How to Win With Prompt Engineering - Ep. 38 with Jared Zoneraich

Prompt engineering isn’t just about telling AI to solve your problems—it’s about knowing which ones to solve. Yet there’s a mismatch between the people who can identify the right problems—experts with deep domain knowledge—and the technical infrastructure required for developing and refining prompts. Jared Zoneraich, the cofounder and CEO of prompt engineering platform PromptLayer, is bridging the gap with a platform on which non-technical experts can manage, deploy, and evaluate prompts quickly. The role of human prompt engineers, however, has been the topic of controversy, with some arguing that AI can optimize prompts better than us, while others suggest that more capable LLMs eliminate the need for meticulously crafted prompts altogether. I spent an hour talking to Jared about why he believes prompt engineering isn’t becoming obsolete. He also tells me everything he’s learned about writing a good prompt and what the future of AI tools looks like. Here is a link to the episode transcript. This is a must-watch for prompt engineers, people interested in building with AI systems, or anyone who wants to generate predictably good responses from LLMs. If you found this episode interesting, please like, subscribe, comment, and share! Want even more? Sign up for Every to unlock our ultimate guide to prompting ChatGPT. It’s usually only for paying subscribers, but you can get it here for free. To hear more from Dan Shipper: Subscribe to Every: https://every.to/subscribe Follow him on X: https://twitter.com/danshipper Links to resources mentioned in the episode: Jared Zoneraich: @imjaredz PromptLayer: @promptlayer, https://www.promptlayer.com/ A couple of Steven Wolfram’s articles on ChatGPT: What Is ChatGPT Doing … and Why Does It Work?, ChatGPT Gets Its “Wolfram Superpowers”!

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

• 0:00Speaker 00

  I really don't like these research papers about prompting. People love sharing them, oh, this research paper came out of this new prompting strategy. Prompting doesn't feel like that sort of science. It's more of, I'm trying a bunch of stuff and as long as I have a good data set and a good framework to iterate and see if it worked, that's what I should be focusing on.

• 0:17Dan Shipper?

  What's the gnarliest prompt you've run into?

• 0:19Speaker 00

  I'm gonna be honest, it's the ChatGPT or Anthropic System prompts. Those are bad.

• 0:24Dan Shipper?

  Wow, hot take.

• 0:26Speaker 00

  You're not going to build defensibility through machine learning for most companies. You're going to win by working with the domain experts who can write down that problem, as we were talking about earlier, who could define the specs of what you're solving. The best prompt engineer is created as a black box and say, the L M is getting more complicated, not less. Let's not think about how it works. All I want to think about is how do I map the inputs to the outputs I want? And if you're not getting the outputs you want, that's a skill issue.

• 1:05Dan Shipper?

  Jared, welcome to the show.

• 1:06Speaker 00

  Thanks for having me.

• 1:08Dan Shipper?

  So for people who don't know, you are the co-founder and CEO of PromptLayer. And how do you describe PromptLayer these days?

• 1:14Speaker 00

  Yeah, we are a prompt engineering platform. We came out about a year ago, a year and a half ago, actually. Whenever ChatGPT came out, not too long ago.

• 1:21Dan Shipper?

  I remember that.

• 1:24Speaker 00

  We, it was an internal tool originally. We just put it on Twitter and people liked it. And we're like, hmm, maybe we'll do this.

• 1:31Dan Shipper?

  Kind of like early Slack vibes a little bit.

• 1:34Speaker 00

  Yeah, exactly, exactly.

• 1:37Dan Shipper?

  So I guess like you've been around for a year and a half, so the prompt engineer is not dead yet. Have reports of the death of the prompt engineer been exaggerated or where are we in the timeline?

• 1:46Speaker 00

  I think so. I think so. This is a, this is a common question we hear a lot. Is the prompt engineering going to be automated? What's going to happen here? I'll tell you my take on it. We, so, so we're focused on building the prompt engineering workflow. How do you iterate on these things? How do you make the source code of the L M application better? There's three primitives we see here. The prompt, the eval, and the data set.

• 2:14Speaker 00

  You can automate the prompt, but you still have to build the eval data set. You can automate the data set. You've got, you need the eval prompt. You could take out one of this, one of these elements from the triangle. But at the end of the day, our core thesis, our core theory is that prompt engineering is just about putting domain knowledge into your LLM system. And whether you have to say please and thank you to the AI is, uh, that'll probably go away, but you still need to iterate on this core source code, we could call it.

• 2:43Dan Shipper?

  Okay. So you're basically like, probably you're not going to have to be like, hey, like I'll tip you $5,000 and like in, in a couple of years, like that thing is going to go away. But like the sort of core primitives that you're talking about. So the prompt, the eval and the data set. Can you make that concrete for me? Like, what do you mean by those things? Give me a specific example.

• 2:59Speaker 00

  Yeah, yeah, let's talk about an example. So let's say I'm building a AI math tutor.

• 3:10Speaker 00

  Pre-AI, there's a lot of math tutors you can go to. Some are good, some are bad. There's almost infinite solutions to this problem. So there's never going to be a one prompt that rules it all where you say, build an AI math tutor and it makes you the only solution. You can imagine 30 companies competing for this. So when I was talking about those primitives, when you're building this, you have the prompt or multiple prompts on how you respond to each question the student asks. You have the ways to test that and the sample data you're testing that on. But all those are a different, like kind of a more technical way to say you have the actual knowledge the math tutor has.

• 3:54Dan Shipper?

  That's interesting. You're, you're making me think about like, um,

• 4:00Dan Shipper?

  I love the kind of like, there's no one prompt to rule them all. There's no one data set to rule them all, and even in a particular domain. And the thing it made me think of is,

• 4:12Dan Shipper?

  Do you know the philosopher Leibniz? I'm sorry to take it to Leibniz.

• 4:16Speaker 00

  No.

• 4:17Dan Shipper?

  Okay, so he's like around Newton's time. He also invented calculus. You know, they fought over that. But one of the things that Leibniz was like really into is

• 4:29Dan Shipper?

  the idea of creating a language where you could only say true things. So the syntax of the language made it so that like only you can only say factual true things. And so if people had a difference of opinion, they would just say like, write it in the language and then calculate. And yet the answer, just like if you have a question about like, you know, how far is the cannonball going to travel? You can use calculus to figure that out. Same thing for like statements of fact or statements about the world with this hypothetical language. And obviously he never figured it out.

• 5:02Dan Shipper?

  And the reason is because

• 5:05Dan Shipper?

  when you try to like produce something that's so

• 5:10Dan Shipper?

  all encompassing or totalizing of truth, like it's, it gets really brittle. It gets really hard. Like a lot of early AI attempts were sort of like that. And I think there's a thinking error in a lot of what people think of when they think of, okay, what's the future of AI? Like, oh, I'm just going to tell it to be a math tutor. It's going to be the best possible math tutor. It's the same kind of thing. It's like, it's going to be, and it's like, what that means is incredibly different and incredible in many different contexts. And, and so there's a lot of room for many different

• 5:40Dan Shipper?

  attempts at finding the truth or representing the truth of what is best in any given scenario. And I think you're kind of, you're, you're getting at that with your, with, with the answer of like, yeah, we're going to have to do prompt engineering because like, there's going to be 30 different ways to make a good math tutor. And like, who knows which one's going to be the right one for which person?

• 5:59Speaker 00

  Yeah, no, I mean, forget about the prompt. In this world of we have the best API possible. I like that reference a lot because it's the left. How do you even define the problem to solve? That's the hard part. Let's assume you had millions of dollars. You had, you just got a big VC round to start a new company. And what would you build knowing AGI is going to come? The hard part is the problem. You would start working on what problem are you going to solve? Because even with the best tool, you need to define the exact scope of the problem. And that's kind of the irreducible part.

• 6:35Dan Shipper?

  That's really interesting. Yeah. And I think we're so used to building things being expensive that finding the problem is like this invisible thing that like everyone has like, I don't know if this is your first company. I assume it's not like every is not my first company. You spend like frigging years trying to figure out what to solve and the building is expensive and hard, but like the finding the right problem is is quite a bit harder. And I guess you can like, you could sort of like posit, okay, uh, the AGI is going to figure out the problem for you to solve, but like, I don't know, that's, that requires a lot of uh, real world interaction and creativity and a viewpoint like there's a, I don't know. It's, that feels quite a bit more complicated than building software.

• 7:19Speaker 00

  Well, this is my first company, actually.

• 7:21Dan Shipper?

  Oh, it is? Oh, congratulations. You're doing great.

• 7:24Speaker 00

  We had a slight product before this that uh, probably was, you know, it was like our pivot. But uh, yeah, I like this simple example. I like to think about if I wanted to build an AI secretary that booked me flights, uh, and I have a conference in Japan. There's a lot of different possible solutions. Do I want an aisle seat? Do I want a window seat? Do I rather a non-stop over a business class with a stop? And these are the differences that mean there's no one solution. These are the, I love the word irreducible here and it, I still have from uh, if like Stephen Wolfram wrote a few articles on ChatGPT.

• 8:02Dan Shipper?

  Like computational irreducibility?

• 8:04Speaker 00

  Yeah, yeah, I love that. I love that theory. And I think the irreducible part.

• 8:07Dan Shipper?

  Can you define that for people who are listening and haven't heard that before?

• 8:11Speaker 00

  Totally. So I, and I'm not going to do as good of a job at defining it. So I encourage everyone to look it up and read the original.

• 8:17Dan Shipper?

  Use ChatGPT to look it up if you're, if you're listening to this or watching this.

• 8:23Speaker 00

  So I would define computational irreducibility as there's.

• 8:29Speaker 00

  When you're solving a problem, you can collapse a lot of parts and speed up a lot of parts of solving the problem. But there's one part that you'll never be able to collapse. I like to think of when in school in math class, you're, was it factoring, right? When you're kind of taking things out into the parentheses. But there's always that irreducible part that you can't factorize or you can't simplify. And in this case, we're saying if you have an amazing API that can solve any problem, the hard part is what do you even tell it to solve?

• 9:04Dan Shipper?

  Yeah, I think another way to think about computational irreducibility is if you imagine the world to be like a big computer or the universe to be a big computer, there are certain operations that like you actually have to run the computer to find out what happens. You have to like run out the reality of the system in order to know. There's no like, you don't have calculus to be like, okay, I know where it's going to be in 10 steps because like the universe doesn't know yet. And I think that's super cool.

• 9:29Speaker 00

  Yeah, haunting problem-esque.

• 9:30Dan Shipper?

  Yeah. Okay. So you kind of have this like pluralistic framework of the future of AI. How did you come to that?

• 9:36Speaker 00

  Well, it depends. I guess what are you referring to as pluralistic? Like multiple models or multiple stakeholders?

• 9:43Dan Shipper?

  Multiple approaches to any given problem.

• 9:45Speaker 00

  Yeah, I think it's just just observation. Honestly, I'm a hacker. That's my background. I'm not necessarily a researcher and not necessarily, to be honest, and this might be a controversial thing. AI kind of doesn't interest me, whether AI is going to take over.

• 10:04Dan Shipper?

  I love that. You're the first person to say that on this show. Hot take alert.

• 10:09Speaker 00

  It is. And I get into, you know, we have a lot of debates in it. We actually, when we interview someone, we usually take them to lunch and ask them what they think about AGI, just curiously, if you could get their takes.

• 10:20Dan Shipper?

  What's the bad response?

• 10:22Speaker 00

  There is no bad response.

• 10:23Dan Shipper?

  There's no bad response. So why do you ask it?

• 10:25Speaker 00

  Curious what they think. This, we have a whole different, lot of different ideas. But what I'm most interested in, not so much is AI going to take over the world? Is it going to kill us all? I would just say like, how do you build with it? I think it's just a really cool technology. And we are, as a hacker, I am like, my mental model of AI is not.

• 10:49Speaker 00

  The reasoning part as much as it's self-language to data and data to language. And how much does that open for us? So that's how I get here. It's just, it's a tool in the toolbox. We have a lot of tools in the toolbox. And what can you solve? In the same way that like my non-tech friends will ask me, why does Facebook have engineers? The site works for me, right? Why do they have to hire more engineers? I think it's the same thing. I look at AI the same way. You're going to always be iterating. There's so many degrees of flexibility in these things.

• 11:21Dan Shipper?

  Yeah. You said earlier before we, before we started recording that one of the things you're kind of excited about or thinking about a lot is the non-technical prompt engineer. Can you just open that up for me? Like, what, what does that mean? And how did you like start, start knowing that that was a thing you wanted to pay attention to?

• 11:39Speaker 00

  Yeah, yeah. So, so when we launched the prompt layer, as I was saying, it was kind of an internal tool. We built for ourselves and people started liking it. So we didn't have this idea when we started, but I'll tell you, there was kind of a light bulb moment I look back on. It's this team parent lab. They're building a very cool app. It's basically an AI parenting coach to help you be a better parent. We got on a call with them, just feedback on our product about a year ago and asked them, how do you like prompt layer? What can you improve? But the most interesting thing is someone on the call, one of their prompt engineers was a teacher, 15 years as a teacher, not technical at all. And we're like, what are you doing? Why are you using prompt layer? What's going on?

• 12:24Speaker 00

  And she had explained to us that the engineers had set it up for her. She would go onto prompt layer, edit the prompts and then pull up the app on her phone and it would pull down the latest version of the prompt. And from there, kind of the gears started turning. And where we're at now is the core thesis of what we're building and prompt layer itself is that you're not going to win in the age of GNAI by hiring the best ML engineers. You're not going to build defensibility through machine learning for most companies. You're going to win by working with the domain experts who can write down that problem as we were talking about earlier, who could define the specs of what you're solving. And in cases like building AI teachers, AI therapists, AI doctors, lawyers, I'm an engineer.

• 13:11Speaker 00

  I don't know how a therapist should respond to someone who's depressed, right? Like, I'm not the right person to be in the driver's seat of building that. Does that make sense?

• 13:19Dan Shipper?

  Yeah, I think it makes sense. Can you give me like a concrete example of like what this teacher was prompting? Like, I don't really understand like what she's going in to change and then what changes she's like able to see in the app.

• 13:30Speaker 00

  Yeah, so she's, when I say she's prompting, she's editing the prop that powers the app. So maybe they're opening the app and it's

• 13:44Speaker 00

  a user could be saying, how do I discipline my child? The AI needs to respond very specifically to that. And maybe it was saying, uh, I'm a language model trained by OpenAI. I don't support any discipline. And then she needs to go into it and say, okay, in that use case, let me make sure the AI responds. I don't have children, so maybe I'm not the best one to say how it should respond, but she makes it respond in the right way.

• 14:07Dan Shipper?

  Smack them.

• 14:07Speaker 00

  Yeah, yeah, hit them with a belt. And you don't want to ruin all the other cases, though, where they're asking about what I should feed them. You shouldn't also hit them with the belt. So it's kind of, it's kind of that systematic process and it's just an iteration. We really believe prompt engineering is just how do you close the feedback loop? How do you iterate as quick as possible?

• 14:30Dan Shipper?

  Yeah.

• 14:32Dan Shipper?

  Okay, let me just play devil's advocate for a second. So like, I think, I think I agree with you. I really, I really like this and it resonates a lot with some of the things that we've been building internally at Everi, which I'll tell you about in a second. But um, I think a like AGI pilled response would be like, sure, you can have like the teacher do that. Or you can just like put the AI in the loop and have it measure like, you know, parents who say thumbs up or thumbs down on, you know, a response and then it can just try new things until it like learns what, what responses to give.

• 15:06Dan Shipper?

  And so you don't really even need like human prompt engineers anymore? Like, what do you say to that?

• 15:11Speaker 00

  Yeah, so let's switch gears to AI therapy because I think that's an even better example.

• 15:17Dan Shipper?

  Great.

• 15:18Speaker 00

  Do you want the therapist that just gives the responses that you want to hear and people are giving thumbs up to? Maybe that's not the best metric. Uh, may, maybe there's a better metric that you can find. Maybe the metric is how long people, maybe just exit survey or something like that.

• 15:33Dan Shipper?

  Yeah, it's like long-term well-being or whatever, you know, something like that.

• 15:37Speaker 00

  Yeah, I think the data-driven approaches need to be used as well, but at the end of the day, I think you're going to reach a local minimum. Uh, you're not gonna, when there's 20 other teams working on this exact problem and you're facing off in the market against them, the differentiation is everyone's going to do this data-driven approach. The differentiation is what, going back again, what is that problem you can write? What is that domain expertise you can bake into the application? And as you have more and more users using it, you have them using it different ways. What edge cases can you find? What is the trade-off you're willing to take between latency and quality?

• 16:16Speaker 00

  And that's, there's no one answer to that. Every company has to decide that themselves and that's the job of the prompt engineer, in my opinion.

• 16:24Dan Shipper?

  Yeah, I think people forget with the data-driven stuff that um,

• 16:29Dan Shipper?

  having a perspective is really important and the way that you're collecting data and the way that you're doing the loop of like, you know, response to like, you know, changing the prompt or whatever, like,

• 16:44Dan Shipper?

  that's not neutral, like that embodies a perspective on the world, even if it's like, implicit or you haven't thought about it too much and having a different loop that embodies a different perspective will get much different results. And will create behaviors in people and all that kind of that are just different and what's really interesting is if you're a human with domain expert expertise, like,

• 17:11Dan Shipper?

  You've developed a perspective over many, many thousands and thousands of like data gathering, you know, loops that maybe you could get with an AI, but it would take a while. We're back to the kind of like computational irreducibility thing like you just have to, you can't just be like a super intelligence, sitting on a server and like theorizing about like, what might make people react well in a therapy situation that you probably have to try it. There's maybe some like some weirdness there where like, if you have enough simulated data, if you have enough data, you can like simulate different things and do self-play or whatever. But I don't know, we can just put that to the side for a second. It's just hard. It's much harder to do that. It's much harder and it's much easier and better, probably to like take people who are smart and have a lot of domain expertise.

• 17:56Speaker 00

  Also, who, who sets up the loop, like you were saying, even the supercomputer setting up the loop to learn?

• 18:04Dan Shipper?

  It's super intelligence all the way down, Jared.

• 18:08Speaker 00

  All right. So the super intelligence all the way down, it's still, like you said, it's still taking an opinion. There's still, I, I think it's the same, I think also to go back, Stephen Wolfram also talked about this in that article of,

• 18:26Speaker 00

  you have a computational irreducibility and then you have the, because of that, there's not going to be one model that rules them all because there's multiple answers. There's the, it's irreducible how to find it. And because of that, there is no one way to gather data and come to the conclusion. And by setting up these data-driven approaches, Instagram is different than TikTok. They're both data-driven, but the users are different and how they make decisions are different.

• 18:53Dan Shipper?

  Totally. I think that's totally right. And so the thing that you're making me think of is this, this product that we have. And, and I think this product, it's like, it's a pattern that we're replicating across a lot of different products. So I think it like actually, it's a general thing in AI and it relates to what you're doing, but it's a sort of slightly different take on the same idea, which is. We have this product called Spiral and it helps you automate a lot of repetitive creative work. So like, let's say, like for this podcast, I'm like constantly taking the podcast transcript and turning it into a tweet. And Claude with a few-shot prompt is like really good. If I have historical examples of podcasts that I turn into tweets, like getting me to like 80% on my tweets.

• 19:37Dan Shipper?

  So I just like throw it into Claude and whatever. But the prompt is like kind of hard to construct and it like, you know, it's just, it's kind of messy to do it in Claude. So we built Spiral and in Spiral, you can just create a Spiral that converts podcast transcripts to tweets. And then you give it examples, it writes a little guidebook for itself. And then you have a little form where you can like just paste in your podcast transcripts and it, you know, it makes tweets in your voice and with your examples or whatever. And there's no one podcast transcript to tweet converter. You can make many different spirals with many different tones or voices or styles or, you know, and you can share it with your team, you can make them public, whatever. We launched that like three or four months ago and I think it has that same flavor of not having one single answer to a problem.

• 20:20Dan Shipper?

  And sort of there's some differentiation in having like this diversity or pluralistic like approach to like what prompt is best for a given situation. But one interesting difference that I'd be curious to like pull apart with you is in, I think in the prompt layer example, you've got like a teacher who is like the expert for everyone else who's using the app. Like she's the expert for the parents more or less. And then in a Spiral example, like each user is their own expert to some degree. Cause they're all either constructing their own prompts or taking a spiral that exists and cloning it and maybe changing it. So it seems like there's two, there's two places to prompt engineer. One is to like turn your users into prompt engineers without them really knowing it.

• 21:05Dan Shipper?

  And the other is to like have domain experts like that have crafted something that users are using.

• 21:12Speaker 00

  Or both.

• 21:13Dan Shipper?

  Or both. That's also, that's, yeah, I mean, for sure. Like, yeah.

• 21:17Speaker 00

  I mean, there's probably a skeleton prompt that you guys are using, right?

• 21:22Speaker 00

  Yeah, I guess my, my take on this is that at the core, what Spiral's doing, what Prompt Layer is doing, what these end user applications are doing, is taking a knowledge expert and distributing their knowledge, whether it's distributing it, distributing it across a workflow and across data and distributing it by writing tweets from a podcast transcript, or whether it's taking someone's knowledge of parenting and distributing it to the user base of their app. So I actually see them as the same problem, and it's less about how much degrees of flexibility do I want my user to have. First,

• 22:01Speaker 00

  what workflows can I provide them? What, what workflows can I kind of scale on my own? Can I hire a team of salespeople and build a sales AI application that does what they're doing? Same thing, I think.

• 22:17Dan Shipper?

  What have you learned about like, the most effective ways to like, make good prompts? So we, you know, you talked about the primitives of prompting data evals. When you think about the people who are most effective at like, making good prompts and improving them over time without backsliding, especially, you know, dealing with new models and all this kind of stuff, like, what are the, what are the characteristics of good prompting?

• 22:42Speaker 00

  Yeah, I have a really annoying answer here, which is, it's just the scientific method, just trying and fixing it. I'll give another AGI-style hot take, which is, I really don't like these research papers about prompting. I think people love sharing them, oh, this research paper came out of this new prompting strategy. Prompting doesn't feel like that sort of science. It's more of, I'm trying a bunch of stuff, and as long as I have a good data set and a good framework to iterate and see if it worked, that's what I should be focusing on. How can I edit my prompts in an operational way, as opposed to, how do I discover the correct prompt for me?

• 23:25Dan Shipper?

  And do you have a, do you have a sense about why it's not that kind of science?

• 23:31Speaker 00

  That's a good question. And maybe, maybe it has to do with it being a language. I mean, like, listen, I took machine learning in school. I understand how it works under behind the scenes. But I actually think the best prompt engineer is kind of treat it as a black box. And say, the LM is getting more complicated, more hard to understand, not less. Let's not think about how it works. All I want to think about is how do I map the inputs to the outputs I want? And if you're not getting the outputs you want, that's a skill issue. That's not something else. And I think there's another, there's another degree here besides just the prompt is, what is the combination of prompts you're using? Now you have exponentially more ways to improve it. Are you breaking down the prompts?

• 24:15Speaker 00

  We have a lot of opinions we've learned about the right architecture to use to build it with a team. But that's more about how do you actually ship a product that works in practice? As opposed to these are the right ways to prompt.

• 24:31Dan Shipper?

  Yeah, I think that makes sense. I mean, if I had to guess, it's like, it's sort of the same things that we've been talking about. Like, in order to say like X technique is better,

• 24:42Dan Shipper?

  you have to define what better is, which, you know, depending on how you define better, you're going to get different results. And also, so it's, it's very dependent on the, on, on that. And then it's also dependent on the data set. And so like, you can kind of find some hand wavy things, but like, it's, it's so contextual that like, it's probably, I mean, maybe you start with some of the things that are like best practices or whatever, but you're probably just going to like discover stuff for your particular use case where group level statistics don't really apply. Which, you know, I, I love that kind of, that kind of stuff. I'm a huge nerd for that.

• 25:17Speaker 00

  Yeah, there, and there's like, can you speak the language of a certain model family? Obviously, I mean, the popular ones is Claude, it's better XML, GPT, it's better Markdown. And those are kind of model level things that you can learn. But I look at it as like, What idioms does the AI understand and can you speak that? But I think the best example here is tool calling and function calling. Whereas personally, I love using function calling when I can, even for things that are not functions, because implicitly, that's the language that it knows. And you're actually by having the AI return or interpret a tool call,

• 25:59Speaker 00

  you're conveying much more information, at least in my mental model, than you would be as writing, this is the data I got back, if that makes sense.

• 26:09Dan Shipper?

  Yeah, I think so. But I don't want to let you off the hook. Like, I agree that it is like scientific method. But like, I just think that as someone who spends all their time thinking about how to help people do this, like, you must have some opinions on like, what do you see that works? And like, how do you think about it? And all that kind of stuff that goes deeper than than try it and see what happens.

• 26:30Speaker 00

  Sure, sure, sure. Yeah, the first thing I always start with, try and see what happens. Then make your iteration loop as quick as possible so you could try as much. Then we can go to kind of best practices that we recommend. The biggest one, I think, is this, we call it prompt router approach, but there's a lot of different words and it's kind of overloaded. So I think there's, the naive approach to prompting is just stack messaging. So I just have one prompt that does everything and the user sends a message and I stack another one, then I stack, stack, stack. I think the better approach is to kind of build a workflow, a DAG, a graph, and route them to the right prompt.

• 27:10Speaker 00

  Because if you, if you can make your prompts do one thing and do one thing really well, it's going to be easier to test, easier to collaborate, and it's just going to work better. As you try to make the prompt do more and more things, it's, it's going to be more likely to fail and more likely to actually build unit tests on it. So it's, it's around operationalizing and then kind of structuring your prompts in this discrete way. Although, I'll argue with myself here a little bit, that as models get better, you'll have to do that less probably, but there's always a trade off. Because if you're building individual prompts to do one and only one thing, it's gonna...

• 27:55Speaker 00

  Work much more of the time and have much less failure cases, but it's also going to have less degrees of flexibility. So you might get new user messages you're not expecting. And then the opposite is also true. If everything's one prompt, you're going to be able to answer any message the user sends, but might not always be right.

• 28:12Dan Shipper?

  Yeah, that's interesting. That makes a lot of sense. And it also sounds like maybe

• 28:20Dan Shipper?

  towards the beginning of the lifecycle of a prompt, when it's less known what the distribution of questions you're going to get or messages from a user you're going to get, the better it is to like, start with a single prompt so that you can iterate more quickly and cover the whole variety of like cases. And then as you're maturing, you're probably, you're probably learning like, okay, these are the kinds of things I typically get. And then you can build that directed graph of prompts and then you can start to tune each individual prompt. Once, once things are a little bit more set. Does that sound right?

• 28:55Speaker 00

  Totally. It's also a solution to jailbreaking, right? If each prompt does one thing. Say I'm making a sales bot to issue. So users can place orders and refunds and that sort of thing.

• 29:09Dan Shipper?

  Yeah.

• 29:10Speaker 00

  I can make individual prompts to do a refund or to do a new sale or whatever. And if I route to the right prompt, if the user asks me to refund a million dollars and I'm in the sales prompt, it's not, there's no risk of kind of breaking out, but then you can ask for both.

• 29:32Dan Shipper?

  What about best practices for evals?

• 29:34Speaker 00

  There's a lot of ways you can go here, I think.

• 29:38Speaker 00

  And I have a very similar answer, which is starting easy. The easiest thing at the 80% case for evals is just back test on your old data and see how it's changing because the most information dense thing you can learn from an eval is that it changed everything. That's what you don't want to do. So what we have a lot of our users do is they just create a back test based on like, all right, our last thousand or our last 10,000 prompt responses, let just run the new prompt using that data. See how much it changes. Maybe you want to do a cosine similarity, or maybe you just want to scroll through and eyeball the diffs. Once you do that, then you can kind of.

• 30:20Speaker 00

  Get a little bit more fancy and it depends on the use case again, if you have a ground truth. So if you can plug into thumbs up, thumbs down, you can plug into did the sale, was the sale made? Did the ticket close that sort of thing? Then you can really throw some A B testing in there and and have anchor it on real metrics or build the eval that gives you a real score. Unfortunately, if you're doing something like AI summarizing or summarizing calls or something like that where there isn't a ground truth, then it gets a little more complicated. And what we recommend is either if you're having human graders read it, you could do that.

• 31:02Speaker 00

  Or you have your prompt engineer, you need to sit down and say, what are the heuristics? When you're looking at the output yourself and trying to decide if it's real and then trying to mimic each and every heuristic and build the metrics based on that.

• 31:18Dan Shipper?

  Interesting. And then what about data sets?

• 31:21Speaker 00

  The three primitives, like we said, you can skip one of them. So if you care a lot about evals, you care a lot about prompts, you could just build backtests and those could be your data sets. But if you don't have the backtest data, you're going to want to focus a lot on building ground truth data sets. And if you can really get that ground truth, then you're sailing. Then prompt engineering is kind of easy because it's do you get a hundred percent or do you not? So data sets, I think there's a lot of room of synthetically generating them as well. I've done that to kind of bootstrap data sets myself.

• 31:55Dan Shipper?

  Yeah. What about situations where you don't have ground truth? That's just something I'm thinking about personally is like, how do you eval on that? And it's funny because I think a lot of the AI labs are like going hard only on problems that have ground truth. So like O1, for example, like they're doing reasoning and the reasoning chains, like they can validate the reasoning chains because it's all math problems. Something about that seems like kind of fishy to me and it feels like important to also be able to like eval on things without ground truth. Like, yeah, talk about that.

• 32:27Speaker 00

  Yeah. Well, first of all, my O1 take is that it's just prompt engineering and that's the cool part of it.

• 32:34Dan Shipper?

  Everything is prompt engineering. You heard it here first.

• 32:37Speaker 00

  I really think O1 is just a bunch of different, just feeding it into it a few times and maybe they do some low level stuff there to make it a little bit better. But I think the core. Of that innovation from it here, but to answer your question of what if you don't have the ground truth, what's important, they don't eval ground truth because they don't eval problems without ground truth because you can't really do that in a generic way. It's kind of hard. That's why I don't trust these kind of eval benchmarks that people come out with. They're not, they're a little bit useful.

• 33:14Speaker 00

  Yeah, maybe we'll talk about example because I think it is very example based. So let's go back to this transcript summaries because I was thinking about that in depth the other week. What I would do is if I'm building, let's say it's an email summarizer. Every day it sends me a summary of all my emails.

• 33:33Dan Shipper?

  We're building that internally right now. Are you, are you reading my mind?

• 33:36Speaker 00

  No, but that sounds good.

• 33:39Dan Shipper?

  It'll probably be out in a couple of weeks.

• 33:41Speaker 00

  So let's talk about how you eval it. I'll tell you what I would do and you tell me what you're doing.

• 33:47Speaker 00

  If I was building that, I would just make something quick and then look at a few summaries and say, what do I like, what do I not like? And start to generate individual test cases based on that. So maybe the first test case is, and I'm sure you've experienced this in this case, sometimes the AI will give you a little excerpt at the bottom. This summary is about blah, blah, blah. That would be my first check. Does it have that excerpt? I don't want it to have that excerpt. Second check, does it use markdown? And maybe it's too indented. Does it only have one level of indents? And then these really simple, the hard part of this is understanding what your brain does when you eyeball a summary to see if it's good or not, and then breaking that down to individual characteristics, but.

• 34:32Dan Shipper?

  That's really interesting. So I'm not close enough to the actual engineering happening on this to like know exactly the evals. The, the person building it is one of our entrepreneurs in residence, Kiran.

• 34:45Dan Shipper?

  I know we have an eval set for another component of the product is drafting. So I know we have an eval set for that and that's basically just like labeled email responses. So it's like email responses that we wished we had gotten and then it like checks. There's a couple of different similarity scores between like the what got sent and what got what actually got sent and what got drafted. So it's like, um, it checks to see if the, the, the prompt is generating an email that is close to the email that I actually generated or someone else in the data set actually generated. For the summary, email summaries, I can tell you like the way that we approached it is I

• 35:23Dan Shipper?

  literally just sat down, looked at my inbox and like wrote in a Google document, the, the summary of my inbox that I wish I'd gotten.

• 35:35Dan Shipper?

  And that's how we started, like handwriting it. And then from there you can pull out all these like little principles, like the summary,

• 35:44Dan Shipper?

  should not, should not be passive voice, it should be active voice or like whatever, you know? And I don't actually know if there's evals yet for that, but um, it's, we'll, we'll have that eventually, like in the next couple of weeks, I'm sure we'll have something. But yeah, like it was, it, it required one person, I guess me, just to like write it out and then from there you can like pull out principles and stuff.

• 36:06Speaker 00

  Yeah. It's like, when you build these applications, you really want to skip, you want, you want to skip the hard work, but the hard work just changes from writing the code to figuring out how it'll actually be correct or not. And in that case, it's like you said, you have to create the summaries and read it. And yeah, I, I'm, I'm a little skeptical of that similarity thing you were talking about. Uh, I mean, I think it's good as an approximation, but in the email, maybe I think there's a lot of, I use a, I mean, the superhuman command, command J, uh, response thing.

• 36:46Speaker 00

  I click retry like six times when I'm doing that. I'm not trying to hate on the team. I'm sure they've done a great job, but I click retry a lot and the retries are all very similar to each other, but very different if I were to send it in terms of wording.

• 37:02Dan Shipper?

  Hmm, that's interesting. Yeah, it's not one similarity score. It's like five different ones. And, and we are like, we are like eyeballing it too and using it every day. So that's one of the things I think is really great, which I think goes to your point is like, we're building all this stuff internally. Like we've launched three apps so far. We have a bunch of other ones that we're kind of in the process of. And what's really cool is we're all building stuff that we want to use. And we're all kind of similar because everyone's a fan of every and sort of like came in from every. And so someone makes something and then we're all using this email summarizer.

• 37:42Dan Shipper?

  And so like every day there's like feedback in the channel being like, hey, like this thing didn't work or I want this or like blah, blah, whatever. So it creates this like really rich feedback loop. Because not only are you using it yourself, but everyone else around you is using it too. And I think that's sort of unique. I mean, there's something unique about it to every, I think right now. But I think what's really unique is just the time right now is one where

• 38:11Dan Shipper?

  the low hanging fruit hasn't been picked. So you can build something for yourself. Like, you know, I assume PromFlare to some degree is an internal tool. So it came out of something that, you know, you, you, you just needed. And we're, we're early enough in the kind of AI wave that like, you can still do that. And it's not like 15 other people have tried it over the last 10 years or whatever. And I think that's so special to like be in a place where you can just make, make stuff for yourself. I love it so much.

• 38:40Speaker 00

  The hacker, the hacker energy is back. We, we don't have to buy a fleet of cars to build a startup in the weekend anymore. I

• 38:52Speaker 00

  I I yeah, I I actually think it also is another phenomenon that's going on, which is L M's unlock a new type of kind of software you can build where, for example, one of the internal tools I have that I use all the time is a very simple natural language to SQL, specifically for our schema. So I have a lot of information on like, this is a request lock, whatever. Uh, I don't know if I, maybe I would pay for a tool like that if it was really good, but this also prob- might just exist in that subset of software, like, uh, people call it like the single use software that you're not really going to sell to other people, but it's easy enough to make. It's one problem, pretty simple. I mean, I could,

• 39:38Speaker 00

  I could really spend months and make it really good, but for my use case, that's as good as I need it.

• 39:43Dan Shipper?

  Yeah, I love that. I think, I think there's the single use thing, and I think that that extends even further away from software than people realize like, there's a lot of consternation right now about, um, uh, like AI replacing writing or like art or podcasting or whatever. Um, and it certainly is lowering the cost of telling stories. But my experience thus far is what that actually does, rather than crowd out people who are telling, you know, like professional storytellers, it allows you to tell stories about, about things that like, it used to cost too much to tell stories about. So like a good example is just Notebook LM. Like, the thing that's so compelling about that

• 40:27Dan Shipper?

  is you can get an NPR style talk show about something that NPR would never cover.

• 40:33Dan Shipper?

  And I think that that's so cool because there's, there are so many stories that are happening all the time in our lives that um, maybe they don't deserve like a Netflix special, but they certainly deserve better storytelling than we have. And I think that's sort of the promise of these things. And I, I think that's kind of beautiful.

• 40:53Speaker 00

  Yeah, my, my friend made a, in the software side, my friend made like a website that helps you just make a button with sliders for the button radius and the color. And she just made that for herself. Nobody's ever gonna spend the time to set up the server and pay for the hosting for that. But if you can make it with one Claude command, why not? But, uh, I don't know. I also like. Uh, not to get too in the clouds here a little bit, but uh, I, I look at it as like a junk food analogy. Where uh, so about like AI music, for example, I think we'll have a lot of AI music. And a lot will be junk food, meaning a lot of people will consume it and will love it and it'll be great.

• 41:35Speaker 00

  And uh, but they'll still be the organic uh, farm to table musicians where a human makes it. And it's just gonna be, it's just gonna be different. It's gonna solve different things.

• 41:47Dan Shipper?

  Yeah, I, I definitely, I definitely agree with that. What's the gnarliest prompt you've run into?

• 41:54Speaker 00

  Um, I'm gonna be honest, it's the chat GPT or anthropic system prompts. Those are bad.

• 42:00Dan Shipper?

  Wow, hot take.

• 42:04Speaker 00

  Yeah, I mean, this is actually, I like this because this is a, connects to the other stuff we were talking about. They don't, if you're building chat GPT or Claude, and I encourage everyone to check out, people leak their system prompts. You should be able to find it with a quick Google search. Uh, what are they evaluating on? Uh, they don't have a specific use case they're evaling uh, or tying their prompt into. So you see these prompts are long and just run on, I call it prompt debt. And uh, it's just they're accumulating all this debt of, oh, do this, if they say this, don't say this. Uh, you know, the classic, uh, the Google, uh, make sure every historical figure is diverse, all that stuff.

• 42:47Speaker 00

  And uh, it just, you keep adding it to your prompt and I don't think that's a good prompt. I think that's a, you, you're not clear and concise with the problem you're solving. You're just tacking on more stuff and maybe that's overfitting, in my opinion.

• 43:01Dan Shipper?

  Yeah. Well, how would you solve for that? Cause like they're real, they really need to build a general purpose tool that hundreds of millions of people can use. So like, how do you, how would you do it differently?

• 43:11Speaker 00

  I don't know. But I, I, I think, uh, I think the fact that they're building a general purpose tool shows you why your prompt shouldn't look like that, because you should have a better way than them to evaluate if your prompt works. And then be able to trim it down. Um, if. I don't think they should do this because this is not their business model. But if I started a company to build a general purpose AI tool, I would probably have different prompts for different types of thing and try to route it. But then you have the latency concern. I spoke with someone from the Snap AI team. And I think theirs, you don't hear about it a lot, but it's one of the most popular AI chat box, the Snapchat AI.

• 43:54Speaker 00

  And yeah, because all the kids who don't know about ChatGPT or aren't old enough to sign up, I think, use it. And they have tons of models from last I spoke to someone there. Maybe it's different now, but they're constantly trying to understand, do I send it to this model? Do I send it to this prompt? And they have a lot of thought that went into making that good, I think.

• 44:17Dan Shipper?

  So do you think that's sort of like the future of this? Like, do you think these kind of, even these kinds of general purpose ones are gonna do explicit routing? Or is like, because there's another way to do this, which is like, it's sort of implicit or internal in the model. Like, the models are mixtures of experts where it's like, it is sort of routing your request to a certain part of the model, depending on what it is. Like, there are trade offs between the explicit and implicit versions of this kind of thing. Like, which one, do you think the explicit one is part of the future?

• 44:48Speaker 00

  I think to the end user, probably, I mean, that's a UX problem. But look at ChatGPT's evolution. You had to select which tools you wanted and which plugins you wanted. And that was just a horrible experience. And I think they quickly moved to this world where ChatGPT will choose whatever tool they want you to use and will let you use the code interpreter. You don't have to use the code interpreter style ChatGPT. From the, from the technical perspective of someone building these applications, bad answer, but it's very use case dependent. It's very, what are you building? What are your trade offs? What's your latency? What's your cost trade offs? Hard to say.

• 45:29Dan Shipper?

  What are like, what's on your mind right now? Like, what are you thinking about currently in the evolution of the business? And like, what's, you know, what does the next couple of months look like?

• 45:38Speaker 00

  I think we're shifting into the next phase of this company where we spent the last year and a half really trying to understand where we fit in, what the prompt layer is, what layer of the stack is this? And what the, what the value prop is. And I think now we have a really strong thesis of kind of what we were talking about earlier, bringing these domain experts, these non-technical subject matter experts and putting them in the driver seat of prompt engineering. And our product does that. But I think a lot of the next few months is making kind of the happy path a little bit easier, making, encouraging that and, and kind of scaling it up from this point, running out of that.

• 46:24Dan Shipper?

  Like, how do you make that happen? Like, this is not a role that I think a lot of people are hiring for right now, or that people haven't realized it yet. So if you're kind of pushing the company in that direction, and you're sort of skating to where you think the puck is going to be, how do you make sure the puck gets there? Are you doing anything in particular? Is it, is it just about building the right tools and people will figure it out? Or like, yeah, how do you think about it?

• 46:54Speaker 00

  So we have a lot of people doing this today already. But you're right, people haven't realized this.

• 47:00Dan Shipper?

  Wait, actually, before you, before you tell me that, what are the characteristics of people who like figured out that this is a thing that they're, they're doing, like your early adopter crowd?

• 47:09Speaker 00

  So there's two things that bring people to the thesis of non-technical prompt engineers. The first thing is, they just are kind of not a visionary, but they, they understand it.

• 47:23Dan Shipper?

  They're just smart.

• 47:26Speaker 00

  They have this like, they come in with this as a preconceived notion. The second thing, and this is the thing we're kind of locking into more, because the first you, you get a few of those, but you can't count on that. Second is companies that are

• 47:41Speaker 00

  Actually building something revenue generated, building something that you're going to put a team behind. Because when you put a team behind something, collaboration becomes a big issue. They're gonna, almost always, especially in non-technical domains, if you're building Cursor, might be different because your experts are software engineers. But in these other cases, they're ending up using, having a lot of handoffs with QA testers, who are asking a lot of legal questions or whatever, and then they have to keep coming back and telling the engineers to fix something. So they experience the handoff and then they're also trying to hire. But I think at the core, the interesting thing here and the big bet is that we don't have to convince anybody this.

• 48:24Speaker 00

  The market is going to do the convincing for us because the winners are going to be the ones who do this. And then everyone's going to say, oh, they're winning, I gotta do that. And

• 48:35Dan Shipper?

  And so if you're building the tools to connect domain experts to companies like, and one way to do that is to kind of rope them into the loop of prompts. Is part of this also like recruiting domain experts and the sort of marketplace thing? Where it's like, if you're building a teacher app, you'll hook me up with a teacher? Is that the user's job or?

• 48:56Speaker 00

  Today it's the user's job, but I have heard from a lot of teams that that's hard for them to, because it's, it's not just data labelers, you're not just MTurking it. It's expert data labelers in a specific category. And people do hire out for this. I've heard of legal AI teams doing this, but you have to find the right person. So maybe we'll, I don't, I don't think we're going to become a marketplace for that, but maybe we'll help people, maybe we'll help connect them.

• 49:24Dan Shipper?

  That's always, that's always been the kind of the pet thing of mine is like,

• 49:30Dan Shipper?

  language models, like their summaries are all like RLHF'd on like human feedback from data labelers that are not professional writers. And even among professional writers, the like level of skill and quality is highly variable. And, and then there's no one good metric for like. What is a good summary or a good writer? But like in my, in my ideal world, like I would be so curious to see a model that was like trained, like or RLHF just on like, you know, my favorite, a couple of my favorite writers, like doing data labeling. I feel like it would be different.

• 50:07Speaker 00

  Would they argue with each other though? Because maybe it's the nature of them being your favorite writer that they're a little unique in their own thing. And L L S not to do that.

• 50:16Dan Shipper?

  It's possible, but it would definitely be, even if they argued and didn't agree like, and maybe that makes the final output like a little more like unreliable. I feel like that's, I'm fine with it. It's better than the bland, like whatever the bland thing is now, you know? Cause I think there's this thing happening with language model summaries where they're very adept at saying a lot without saying anything, which is a specific kind of bad summary, you know, that a data labeler, because I, I know this because

• 50:49Dan Shipper?

  I edit a lot of articles at every, and I mean, I don't edit as much anymore, but like I used to. And there's such a big difference, like you can glance at an article and be like, yeah, this is pretty much, pretty much right. And I would just like give it the, give it the check mark and just be like, yep, approved, human approved. And then you actually like sit down with the article and you're really like pulling it apart to be like, does this really make sense? And then you just see all the stuff that you would not have seen before. And I can imagine that that's happening with these labelers, you know, it's like, you're, you're sort of like tuned out at some point and you're just like, yeah, this isn't like garbage. And then you just get this sort of milquetoast thing, especially with all the like,

• 51:32Dan Shipper?

  they want it to be like not controversial and whatever. I don't know.

• 51:35Speaker 00

  Interesting, interesting. Maybe, uh, maybe you do that with prompt engineering. Maybe you don't need to train your own model.

• 51:41Dan Shipper?

  Maybe, I mean, I can get Claude to do some of this, but like not, it's still not amazing. It, it depends on the thing. It depends on the task.

• 51:49Speaker 00

  It's hard. Yeah. It's hard to know, will it ever?

• 51:53Speaker 00

  I'm sure it'll be, it'll get better. I, when I summarize with AI like I have, I almost always say use bullets, don't use complete sentences and like, it helps it a little bit, but I totally see what you're saying. Yeah, I don't, can't, that's interesting one. That's interesting, but it, because it's, what is the core that makes the writing good? Is it that it's concise or is it something a little bit more to piss off the AGI crowd or again, a little more human

• 52:24Speaker 00

  that, that you can't really codify and that's not just logic. Uh, I don't know, maybe not.

• 52:29Dan Shipper?

  I think that you can, but it's, if you said it, it would just be like a high level thing like, a good summary has to be concrete and it has to tell you the most important thing that you need to know and leave out everything else in a concise way, right? But like, what's the most important thing is like incredibly dependent on the situation and the person. And language models are, one of the beauties of language models is they're super sensitive to context. So they should be able to like pick that up.

• 52:59Dan Shipper?

  But yeah, they're like, there's no generically best summary. There's just like, there's a principle, but the, the way that that principle like plays out operationally is like

• 53:13Dan Shipper?

  very different depending on the context.

• 53:15Speaker 00

  So what you said makes, it sounds like multiple tasks to me and maybe that's the solution.

• 53:25Speaker 00

  When I was in, I remember when I was, when I was in high school in some classes, I tried a little too hard. Then I learned it doesn't really help, but I, I would make summaries of my notes and then make summaries of that to learn. And it is hard to figure out what the most important thing is, right?

• 53:42Speaker 00

  So maybe that needs to be the pre-processing task. Maybe the fact that the L M doesn't generate what to you is a good summary is because it's, task is just to condense the data, but not to make the reader understand the important thing.

• 54:00Dan Shipper?

  Yeah. This is really interesting. I just spent, I was in SSF like twice over the last like three weeks. And so I got like super AGI pilled because you just like, you just see enough of the like thousand yard stare of the like internal researcher at the like AI labs and you're just like, oh, fuck.

• 54:18Dan Shipper?

  And you're just like, no, this is gonna be fine. Like.

• 54:22Speaker 00

  Maybe not.

• 54:23Dan Shipper?

  I mean, maybe you're not saying it's gonna, it's all gonna be fine, but like, I think that you have a different perspective, at least on the fundamental limits of whatever, like at least near term AGI might look like, or you know, more advanced AI might look like. I'm getting that vibe from you. And I actually, I'm swinging back a little bit and I always sort of swing.

• 54:47Speaker 00

  Come back to the dark side.

• 54:50Dan Shipper?

  I think it's, no, I think that there's something.

• 54:55Dan Shipper?

  It's really important to realize that like, people have been predicting the end of the world because of the latest new technology since the world started, basically.

• 55:04Dan Shipper?

  And so if you're gonna place bets, like, you gotta bet against those people probably. Is my, is my thought. And there, there are way more limits.

• 55:14Speaker 00

  Until they're right.

• 55:15Dan Shipper?

  Until they're right. Well, you know, but yeah, so you never know when you're the turkey and it's Thanksgiving, but like, and if they are right, like the bet is, you know, it's like a Pascal's wager type thing. Like, if they are right, then the bet doesn't matter because you're dead anyway. So you might as well believe that.

• 55:30Speaker 00

  Yeah, yeah. I, it's also, how do you define AGI? Because when I kind of, when I like to rail on AGI, I don't know. I'm using kind of the hypothetical, oh, it's this human like AI that takes everything over definition. But I think the real definition is that it's just like, an AI that can do most tasks humans can do. And I think we're there already. I think you could do that with enough prompt engineering. And you can kind of build systems using that toolbox, but does that matter? Yeah, it matters for the economy for a little bit. Maybe it'll correct after that, but.

• 56:11Dan Shipper?

  But I think by the time you're like talking about prompt engineering though, like that's already you're eliminating the AGI thing.

• 56:17Speaker 00

  What about O1 or Sora?

• 56:19Dan Shipper?

  O1 doesn't even know modus ponens.

• 56:21Speaker 00

  I don't know what that is either.

• 56:24Dan Shipper?

  Exactly. You do know what it is. Basically, for example, with O1, you can be like, who is John Mayer's mother? It'll know the name. It'll think it through and it'll know the name. I don't remember what her name is. Let's say it's Cynthia Mayer. And then if you ask, you just start a new thing and you're like, who is Cynthia Mayer? It won't know. And you can even be like, I want you to like, think about like, other, like, famous people who have the last name Mayer. So it gets in the, gets in its thought process and it still doesn't get it all the time. And that's something that like, yeah, like a human just knows, basically.

• 56:58Speaker 00

  We don't know her name, though.

• 57:01Dan Shipper?

  I don't know her name both ways, but if I knew it one way, I would know it the other.

• 57:06Speaker 00

  I see, I see, I see. Well, yeah, I mean, this is why I, I know this is not a complete way to think about it, but I think the simplest way to think about LLMs is just, again, solving language. And language doesn't know who John Mayer's mom is. But if you give language the repository of information, it can tell you that.

• 57:30Dan Shipper?

  I don't know what it means to solve language. Like, what is, what do you mean?

• 57:35Speaker 00

  So, I mean, a big part of computer science for many years is processing natural language or language and turning it into data. LLMs are really good at that and basically have solved the problem of going from human speech or human text to information that a computer can understand and vice versa. Before that, that was completely, at least vice, going from data to language is basically impossible to do it well. That's solved. I'm, I think, anything with regard to bringing in real world data, John Mayer's mom, that sort of stuff, is just what you hook this box up to,

• 58:20Speaker 00

  what you hook this technology up to. I'm, I think like, how you get to the...

• 58:25Dan Shipper?

  I think what you're saying is like, you could probably take the same architecture that we have now and with a different training method or, you know, dataset or whatever, you could get it to answer both ways. And I totally think that that's right. But what I'm trying, I guess what I'm trying to... Point two is like, that's why I'm like, I think it's possible we're gonna get there. I just don't think we're there right now. O1 is like clearly, the thing that people miss is like, in my opinion, is, oh, they're like, oh yeah, oh, one can do physics problems better than I can. And it's like, yeah, but like, it gets all this other basic stuff wrong. So, um, so for, so I'm not ready to call AGI. I'm ready to call it's smart on certain, in certain domains. And I think that's really important.

• 59:09Dan Shipper?

  Like, that shouldn't be overlooked. It's not like, I'm not like being Gary Marcus here where I'm like, you know, they suck or whatever. Um, uh, I think that they're, I think that they're truly incredible, but I also think there's a lot of, there's a lot of worship. I wouldn't, I wouldn't call it AGI.

• 59:26Speaker 00

  So hypothetically, say we, all we needed to do is train the LM on 100x or 1,000x the data size. Probably not, but let's say that was the way to do it.

• 59:41Speaker 00

  What's the difference between having an O2 model that can answer John Mayer's mom both directions versus having a one that's plugged into the data set and that kind of end user application that lets you ask it both ways?

• 59:56Dan Shipper?

  Uh, you have to predefined the, uh, data set. So, um, like, you have to have like a, an explicit knowledge store that has all the knowledge in it. So you're kind of like cheating and, and also like having that predefined knowledge store is a impossible problem. You can't, you can't have all of that in a, in a data, database. And that's why it took so long to get even the models we have now. Cause you, you have to have, for a knowledge store like that to work, you have to have a more bounded problem, a less general problem.

• 1:00:30Speaker 00

  A lot of hypothetical, I mean, the weights are technically a data set.

• 1:00:34Dan Shipper?

  They are, but they're, it's inexplicit. It's a totally different thing. It's not the same thing as a database. It's much more flexible.

• 1:00:41Speaker 00

  Right.

• 1:00:41Dan Shipper?

  I, I like where this is going. I think this is a, this is a very fun conversation. We're, we are at time. Um, and I really, really appreciate you coming on the show. For people who want to try out prompt layer or find you, where can they find you?

• 1:00:53Speaker 00

  Yeah, you can find me on Twitter, on X at I am Jared Z. I am J A R E D Z. Also check out prompt layer, promptlayer.com. It's, uh, free to sign up and use for hackers and hobbyists and small teams. And then, uh, not so expensive for startups either. And it's a Great way to build AGI.

• 1:01:15Dan Shipper?

  Sweet. Thanks, Jared.

• 1:01:16Speaker 00

  Thank you.

• 1:01:24Speaker 02

  Oh, my gosh, folks. You absolutely, positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard. But instead of gold, it's filled with pure, unadulterated knowledge bombs about ChatGPT. Every episode is a roller coaster of emotions, insights, and laughter that will leave you on the edge of your seat, craving for more. It's not just a show. It's a journey into the future with Dan Shipper as the captain of the spaceship. So do yourself a favor. Hit like, smash subscribe, and strap in for the ride of your life. And now, without any further ado, let me just say, Dan, I'm absolutely, hopelessly in love with you.