Jesse Clark:
There's kind of moments in your life where you're sort of reflecting on what is it that I'm sort of doing? And then you're like, hang on a minute, there's opportunity here. Life is kind of short. I've got to kind of take the leap here.

Eric Anderson:
This is Contributor, a podcast telling the stories behind the best open source projects and the communities that make them. I'm Eric Anderson. Jesse Clark is a co-founder of Marqo AI and joins us today from Melbourne, Australia. I am joining you from my home where we are having a piano practice in the background. If you hear anything, that's what that is. Marqo is a multimodal vector search engine. Jesse will tell us all about its horizontal scalability. They've got a fancy query DSL that helps with pre-filtering, has some search highlighting features. It's quickly grown to almost 3000 GitHub stars, yet being only less than a year old. Jesse, did you anticipate this amount of enthusiasm for vector search and its ilk when you set out to start Marqo?

Jesse Clark:
Of course I did. No, I didn't. No, I did not probably quite expect it to sort of reach the point it has now. I think we knew, certainly when we were thinking about Marqo and starting it, that there was a huge opportunity around vector search and machine learning based search. But I think since Stable Diffusion came out with the generative AI, and then ChatGPT, the role of vector search, vector databases has just kind of exploded. So it's been, I think really fortuitous timing. And yeah, it's really exciting to be kind of really amongst it.

Eric Anderson:
I imagine when anybody starts a new effort company project, they wonder, is anybody even going to care? Maybe it's just going to be crickets. And in fact, the whole world attention seems to be on just what you're doing, so kudos for your prognostication.

Jesse Clark:
Thank you. Yeah, I think you've sort of got to get through that, I think. And sort eventually, you get to a point where you've got something that is usable and people really can start to actually create value and make things out of it. And I think that's... You just sort of got to have enough conviction to get to the point where you can build something that is usable, take it from that kind of idea to implementation, and actually then becomes tangible. So getting through that barrier, I think is really key.

Eric Anderson:
So I want to get into your story, but first I want the audience to have some context for what you do and what we're discussing. And so we should talk briefly about the project. And maybe the best way to do that is... I imagine some people come to this conversation with some context on the industry. Maybe you could disambiguate for us kind of what Marqo does versus other vector search or vector database solutions.

Jesse Clark:
Yeah, absolutely. So sort of succinctly, Marqo is end-to-end vector search engine. And sort of in that name, I think it sort of encapsulates a lot of what we're trying to do. We really want it to be end to end. So it's documents in. It's documents out. It's very developer friendly. A lot of the abstractions, the transformations, the machine learning inference, the vector storage is really taken care of on behalf of the users so that they can really get started on building the actual search applications. And so yeah, really focusing on being able to actually provide this kind of end-to-end solution. And it works, again, then across modalities. So we have... You can index text. You can search across text. You can search across images. You can extend that to videos, and then the abstraction is very easily taken over to audio and things like that.
And so really about bringing this end-to-end vector search experience to developers, couple of lines of code. And then certainly, the other focuses are... I think we had, from the start, really about production workloads and really looking at real time search because this is also incredibly tricky from the sort of production engineering, cloud deployment side of things. So not only do you have the kind of decisions to make around the machine learning and the inference and all those kind of design decisions. It's about how do you actually make it scalable, real time search, low latency. So that's also been a big focus for us.

Eric Anderson:
Yeah. Maybe to repeat some of what you said, so if somebody wanted to have vector search, they're going to need a place to store document, a database, a place to store embeddings, another kind of database, and they're going to need a framework, a functions and a programming model to kind of tie these things together. You offer all those things. That's the end-to-end solution.

Jesse Clark:
Exactly. Yeah. There's a few different pieces to make the kind of end-to-end vector search engine work. And so one of the critical pieces is this vector database, which we've heard a lot about, sort of takes in vectors or embeddings. You can sort of store them, and then you can do this kind of fast lookup. But that's sort of the quite primitive operation that kind of powers a vector search engine. You've got to have the generation of the embeddings. People have metadata. They want to do keyword search. You want to do filtering over that as well. So all of these other functionalities are also required to make it the sort of production system. It's not just this kind of simple vectors in search for them as quite a lot more to actually make it useful. And that's what Marqo was aiming to solve. So we take care of the vector storage for you, all of the transformations, and then all of that kind of just bookkeeping that's required as well for how do you actually format the data, keep the metadata. We want to have consistency, asset transactions, that kind of stuff.
So it's really trying to make it this production system that can really make developers just get up and going very quickly.

Eric Anderson:
Well good. We'll circle back to the product and projects and more and its capabilities in a moment. Now that we've got that context, what brought you to this at this point in time? You have a long history in machine learning, and maybe that helped you to see around the corner and realize the need for this.

Jesse Clark:
Yeah, I can... Yeah, sort of where do I start? I think, yeah, I've had a reasonable history in machine learning, although it felt like I was late when I started, but now I feel like I was probably early. Yeah, maybe it was good insight. It could have been good timing. It could have been a combination. But yeah, my background originally was in physics. I sort of was quite enamored with physics when I was growing up in the Australian country and used to look at... There was no light pollution, you could look at it at the stars and it was incredibly clean sky with all these galaxies. And that was sort of where I was like, hang on a minute, I think I want to try and understand that a bit more. So yeah, I went to university, studied space science, maths and physics, and then fell into a PhD into an area that I wasn't so familiar with, physics. But that was where I started to fall in love, I think, with computers and actually this kind of computational analysis and numerical computing and optimization algorithms, which is just the foundations of machine learning.
And so that was where I also sort of started, I think as well in sort of dabbling in sort of open source, although I didn't quite realize it at the time. I was developing after my PhD, developed some software to help other people do analysis for the stuff that we were doing. We were collecting huge amounts of data, very much a big data kind of problem. We had petabytes of data. And this was probably 15 years ago, we used to carry suitcases of hard drives back from experiments. It was rough and ready. We had just incredible amounts of data. And so yeah, started building out programs to help other people analyze their data. And it was interesting in our kind of field, that sort of people were very protective about their algorithms and their code because it was a competitive advantage from a kind of publishing and experimental point of view, which was quite interesting, because the motivation for doing a lot of the science was kind of to be open and actually be able to share these kind of innovations.
And so that was where I got a flavor for it after my PhD, was to do this open source development. And then, yeah, I stayed in physics for a while. I was living overseas, was in London, and then the US. And then it was around 2013, I was living in Palo Alto. I was working at Stanford doing physics research. And then sort of looking around, I was literally walking down University Avenue, and it's like, hang on a minute. There's this huge kind of opportunity here. There's a lot of data science coming out. It was really starting to become quite popular and impactful, and I think people started to see a lot of value in these kind of data analysis, the early stages of machine learning and how that could impact businesses. And so I was in the best place at the best time. I felt really recognizing that and sort of decided to leave my career in physics, which was going quite well.
And I joined Stitch Fix at the time. It was going through rapid growth. It was really innovative in terms of how they were using data. Eric Colson and Katrina Lake had built out a big algorithms organization. There's a lot of physics PhDs there. It's a really good mix of people. And so that was great kind of introduction. That's where I started a lot of my industry-based machine learning. And then went from there, spent time at Amazon four and a half years, did a mix of things, working robotics, doing a lot of intelligence development for item manipulation robots and computer vision as well there. And then the sort of final piece that I think really spurred me on to start, Marqo was working a bit in visual search. So this was around looking for using multimodal search effectively, like images and text. And shopping on amazon.com or any other...
Not to necessarily call out Amazon, but other retailers, it's like I can search for a t-shirt, and that's great and I can find one. But then there was this problem. It's like, I want that t-shirt, but I want it with a pocket, or I want it to have long sleeves or stripes. And this kind of navigating a catalog like that just wasn't possible. You have these kind of rigid taxonomies. You can't cut across it in these kind of semantic directions, which mean a lot, I think, to users. And so that's where these ideas started to come about that. Hang on a minute, there's this opportunity to do something kind of more. And we've seen some innovations around multimodal representations, so being able to sort of image and text living in the sort of same space. And so you could query texts with images, images with text. And so yeah, that was sort of the precipitous for Marqo and where it all started.

Eric Anderson:
Before we get back to Vector search, your background, Stitch Fix is seen as premier machine learning institution, maybe surprising to some, and Amazon similarly. I'm curious, do you learn more in industry or in academia if you want to be cutting edge in this field?

Jesse Clark:
Yeah, it's a good question, something actually I've thought a lot about. I think they have different strengths. The short answer is the rubber meets the road, and sort of the money is generated from production systems in industry. And so I think that's a really good opportunity to learn because that's just where it happens. And when you put stuff into production, there's really no hiding. It works or it doesn't work. You AB test it. You do other testing. It's very clear. There's no kind of hiding. Whereas it felt like in academia doing research, you could always kick the can down the road. This hard problem, which we couldn't solve now, we can come back to that later. But certainly in industry, there's no real... A lot of that, if you've got an MVP out, it's still got to create some kind of value to...

Eric Anderson:
You just got to ship something.

Jesse Clark:
You got to ship something. So I think industry's a great place. And I think certainly, as everything's evolved, a lot more of the flavor of academia is crept into parts of industry, so it's got a really good mix, at least from my perspective about some of the academic kind of learnings, but it's about really making things work and sort of iterating fast.

Eric Anderson:
So you launched Marqo soon after Dall-E, and maybe similar to when Stability was happening and Midjourney. You launched as being multimodal and image was everything. And now, we're seeing we're kind of living through a wave of natural language, it seems. And maybe these things, we're looking at too close zoom. And if we take a step back a couple years, we'll realize that everything was moving at the same pace. Is this a language thing, an image thing, or is it everything? And are they moving at consistent paces, the two of them?

Jesse Clark:
I think so. I think there's been a huge crosspollination, I think as well, between the language and the images. Yeah, the language has come back. I think the future is multimodal. I don't think that's going to change. I think we can see just how... Having this additional grounding from the other signals sort of... I think we've seen is just so powerful. Do you need four or five modalities? I don't know, but I think certainly greater than one. Yeah, images is... I've always had also a love with images. I've been somewhat biased with them, I think from my sort of earlier days. I think it ebbs and flows, but they're moving both very fast. I think it just depends on where the kind of spotlight is a little bit in the media as well. You look at the innovation still in the image generation, it's still absolutely incredible. And some of the recent ones around just reconstructions from 2D views, Nerf, all of this kind of stuff from the image domain is just still absolutely speeding ahead.
So yeah, I think they're both just moving really fast. I don't think that... It's hard to pick one, I think.

Eric Anderson:
Yeah. Yeah. Yeah. Circling back to Marqo, and we should throw in elements of your story in the last year, but I want to float some use cases by you. One use case, which I think is the bread and butter for Marqo, is like I have a bunch of documents that I want to search. And the state of the art from a year ago is I put it in elastic or something and I get keywords, but I miss out on a lot of similarity and meaning and context rich searching. And maybe one of the hard parts of doing context rich searching has been the variable length of documents, and what exactly do you return? And Marqo helps you kind of... It has an opinion and a method for serving your results. Is this the bread and butter? Do I understand that correctly?

Jesse Clark:
Yeah. I think the text-based search, and particularly for vector search, is still a huge... It's a huge use case. And absolutely, we see a lot of users and a lot of customers using it for this way. And like you said, again... And I think this is where the kind of end-to-end vector search engine comes in because there's a lot of decisions that you need to make about how actually... The models themselves have a sort of fixed context length for the most part. And that means how much you can kind of... It can see at any one time. And depending on the model, this can be quite small. It can be quite large. And that's one of the big innovations, I think we've seen with the GPT [inaudible 00:13:09] kind of recent innovations, is these large context lengths from these large language models enables these really interesting use cases.
But it's the same kind of context length, exactly the same that we see with these embedding models. And so you've got a fixed context length. So it's sort of how much text at any one time can it kind of understand? And so yeah, we make decisions about how we can automate that kind of process. And so you can take long documents and we can automatically break it up into smaller ones. You can create embeddings and vectors from those smaller chunks, and then you can search over all of them. So then even a document itself can return just a subpart, which is effectively the highlight. And so becomes incredibly convenient for a user to just insert documents. And then when they search and return, not only do they get the relevant document, but they get the relevant highlights. There's also lots of other ways that you can use this kind of paradigm to improve retrieval performance as well.

Eric Anderson:
What about the use case of fine tuning an existing model? I think people are using some vector databases and resources for that. Does Marqo play a role there?

Jesse Clark:
Yeah, absolutely. I think what we've seen is the sort of rise of foundation models or these zero shot models that are really good at general tasks, right? They're kind of been trained on a large corpus of data, and then they sort of generalize reasonably well to kind of most domains, specialist domains. But then even what people find is that's a great way to start, and they can get very good performance. But then almost always, everyone wants kind of better performance. And so that's where the fine tuning comes in the domain specific kind of modeling. And yeah, and absolutely Marqo really encourages that. And you can load in your custom models, and then you can get going on your own. Exactly the same thing. You just configure the settings, you put in the custom model, and off you go, documents in, documents out. But again, now it's your fine tuned model.

Eric Anderson:
So I guess this is just my opportunity to clarify things on my end. In the first scenario we talked about, you mentioned an embedding model. So in that instance, I don't need a ChatGPT or a GPT4 type model. I just need a model that's going to convert my documents into embeddings. Marqo wouldn't offer me that. I bring my own embedding model. I'll bring it to Marqo, and documents in, documents out.

Jesse Clark:
Yeah. So both, actually. So we have a lot of default models, sort of a registry that you can choose from to get started. Because yeah, like I said, there's a lot of really good models now that have been contributed that are open source that allow people to get really started with pretty good results from day one. And so we support all of those models. In fact, I think... And certainly the state of the art, most of those are out of open source at the moment. So we support all of those kind of models so people can get started and really try and build that MVP, right? I think, do I need to find tune model? I don't know. I've got to build my kind of... There's a whole lot to go before you go down that kind of path. And so yeah, really lowering that barrier to entry, and then they can assess it, right?
It's like, hey, it's failing in these kind of areas. This is actually where I need to spend the time in fine tuning it. And that can also really cut down your iteration cycles by getting that feedback and just seeing where are the challenging parts and tuning the model to get over that.

Eric Anderson:
While we're on the topic of use cases, I imagine almost any organization in the world could benefit from Marqo, and I also imagine that few organizations know exactly that thing that would benefit them, what exactly they're looking for. I imagine you're left with the challenge of helping people map their general interests and curiosity to a, here's use case one, or two, or three. And is that something the industry is solving for people, or is that something you're having to block and tackle on your own?

Jesse Clark:
Yeah, absolutely. I think it's a bit of both. I think the education piece, I think it's been incredibly... Yeah, the recent wave of innovation in AI has been incredibly beneficial from that kind of educational piece because people's awareness of the sort of concepts around artificial intelligence, machine learning are much better than what they were say a year ago. And that includes things like vector databases and vector stores. But I think as well, we're sort of trying to really join the dots and help people with that self-discovery. Like you said, I've got a problem, and then it's like, hang on a minute, that's my problem. I can use this to solve that. And so I think that's actually been incredibly... It's been such a huge part, actually, for driving people to adopt it is actually through this kind of content generation that we do in the blog and which we really want to keep doing. It serves many good use cases. You can really just point people to these kind of examples, these end-to-end examples, and this self-discovery has really help.
So it's definitely a combination of, yeah, the industry is really helping a lot, and then certainly just actually putting in the effort and really trying to generate high quality content, I think as well, to help users do that self-discovery process.

Eric Anderson:
Maybe we go back to the story some. You had this idea to leave into your jobs to go and kind of pursue this. What were the initial steps? And how soon maybe did this become open source? And what was that kind of decision process like?

Jesse Clark:
I think it was just sort of something that was burning in my mind, and I was looking around. Also, there's kind of moments in your life where you're sort of reflecting on, what is it that I'm sort of doing? And then you're like, "Hang on a minute, there's opportunity here. Life is kind of short. I've got to kind of take the leap here." And so that was what happened. And I think it was really... I was sort of poking around trying to think how do I... I had these ideas about vector search, and particularly multimodal search. And it's sort of very much knew the machine learning, but it's like, hang on a minute, how do you build a managed service around this? How do you kind of actually build a sort of company? And so I sort of started to speak to a few people I knew in the industry to sort of scope it out. What's the kind of process here?
And that was kind of chugging along. But then I got introduced to my co-founder Tom through a mutual friend from Amazon actually as well. And that's kind of where everything kind of really happened. So Tom had been trying to been developing his own startup as well and was looking at sort of serverless elastic search and databases, and really looking at managed cloud and how to make that easy for developers and was looking to leverage a lot of the vector search as well. And so I was looking into, had a lot of the vector search and was looking to enable the cloud, and how do we actually do that? And so that's kind of where we got together, and it was just a really good, I think match in terms of what we wanted to do and our kind of skill sets we could cover off on the two big components.
And so yeah, we sort of just iterated on the ideas, I think, and then we're able to get some early stage funding, and sort of start from there and quit. I sort of decided to just... This is it. I'm all in, right? I'm leaving my job and I'm going to build Marqo into something that's incredibly, I think beneficial and useful for people. So that was kind of... Yeah, that's how it started. And I think in terms of the open source, it wasn't the first thing that we kind of decided, but it became really natural evolution.
I can't remember what it was, but we were sort of thinking about it and how we're going to develop Marqo, and then just became evident that actually open sourcing it was going to be a really good option and that there was a lot of benefit to actually doing this in terms of just not only as a contribution, beneficiary of open source for so many years, not just the contribution kind of back, but you can get very early feedback, right? Launch early features in this fast iterative loop, just fast feedback. I think that's kind of also a huge benefit.

Eric Anderson:
We talked a bit before we started recording today about the language people are using to describe this field. I think tensors, embeddings, vectors, are these all the same word? And I think Marqo has described itself as tensor search in the past, vector search some. How do you keep up with where the marketing is headed?

Jesse Clark:
Yeah, it's a bit of a hot topic, and it's somewhat contentious. The use of tensors in machine learning, I think now, it's got a reasonably well established kind of terminology, effectively is used as a placeholder for multidimensional data, multidimensional arrays. But from the sort of pure physics point of view, this is somewhat incorrect. And so I think that can certainly rile a few feathers. But I think at this point, machine learning has adopted it to mean this kind of higher order abstraction of vectors and matrices. That's sort of how it's been represented. And so that was the spirit that we were taking it in. And so instead of just a vector... And the idea, as well as that the information that we have is... It's really rich. Documents aren't just a single representation, particularly long documents. And so actually having multiple vectors, there's much richer representations, much higher dimensional. This kind of tensor based representations made a lot of sense.
And so that's where the idea from this sort of tensor search language came from, really trying to show that it's this kind of slightly more generalization of this vector search and that you could... In fact, you don't even need to have... It can be different modalities. You can imagine a video. You can have embeddings and vectors for different parts of the video documents. You can have for different parts. Even an image, you can chunk it up and have different vectors for different parts of the image. And so there's a lot of... And even in queries, you can have multiple vectors for a query. You don't have to have just a single vector for a query. And so these representations and queries that just made a lot more sense, I think, to start talking it in this tensor based way.

Eric Anderson:
Whenever somebody's kind of described embeddings to me, it sounds like this abstraction from our language. Our words have meaning, and these learning represents kind of concepts or ideas, and embeddings have this kind of mathematical representation of concepts and ideas. Does this mean that embeddings are kind of a universal language? And are those ideas and concepts... They're not standard, right? Your embeddings don't look like my embeddings, and so they don't mean the same thing, or do they?

Jesse Clark:
Yeah. It's a good question, actually. They probably don't right now, but maybe they do in the future. Yeah, the embeddings, yeah, like you said, are able to... They sort of capture a lot of meaning and they represent... It's a really nice interface basically between sort of rough unstructured data and then something that's kind of well organized. And that's really a key point of the whole embedding in vector representation, is that you can take this rough and ready data and put it down into this kind of common interface, which allows you then to have all sorts of other operations over that because you can kind of standardize things. But yeah, exactly, the different models trained on different data will have different kind of representations of their internal space. Different concepts get mapped to different parts depending on what the data is. Some concepts won't even be present or won't have really any good understanding of those, again, depending on the data.
And so yeah, I think at this stage, depending... Again, it sort of depends, but you're embeddings might be the same as my embeddings. It depends a lot on the data, and how the models are trained. But certainly, I think we can also see that there's a lot of foundation models that have somewhat universal representations. But I think the same with language, right? It's very nuanced. It's hard to have kind of universal descriptions for things. A lot of it depends on context, I think as well.

Eric Anderson:
Well, I guess presumably, these embeddings describe relationships between concepts. And so if your data covers different concepts than my data, then we won't have overlap. But if it's similar data, they'll probably arrive at similar conceptual nodes, and then they would have cousin concepts.

Jesse Clark:
Yeah. Indeed. And I think that's seen as well with some of these other models where you actually take embedding models from two different domains, and now you can actually learn as a joint representation with relatively few parameters or little effort. So you can take a pre-trained text embedding model. You can take a pre-trained vision model. They've both got embeddings. And then you can use the right data. You can actually start to map these to the same space. So yeah, certainly, we're seeing a lot more of that. I think we'll see a lot more of that, because certainly this composability as well that we've seen in machine learning. So if you can train a text model in isolation just on text data, you don't need this really paired data, for example, and you contain a vision model on its own sort of vision data and do that at scale, and then you can sort of use a smaller representative set to join them together. I think... And then actually switch these models out.
If you've got a better text model, you can switch it out later. I think this is actually what we're seeing as well as being a really powerful kind of paradigm.

Eric Anderson:
The use cases I brought up were mostly text use cases. And I think it would be great to hear from you, some of these multimodal use cases. Image search seems like an obvious kind of text, return images, and maybe image similarity search, like here's an image, can you find me similar images? What are some things that people probably don't realize they could benefit from a multimodal search?

Jesse Clark:
Yeah, I think it's a huge, huge area of applications. And an image is worth a thousand words kind of thing. So there's a huge amount of data that is also contained in them. But again, I think this combination of text and images is really powerful. We're seeing, I think a lot of opportunity in catalog search, e-commerce. It's traditionally been a lot of text-based search. But the images themselves, and the one thing these verticals benefit from is they have a lot of rich image data. For example, I might have 10 images for an item, maybe a video, plus the structured data. And so the multimodal search is really being able to use all of that data together. And so it's not just the text, not just the title, it's not just one image, it's like actually this kind of combination. And so the search experience becomes much better, because now you can do your vague queries, shirt or something like that. And you can populate it with shirts, but then you can do long sleeve shirt with stripes in a pocket. And then you also get that.
And so in terms of that kind of application, I think it's very well suited. We see a lot of other sort of adjacent ones, things like compliance, people looking for particular concepts and images. Image search itself is just a huge, huge area. A lot of people have lots of images. Think about real estate, for example. There's huge volumes of data that gets collected, and often without metadata. And so these methods, the multimodal search really allows a lot of that kind of discovery aspect as well. You've got a hundred moving images. What's in there? And we also see a lot of... Yeah, actually able to find content that people may not want as well, it's very good at finding that. Even if you've got metadata that people are trying to do, if someone's trying to circumvent systems and so they're trying to, maybe some content's not allowed, the images are usually pretty good at actually picking that up as well, even if it's not in the metadata. So yeah, it's been a broad range of applications. But certainly, large scale sort of product catalog, e-commerce has been a big area.

Eric Anderson:
So we have a real estate insurance company in the portfolio, and your example there kind of spawned an idea. Presumably today, people gather information about property, usually just verbally. Tell us, what kind of roof do you have? Is it pitched? Is it tin? Is it... And you could gather that metadata through images and map it to words and kind of have a synthesis of auto-generated risk information.

Jesse Clark:
That's exactly right. And so the vector search paradigm actually is very flexible. And so I think people have this notion of what the search should be. You sort of have a query comes in, it goes against some documents, and something's returned, but there's lots of ways to kind of frame it. And like you said, your query here can be the image, and then what you are actually searching over can be all sorts of attribute labels. And so now, you're finding what are the best matches in terms of attribute labels. And so you can do this kind of zero shot classification at scale, which is exactly like you said. So you can enrich the data through this process, or you can just simply search across the images themselves. So yeah, it's very flexible in that sense as well.

Eric Anderson:
What's the state of Marqo today? And how can folks interested in this industry level up through Marqo? What's the next step to get involved, that sort of thing?

Jesse Clark:
Yeah. If they want to level up with Marqo, I think, yeah, no, head over to the GitHub. We have a really active development, and really just start getting, we've got a great getting started guide. We sort of pride ourselves on the developer experience, getting started in three lines of code. We take that very seriously and we really want people to be able to get up and going with Marqo. So yeah, head over... GitHub has a plethora of kind of examples to get you started. And then yeah, we certainly have a lot of examples and the blog as well to go into more depth as well. And we kind of continue to add content there, and we'll keep doing that. I think that's again from the sort of Stitch Fix days as sort of saw how valuable that kind of really high quality content was on a blog, and I think we want to keep that going.
So yeah, that's how you get started. Where is Marqo today? Yeah, we're sort of developing, trying to really develop as fast as we can and get a lot of features in there. I think we have huge amount of ideas. And really, it's just about getting them now into the product and sort of keeping it the high quality. And then we also have... For people we've seen as well, building applications is managed services. People just want to be able to get started and going. And so we've got a pre-release of our managed service. And so people can also get to that from our website. There's a link and you can sign up. And there's also links from the read me on our GitHub. So if you're just looking for managed cloud, managed service of Marqo, you can also get in contact and do that. So open source is a great way to get started. And then we've seen a lot of users start with open source, and then be like, "Yep, this is what I want. Now, let's get the managed service and let's just keep building."

Eric Anderson:
Anything we didn't cover, Jesse, that you wanted to cover? We've kind of bounced around quite a bit.

Jesse Clark:
The only thing was potentially where everything's going. I think where we started... There's a bunch of use cases which have emerged, which I probably didn't even think about particularly closely a year ago. Certainly, the impact with generative AI, I think is immense. We've seen Marqo is absolutely perfect to use as kind of this memory or repository of knowledge for large language models. Even before ChatGPT came out, we were sort of experimenting with the retrieval augmented generation and things like that. And so being able to connect these generative models to Marqo, so for example, connect ChatGPT, now you can store information when you do a query, or ask it a question. You can search and you can provide context to the large language model, and you can get this incredibly rich experience from these processes. And you can do it with images as well.
You can have this kind of conditional image generation. So if someone's been using Stable Diffusion or Midjourney and you sort of ask for an image, and it sort of generates it and it can be all over the place. And so there's been other, the model developments which provide a lot of context, like control net. And so now you can use Marqo like a memory as well. So it's like, okay, I want to generate an image of someone walking on the beach. And so you can actually use Marqo to search for prototypes, for example, of people walking on the beach so that a lot of the actual scene construction is already kind of there, and then the model can fill in the rest. And so providing context for these generative models, I think is a huge opportunity for products and applications like Marqo.
We're going to see a huge evolution of how this actually plays out, particularly around how do we actually... Yeah, again, it'll be the same kind of problems that we see in search and information retrievals. If you start having language models, how do you find the best information to propagate for context? It'll be a lot of relevance and search ranking, I think coming into this as well.

Eric Anderson:
What I find fascinating with all the change happening is not only are people trying to figure out what the future looks like, but there's that adage like skate to where the puck is going. Where does the value accrue in the future? And every day, it's a new theory. There was a time when we thought OpenAI was going to rule the world. And now, the Google memo is saying that actually, foundation models are basically commodities now. And the thinking was you couldn't really build an app on top of OpenAI, and now all the value goes to the apps built on... How do you, Jesse, sleep at night knowing that tomorrow, the world's going to be different again?

Jesse Clark:
Yeah. It's definitely been a rollercoaster, and certainly some restless nights because you've seen the innovations come out, and almost entire verticals are completely changed overnight. In terms of the impact of AI, I think the biggest impact that's happening at the moment is on the people working in AI, developing... Especially someone has been spending years developing particular models. And then all of a sudden, a free open source or other kind of API comes out and it makes, it just lowers the barrier to entry again. I don't have answers about the future. It's all very exciting and ever changing. But I think the way we think about it is just focusing on in variance, right? What's not changing? And so we know, for example, exponential growth of information still... Majority of that's unstructured data. People are still going to want to search that data. Machines are going to want to search that data.
There's a huge need for real time search. That's not going to go away. And people still want relevant results. And so I think focusing on those kind of in variants is a way we're thinking about it, and sort of that's how we're navigating, I think a lot of the change at the moment, and I still feel like that's really the right approach.

Eric Anderson:
Well, your gut served you well when you got into this endeavor. You were kind of spot on when you launched Marqo, and so I expect you'll continue to see around the corners and know where to go next.

Jesse Clark:
Thank you. Yeah, let's hope so. Yeah, you got to be nimble though as well. You got to kind of see these opportunities, and you got to kind of be aware of what's happening. It's very dynamic, but yeah, exciting at the same time.

Eric Anderson:
Jesse, thank you so much for joining us today. Given the pace of innovation, maybe we'll have to circle back with you in a few days even and figure out where we are next.

Jesse Clark:
Yeah, absolutely. It'd be my pleasure. We'll see what happens.

Eric Anderson:
You can subscribe to the podcast and check out our community Slack and newsletter at contributor.fyi. If you like the show, please leave a rating and review on Apple Podcasts, Spotify, or wherever you get your podcasts. Until next time, I'm Eric Anderson, and this has been Contributor.