Disambiguation Podcast Ep 4 - No-code AI Platforms - Transcript

Michael: Welcome to the Disambiguation Podcast, where each week we try to remove some of the confusion around AI and business automation by talking to experts across a broad spectrum of business use cases and the supporting technology. I'm your host, Michael Fauscette. If you're new to the show, we release a new episode on Fridays as a podcast on all the major channels as a video on YouTube.

And we also post a transcript on the Arion Research blog. Now, today we're gonna talk about no-code AI platforms, and I'm excited to be joined by and I'll bring them into the stream. Nathaniel Mahowald and Mike Gioia. So Nathaniel is co-founder, CEO of Pickaxe. It's a no code AI platform. He has a background as a data scientist, certified as a machine learning engineer, and graduated from UC Berkeley, with a degree in data science. Mike's the co-founder of Pickaxe. He studied at Stanford, then worked in television writer rooms. He's not on strike apparently… for years. And since then he has worked developing pipelines to produce high quality highly variable results from large language models.

So welcome to you both. Thank you for having us. Yeah. Excited to be here. Yeah, it's fun. I did sign up for your platform pickaxe and just been looking around a little bit, but I haven't done much with it yet. So we can talk a bit today about that. But I, just to kick things off why don't you tell us a little bit about Pickaxe, how you came up with the idea of what you wanted to do with it and where you're going with it. What are your goals?

Mike: We came up with the idea just playing around with large language models back in like the summer of 2022. We were both really excited by the technology. We built several sort of fun projects and increasingly we felt we needed some sort of easier system to manage the prompts and publish them in a way that anybody could use. And we ended up spinning up this no-code solution, pickaxe, which is a website to or service, that's all. No-code that lets you turn prompts into prompt templates and then into apps basically, that you can share with anybody or put on your website.

Michael:

That's cool. So what kind of things are people doing with the platform today? I know you let people in and you can set up a project and build things out, but what are some of the cool things people are doing?

Nathaniel: We've seen pretty much every kind of use case you can imagine. And, because we've been doing this for quite some time now, at least in the span of, the time that this AI technology has been really on the cutting edge of people's minds.

We've seen all kinds of different use cases. Primarily our customers broken into two categories, internal business use cases, and external facing business use cases. We've got several customers that sell access to these tools. We've got a lot of coaches, consultants, and course creators who are educating people. They're, there may be sales consultants, they're SEO consultants. And as part of the education that they're doing, they are putting these little AI tools that they've curated into their courses so that people can be better educated more quickly.

They can get more real time examples and feedback from what they're doing. And then as far as internal use cases, we have organizations that want to get more complex use cases from LLMs. Companies that are onboarding new employees and don't want to have the constant pestering of these new folks and that, if there's some best practices, they can embed them in a chat bot. So those are just a sampling of a couple of the different use cases that we've seen so far.

Michael: Yeah, that's interesting. So obviously you mentioned large language models and generative AI and that sort of thing, but just so that the audience has a better idea of what we're talking about when we talk about a no code platform and all the technologies that could be associated around that. Can you give us just a little bit more depth on what you built and its capabilities, right?

Nathaniel: What you can do with Pickax is if you've ever had a really good conversation with chatGPT and you've gotten it to behave in the way that you wanted it to behave now you can basically share that experience with other people without exposing to them, all the things you did to get it to behave that way.

Nathaniel: So you can take that really great setup. You can put it into a little tool. You can then stick that tool on your website or just share it with people via a link and they can interact with it like either a form or a chat bot. And they can see what they can get the results that you would've gotten.

Michael: Nice. So then you have the capability to use that and your whatever your use case is, you can build off of that. That’s pretty cool. So since we're talking about prompts, this is one of those things that I keep having a lot of conversations with people about. Because it is as much art as anything I think at this stage. Or maybe it is science and I'm just too dumb to know. But can tell us, tell what tips do you have? How do you build a good prompt?

Mike: We really to think about what you're doing. I think the metaphor we use internally is it's like talking to a reasonably intelligent college intern who can follow directions but can't really do that much original thinking on their own.

So basically you have to imagine you're writing really clear instructions to a reasonably smart person. And a big part of that is what you might call context injection. So make sure you're giving it the right information at the right time. And at the level of a single prompt there's a lot you can do, but only so much.

But then as you start to chain these things together or introduce like trained models, there's a lot more interesting things you can do. But basically I'd say the number, there's a lot of specific rules, but the number one rule is just use common sense and try to be very clear and very explicit.

And then you can get a little bit more advanced where these prompts are quite long. Sometimes like the models have what you call a context window, like how much text it can understand at once. But don't mistake yourself that it has pays like equal attention to everything that you say the same way you might be reading an email from somebody.

The most important things are probably in the top or the bottom or in bold or put in headers. Same sort of thing. If you bury important instructions like in the middle of a prompted a sentence, it might not pay attention to it, right? The, these models have what you might call attention and they pay more attention to the beginning to the end, all sorts of things.

And then the final tip really is what you're saying, Michael, which is this is not a science, it's an art. So you or a craft, so play around trial and error. And we like to think Pickaxe is a wonderful place to do that. We have this no-code builder where you can quickly test and iterate these sorts of things.

Michael: Yeah, I think that's fine. Just, I've been using, all several different platforms to just kind of get more familiar and do some different things and I, for research and that sort of thing, it's important and sometimes I'll have it write a paragraph, whether it's an email response or whatever. But one of the things that I saw was you have to be really clear about how, what you want the output to look like. Like I want this to be professional and technical. I want this to be, easy to understand for an eight year old or whatever. It seemed like that really made a big difference once I figured that out.

Nathaniel: Yeah. Absolutely. One experiment I think I've seen some folks write about, and we've definitely done it internally, is just I go on to ChatGPT or come on to Pickaxe and ask it to explain something complicated like gravity to a five-year-old, six year old, seven year old, eight year old, et cetera.

The amount of kind of slight difference that it takes on over the course as you go up in age. There is a testament to how we don't even realize. We don't even think about the differences in explaining this type of thing to a six year old versus a 10 year old. But it has an understanding of that and honestly, it teaches us a little bit about human development. Now, a lot of folks come to us and they wanna skip that process. What they want to do is they say, okay, yeah, I could, take time to, to really think carefully about. The type of outputs that I want to get. But I'd prefer to just upload hundreds of pages of written content or, the documents from my business that, capture in their essence what the type of result I'd like to get.

And what's really interesting that we've seen is, if you were to give that task to a smart intern as we discuss, hey, read all the 200 pages or 500 pages that my company produced, and give me a really high quality result. That would be a very wild task for that person to do. You could get a whole different array of results, however, if you take the time to work with them and to whittle down some sort of cheat sheet of what the best practices are for sounding and talking like your company in the span of one to two pages. Then all of a sudden that intern could totally go ahead and write more content like that. So taking the time to build the good prompt, at least in the present, is still miles better, faster, more effective, more controllable than just handing over a bunch of documents and crossing your fingers.

Michael: Yeah, that I noticed some of the tools that I've played around with have started to have things like input information about your brand. So it builds guardrails around it, to say, oh, you use your default style is usually this, it's conversational or it's, and whatever that seems like that works that, and I assume that just augments the prompts that I'm putting in.

One of the things in the survey that I did earlier this month about AI adoption, one, one of the questions that I asked was are you using a publicly available, open model? Are you using your own proprietary model? Are you doing some sort of around it? And it was pretty mixed. Which I think, people are doing lots of different things, but one of the, one of the use cases that I've heard a lot is, oh, I want to use, one of the large, like ChatGPT or Bard or Claude or something. But I wanna use my data too. So how do you do that? And how do you do that safely?

'cause obviously you don't wanna train something publicly on things you don't want public, but at the same time, having your data involved is an important part of the project.

Nathaniel: That's a really good question. It's the question of the hour as I think we've been seeing and there's a couple of different answers. So I, we can go through maybe, three phases. Phase number one is, most of these tools have pretty good privacy agreements with you. If you use open AI's, API they won't train based on your data. OpenAI just released enterpriseGPT. Yeah, I believe, or and they say they won't train on your data so you can trust them with that.

That's maybe step one. Step two as we've been seeing is don't really train them on your data, but use something like pick a's document interrogation feature that basically pulls in just the relevant bits of your data right when they're needed and injects them into the prompt to make to give the impression that this bot has full knowledge of everything that you're doing.

But then the real cutting edge that we're seeing with step three. Just last night actually, we launched our pickax model that we built ourselves on open source technology. But that, we're able to completely own and license and run ourselves. Is that very soon, within the next three or four months?

ChatGPT quality results are going to be easily obtainable from models that are completely self-hosted. And so you are not gonna have to go to OpenAI or Bard to do that. And, we want to, we are becoming a provider for spinning up and training those models and making sure that they're bespoke and tailored to an individual company.

But I think we're gonna be really wowed at the speed with which we forget about ChatGPT once the proliferation of these models starts to occur.

Michael: That, that makes sense because obviously if I'm doing something. Say, I wanna put a chatbot on my on my research website, and I want it to answer questions about my blog and my research reports. That seems like a really good use case for it. You could have, people could ask questions or interrogate or whatever. But I'm, but I would be, I would be concerned about the data, but that makes a lot of sense that I could actually contain it into my world of knowledge and build it off of that.

Which leads me to another part of this 'cause I know you are, some of this some of the way people interact with pickaxe just come on the site and do a project for yourself. But you're also doing some really interesting projects. And we talked a little bit about a few of them when we were prepping for this. So I'm just curious if you could share a few of those use cases for, for businesses to just help put it in context for people. 'cause we, we hear so much about generative AI and ChatGPT and it's supposed to save the world and, all those good things and yet I don't, it seems like we're a little bit light but you could do this thing with it and some of the shows I've done, like I did one for marketing and that was really interesting. There's a lot of use cases there but what are some of the things you guys are seeing

Mike: totally? So for us it's very important to provide like the best no-code solutions and to do that, we have to stay at the edge of this technology. So we're constantly working with businesses to do very difficult things for them. With AI solutions that we can then, bring back as features into our platform. And we can maybe give you a couple examples, a more hardcore business.

One is there is like an auction house and they have, we can't really say which one, but there's an auction house. They're pulling data from all sorts of places, like 15, 20 places. So all the data they have about their auction items are formatted differently, and they need a system that can take in badly formatted dissimilar info and always guaranteed like 99.99% success rate, spit out a uniformly formatted JSON object that has all the information that they can just drop on their website. And then the item information appears. So just to, I don't know if that was like too much, but basically, no, it's messy information, but they want every of these millions of items they have, they want all of 'em to appear on their website the same.

Nathaniel: And just on the note of like, why would a company use Pickaxe for that when there are companies out there that for years, that have done similar types of data processing work and have specialized in that space. The reason is because those companies are behind on what these new solutions can do. So you need a no-code builder where you can work by putting pieces together to build the solution of the modern age. As opposed to using these kind of solutions that have existed for some time.

Mike: So these large segment models are very smart and can handle really bad data, basically in a way that other stuff couldn't so basically by training a model on millions of examples, you can then guarantee you could have this thing that just is outputting all these samples. And then one that's maybe a little bit more fun is we've also been working with some screenwriting, some production companies to help, augment the early stages of screenwriting, so thing called coverage where they read tons of books and the need to basically write a book report about it and talk about how it would look if it was turned into a script. So basically we've started to automate that process for them so they can more quickly cover material and then get writers working on the scripts faster.

Michael: Yeah, that's interesting. The first use case I thought was really relevant from an e-commerce perspective because I think. And there's some subtleties in that, right? You can get an e-commerce platform, it can do some similar-ish things. But what you're saying is, if I understood this correctly, that you can take all this unstructured data and turn that into a repeatable format that's gonna always deliver the way you want things to be displayed on your website, whether it's merchandise or whatever.

Yes.

Mike: Remember the reasonably smart college intern if you can teach a college reasonably smart college intern, how to look at data and then create like a simple JSON object of 10 fields always outta the data. Then you can teach a model how to do it.

Michael: Yeah. All right. That actually that's maybe one of the better ways for people to think about this too, is think about it at the level of a good college intern that you could that you could track.

Mike: Yeah. It's more complicated than just, mundane, repeatable work. It requires a little bit of brain power, but these GPT models have enough brain power to do that.

Michael: Yeah, no, that's great. You know what? Take this out of the, what you've done and just let's just think about practical application at some level. What is the potential for businesses today? What could they do with it today versus, I don't know, some of the hype seems like it might be a bit of in the future it would do, which is fine, but if I'm a business person and I wanna solve problems, what are some of the top business challenges that companies can address with generative ai?

Nathaniel: It's a good question. I think that people, we go through this hype cycle. People say, hey, I think maybe this is over hyped, but it's not. It's not gonna take everybody's job in the entire economy, but it will start to be sewn into the fabric of many of the different software tools that we use.

If we're talking about what the earliest things that are gonna hopefully go away are. This is something we talk about a lot internally. It's anything that people don't wanna write and people don't wanna read. So in business and in life, there's tons of documents that people either don't want be writing or other people don't wanna be reading.

One of the big organizations that uses Pickaxe in a kind of a self-serve capacity are college essay coaching services who are, helping people to. Fine tune and work to generate better college essays. And I think it's only a matter of time before we start getting universities who want college essay reading services.

So that no human is involved at any stage of that process, or at least not in the initial filtering. So to answer your question if you are in business and you're doing a task where you're like, I don't know, why I need to write in this weird way or in such detail. A customer of ours is in healthcare and they are they are using this to help folks to write medical reports that would be really time consuming otherwise. So anytime when you're in business and you're doing that you can just think this. Can be automated in the next few months. Really?

Michael: I see, you got me excited now. 'cause I want you to, I want you to write one that goes through my email every day and deletes all of it except for those two things I actually needed. That'll be, talk about saving time. No, the other question that comes up for businesses I think right now is, Should I jump in? Should I do something now? And people, individuals too. Not just for companies. This is an interesting question for everybody.

Do you need to know, should you learn now? When's the right time to get in? And what do you tell, because you talked to a lot of prospects and customers. What do you tell them when is it time now or is it not quite there,

Mike: It’s not the time to go full ai. That's a little crazy, but it's absolutely the time to start experimenting with this, at least internally, to see what you can where you can affect your business or you can be more effective, where you can work faster.

So it's absolutely time to spend, invest some time and reasonable resources into exploring it. But also the space is moving so quickly. I'd be a little hesitant to build like a full AI stack or bring in a full research AI research team. Which is really where pickax shines. It's a pretty easy way to like quickly prototype some AI workflows, right?

You can quickly spin up tools start connecting them, maybe even dabble in like training a model without having to make your whole system rely on it. A lot of our customers really are experimenting with AI workflows via pickax.

Nathaniel: We see folks. Are they realize that their competitors are all using ChatGPT. They understand that this AI is changing the way that they do business. They don't wanna spin up a whole AI research team. Pickaxe is the low barrier to entry point for, starting to see where in your business this technology can be helpful and where maybe it's it won't be helpful yet for another couple of years.

Mike: And I'd say the limit a lot of people hit is that ChatGPT is a generic model for general purpose applications. So it seems really cool when you first get in there and it writes a blog post immediately, but you realize it's only about 80% of what you need. And in the business world, you can't be getting B minuses on your jobs. You need to be, you need to be getting 90% or higher. So really businesses should just be exploring how to achieve that 90% success rate or higher. And, prompt templating or prompts, chaining, these sorts of things are probably the lowest hanging fruit to achieving that.

Michael: Yeah, that makes sense. I know for some of the other folks that I've had on when I asked, 'cause I almost always ask this question, when's the right time? One of the things that came up and I know there's a lot of fear around jobs, and you mentioned it already too it's oh no what jobs are gonna go away? And actually even in the survey when I ask it, it was interesting because more people thought that there were gonna be a lot of new jobs created around it then necessarily thought that it's gonna automate, lots of stuff away.

So I think that's definitely interesting. But from an individual standpoint, do you do business people today do you need to understand and be at least dabbling with generative AI right now? Is it important from a career perspective, do you think?

Nathaniel: Look, we're biased here, but we would say 150%. Absolutely. I mean that the customers that we work with are primarily early adopters, if we're gonna be honest here these are the people that are always dedicating some time and some resources to looking at whatever the new thing is gonna be, but even if you don't want to use it internally within your organization, you need to spend time figuring out how people are using it to understand what your competitors are doing, to understand when. You are getting a business deal and the contract seems strangely like it's been written with some AI tool. There's a recent case of a lawyer who I think was dis barred because he used this ai. So it's going to be a for every professional working in America, it's gonna be essential to understand the type of ways these technologies work at a high level.

Michael: I think that makes a lot of sense. I will say that attorney, that case I've used a few times, 'cause I think it's really interesting that he generated an entire brief and filed it with all the citations and a hundred percent of the citations were made up.

Crazy. Yeah, I know. It's not always perfect. There are issues there, right?

Mike: We're in this space. So right now we're training a model and we're trying to basically the technology people are using to do it and the techniques are changing week by week, day by day almost. There are a lot of people, like for my uncle has a substack and writes a lot about culture and music, and he is against ai or he wrote a piece about how, oh, the verdict is out, AI is not that useful. And it's true that AI has a lot of limitations right now, but it's the worst it's ever gonna be today, pretty much. And it's only gonna get better. And there are people all over the internet, all over the world working really hard on improving it. And then they're sharing all their, a lot of their research online in like subreddits and YouTube videos, these sorts of things. So it's improving like an astonishing rate, I would say.

Mike: So it's not super helpful for you today, it very well could be even next week.

Nathaniel: Honestly, if you read the that Google piece that I sent you, and maybe we could even link it in the podcast here they go over how they're basically saying, Hey, we are gonna fall behind.

Everybody's gonna fall behind. Open source community is moving so fast, they're even outpacing us, this giant billion dollar company with the most resources of any organization, pretty much in the world, as far as AI development. And that's just a testament to how quickly things are getting better.

Michael: Yeah. I was I was writing something the other day and I, and as a part of it, I ended up going back and reading the Gary Kasparov chess game story against IBM's deep blue in the nineties. And it was funny 'cause I, my memory of it was very different than I think once I went back and I researched it. What I learned was the Kasparov lost six games in a row. He was very upset by it at first, but, pretty soon after he started to realize that, hey, here's an opportunity you could actually change the game and do this sort of human machine, combination and it'd be great. And they'd be better than just a person or a machine.

And for a while that was true. But in the mid. 2000 or 2010 ish or so by then, everything had advanced to the point where that wasn't true anymore. Putting the human in the mix with the machine made the machine badder worse than it would've been if it was on its own and it a hundred percent of the time beat every, anybody that was playing chess. So my point I guess is that you being in early and learning it is important, but also realize this is gonna continue to evolve. So you can't just learn a thing and you're done. This is gonna be an evolutionary process for all of us. Would you agree with that?

Nathaniel: Absolutely. And the funny thing about chess versus the world economy is that, the rules are constantly changing in the world economy and it's made of people at the end of the day, the AIs don't currently have money and buy, for the most part, people are the ones that we're trying to provide goods and services for so at the end of the day, it's not going to, at least from our perspective go in the direction of fully automating out in any kind of a way, but working with ai. It's something that people are, going to have to get more used to.

Michael: Yeah, I mean that, that makes a lot of sense to me. Now I wanna we're getting close to time on this, but there are a couple other things that I really wanted to make sure covered. And one of them is there are a lot of No-code platforms popping up. There are a lot of startups doing AI no-code platforms of some sort. And so I'm curious for you guys, 'cause you have a set of capabilities you've built out. You're adding to 'em, you got your roadmap. What sets Pickaxe apart from some of these other no-Code platforms?

Mike: Totally. I think probably just where we focus and what our emphasis is on. So our emphasis is very much on getting you the best results possible. Basically you might say prompt improvement, results improvement. Like we talked to a ton of our customers. A lot of 'em want this tiny feature, this tiny integration or this, but the thing all of 'em are curious about is like, how do they actually make their prompts better? How do they get better results? These things. So that's our builder's a pretty good place for that and pretty much all of our future features are focused on that. Like the reason we're training models right now is to figure out how to take, let people give feedback on their responses and basically train the prompt, train a model to do what they wanna do. We view a future where people are gonna want to deploy these things and they're gonna have very strong opinions on, this is good, this is bad, this is good, this is bad. And we wanna be able to reinforce the model in that. So again, like ChatGPT, very generic purpose thing. A lot of these no code platforms, they only plug into generic models. Very much our focus is how can we give you your personal area and research model, right? As opposed results, that sort of thing.

Michael: Yeah, that makes sense. 'cause I think eventually that if you're a business, that's, you're gonna need that, right? You're gonna need it tailored to specifically what you're trying to do. Like ChatGPT might be great for a marketer that's, a content marketer that wants some ideas about what to write about or way to improve a little. But you're not gonna get the same thing that you get if you have something that understands your branding and your brand and your data and customers and all those things. Yeah, that makes a lot of sense.

Mike: And the other thing I might throw in there is I think we are easier to use, maybe I'm biased, but that's what a lot of people tell us, a lot of people come to us after trying another no-code AI solutions, but we're very easy to use and very easy to embed. So if after you make your app, you can, your website or your, share your dashboard in about five seconds. So it's just fast and easy.

Michael: That makes sense. And today that's obviously we're all learning and so that's, I think that's really important to the simplicity and making it so that it's easier for people to get their hands in because that's how they're gonna learn. Yeah, that makes a lot of sense. So what, where's Pickaxe going? What do you guys see two years from now, five years from now? What would you like to have? Besides. Selling a lot of subscriptions and making a lot of money, which where's it headed?

Nathaniel: What we have really enjoyed is working with smaller organizations. We started working with just solopreneurs and individual consultants, and now our largest customers are still organizations of no more than, 20 people actually onboarded onto the product. And we want to continue to focus on that and help these small organizations act like they have 10 times as many people as they as they do. As they transact their business at the present moment, our consumer and like marketed to the public product is more about prompt engineering and prompt improvement. And we have a document interrogation feature that kind of allows you to get a little bit more of your information in there.

But our goal for the next year is to completely flesh out this smooth transition between coming in with. Nothing no training data. You don't have an ML team starting with prompt engineering and explaining, adding in your own documents and then transitioning to a completely custom model that you know is basically owned by you and and only provided through us.

So getting people from zero to completely custom bespoke and amazing model for them in a no code way. Is the product that we are building over the next year, and we hope to be selling that to hundreds of organizations over the next year so that they can all take that journey with us together.

Michael: To me that, that seems like the biggest challenge for a lot of people is the base level of knowledge. You need to write a good prompt to get the things you want out of the model that you're, that you're in that, that you intend. And to make that simpler for a small business, that has to be a great force multiplier for them. So that really makes a lot of sense. So that's all the time we have today. So first of all, Nathaniel, Mike, thank you so much for joining. I really appreciate, it's very interesting what you guys are doing and I said, just did set up an account and I'm gonna put my hands in there and do some work on it myself.

But before I let you go, one question I like to ask is could you recommend someone, thought leader, an author, or some mentor that's influenced your career that you'd like to share with people so they can, take advantage of that?

Mike: For me, the one that comes to mind is just what I've been reading recently is probably Charlie Munger, very smart guy, always learning 90 in his, he's 93, 94, still learning a lot. So his advice is just keep it simple, stupid and focus on what you can be really good at. So yeah, if you have, if you don't check out his interviews on YouTube. Read the Tao of Charlie Munger. Really infectiously cool guy.

Michael: Yeah, I'll check it out. That's great. It gives me something to shoot for 93.

That's impressive. How about you, Nathaniel? Anybody you want to add in there?

Nathaniel: I guess I would say that reading Paul Graham is a standard answer for a lot of people, but he's another big proponent of keeping it simple and paying attention to what matters. There are a lot of mentors for Pickaxe that mentor us interpersonally, and I would love to talk about them but in terms of what you could actually read or anybody could actually read Paul Graham, I think is the number one

Michael: Great recommendation. Thank you.

Mike: one other there, maybe Cobus Greyling has a medium, talks about how to get good results from LLMs, how to. Really great. I would check him out.

Michael: Yeah. Definitely worth checking out. Thank you. So thanks again for everyone for joining us today. Thank you Nathaniel, Mike for sharing with us. Just a reminder, remember to hit that subscribe button before you leave today. And for more on ai, you can check out on the Arion Research website. We published a report a couple weeks ago on AI adoption that's based on a survey from this month. So it's very current, which frankly, you have to be in, in the world of ai or you're outdated. And it's a free download Who doesn't like free research? And then next week join us for a discussion of AI and intelligent chatbots.

I'm Michael Fauscette and this is the Disambiguation Podcast.