends what happens in the model market.”

Q was my favorite of all the answers. Of all the show, of all the show. That was the one. Final one, what are you most excited for? So like, when I think about my mother's got multiple cirrhosis, I think that we will have incredible cures to some of the world's most challenging diseases in 10 or 15 years, which, it excites me immensely. What do you find most exciting?

A Yeah, it's, it's that type of thing. Just to be honest, like, I'm getting older. I think we're at this really interesting spot where we're on the precipice of like, Amazing, crazy technology that truly could like help a ton with healthcare or help a ton with people's prosperity or learning or whatever. And on the flip side, I think we're at this very precarious, like political spot. And I feel like those two things are like racing each other is super unstable. We're not like a spot of equilibrium. My hope is that the technology, like the advances in technology can really, really have a big impact. I think if I had to pick one area where I'm most interested, most Hopeful as well is, like, in health. Like, that would truly be transformative for, like, you know, me and my family and just, like, having some sort of real medical progress, but I'm also scared because it's, like, it's, like, a very, like, unstable situation right now in the world, in my opinion.

AI assessment note: “if I had to pick one area where I'm most interested, most Hopeful as well”

Q Do you think prompt is a sustaining input to get desired output? Like in three years time, will we still as users be prompting?

A That's funny. It sounds kind of stupid. It's like, oh, start with a prompt. Um, the general thing is like, are you starting with some expression of human intent? And like, I'm using prompt for shorthand of that. And like, it's more than just you're writing some paragraph or something. It's like, what's going to matter more and more is the total context that the model has available to it. And that context is going to come from more and more places. It's not just going to come from like a person saying to do something. But yeah, I think, like, generally, if you're doing something creative and it's not totally deterministic what needs to be done, there needs to be some expression to the computer from a builder to do the thing. It's not going to read your mind. So yeah, I think it's like, even though it sounds kind of stupid, I think that is the thing.

AI assessment note: “So yeah, I think it's like, even though it sounds kind of stupid, I think that is the thing.”

Q Do you think prompt is a sustaining input to get desired output? Like in three years time, will we still as users be prompting?

A That's funny. It sounds kind of stupid. It's like, oh, start with a prompt. Um, the general thing is like, are you starting with some expression of human intent? And like, I'm using prompt for shorthand of that. And like, it's more than just you're writing some paragraph or something. It's like, what's going to matter more and more is the total context that the model has available to it. And that context is going to come from more and more places. It's not just going to come from like a person saying to do something. But yeah, I think, like, generally, if you're doing something creative and it's not totally deterministic what needs to be done, there needs to be some expression to the computer from a builder to do the thing. It's not going to read your mind. So yeah, I think it's like, even though it sounds kind of stupid, I think that is the thing.

AI assessment note: “It's not going to read your mind. So yeah, I think... that is the thing.”