AI as an interface to the world

Tristan Zajonc is the cofounder and CEO of Continual, a developer platform for generative AI applications. He previously cofounded Sense, which was a platform for data science and machine learning. It got acquired by Cloudera in 2016. He spent 3 years at Cloudera building ML software. He has a PhD from Harvard.

In this episode, we cover a range of topics including:
- AI as an interface to the world
- Tech stack of the future
- The founding of Continual
- AI product delivery avenues (cloud, on prem, model hubs, packaged solutions)
- Traditional MLOps vs LLMOps
- Open source AI
- AI compute market

Tristan's favorite book: Suburban Nation (Authors: Andres Duany, Elizabeth Plater-Zyberk, Jeff Speck)

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.255)
Tristan, thank you so much for joining me today.

Tristan Zajonc (00:04.398)
My pleasure. Thanks for having me.

Prateek Joshi (00:07.043)
Let's start with the premise that AI will be an interface to the world. And we're at a point where it's being used everywhere. It's being used as a way to interact with software, with many systems, and it's making it faster, better, and cheaper. So what's your view on how the interface positioning will pan out?

Tristan Zajonc (00:34.958)
Well, I think one thing is that it's going to be everywhere. So there isn't one interface, right? You know, one AI system will be your car. Another AI assistant will be whatever the next Siri is on your phone, or if you even have a phone, whatever the device is that you kind of talk to, that's that sort of super intelligent personal assistant. And we're going to see that sort of everywhere. Our kids are going to be like, oh,

have devices that help them learn and maybe follow them throughout their entire life and deeply understand the development of the child over their life. So I think there isn't necessarily one interface. I do think some of the existing interfaces that we have, particularly around the way we consume and find information, are going to fundamentally change. So for example, one of the things I'm very intrigued on is sort of what is the future of the internet.

The internet today is sort of like this two-sided marketplace for information. You know, on one hand you have website creators who sort of put information out there and, you know, obviously there's different ways to put it out. You can put it out on social media, the channel. And then there's consumers where you go around and you try to find information, things that you're interested in or answer questions that you have. And in this two-sided market bull, the internet, the invention of the internet really created that sort of that interface, right? That interface, which was, which was honestly incredibly transformative because it unlocked the shared.

sort of knowledge of humanity to sort of come together and be shared and collected and curated. And Google in that world, I mean, if we're honest, is sort of the middleman or the market maker here. We don't really type in addresses anymore, we go to Google. And I think one thing that, for instance, AI excels at today is really this information synthesis, finding digestion compression. And I could see a very profound change to

sort of the two-sided marketplace for information, sort of how information is entered into sort of the shared brain of the world, which is now the internet, but is that what exists in the future? Very much not sure. Every time I go on a recipe website and I click through 23 banners and whatnot to find a recipe, I can see an alternative, right? And so I don't think that's happened, but that's one that I've been mulling on and I could imagine just waking up a year from now and-

Prateek Joshi (02:44.426)
Hahaha.

Tristan Zajonc (02:54.798)
You go to something dot stand, the equivalent of Google dot Stanford dot edu and you just realize wow, or the equivalent of something like chat GPT, although I think it's not, you know, chat GPT I don't think is quite the version of the next internet, or at least the current incarnation of it's not. But I think yeah, these interfaces are going to fundamentally change not only the physical interfaces but the way we can consume digital information for sure.

Prateek Joshi (03:20.659)
got it. So the key takeaway here is that you love banner ads. So just love. No, no, it's the...

Tristan Zajonc (03:26.798)
Well, I think, yeah, no, it's, yeah, it's true. I mean, I think I remind myself back to Google. I mean, I was, I'm old enough that I remember this, the search engine wars. And, you know, one of the amazing things that Google did was it just had this, you know, single search box, right. And had, you know, these 10 links and it was amazingly simple. And they preserved a lot of that, not all of it, but they preserved a lot of that. And I think, I think, I think honestly, I mean, one of the successes of chat GPT or one of the things that makes chat GPT.

sort of delightful is honestly that it just gives you what you want without any garbage. And that's a wide open opportunity, I think, as the internet has sort of decayed a little bit from a user experience perspective to many people. I think that opens, not only does, I mean, obviously, intelligence opens up a tremendous possibility, but also sort of the frictions that currently exist in our existing systems. Leave open some doors to explore.

Prateek Joshi (04:26.311)
When it comes to all the applications that are getting built, there are so many different use cases and the infrastructure layers are also just shaping up. It hasn't been super standardized just yet. So when you look at, maybe, I don't know, two, three years from now, where things are a little more standardized and AI application, it works well, it's nice, what does the stack...

look like. So if you're talking to a developer in the future, what stack are they using?

Tristan Zajonc (05:00.334)
Well, this is something that I've personally been thinking a lot about recently. I think there's really going to be sort of two separate stacks. Um, and obviously they'll talk to one another, but one I think you could think of as, as the engineering stack, you know, whether that's software engineering or mechanical engineering and you know, it's really a stack that is grounded in sort of a more mechanic, you know, mechanical or logical, a physical, um, sort of set of tools, right? You're trying to build tools, a car, a hammer.

or a software application, right? Which is often a system of record for some business process or codifies in some logical way, a business process. And that's not gonna go away, right? If we look around the world around us, we're these intelligent beings in this world and we use these tools, they're tremendously helpful for us. In fact, a lot of our civilization comes from the fact that we've used our intelligence to build mechanical tools or physical or logical.

tools that augment us in various ways, calculators, software applications, cars, machines, everything. And so that I don't think, the rise of sort of this intelligence era, this AI era, doesn't replace the need for tools. Intelligence is augmented and enriched and enhanced by tools. They serve a different function. So we have a rich obviously set of tools for engineering, and I think that is going to stay the same.

What we have that's new, I think, and is going to emerge is an intelligence stack. What is the stack on which we build, you know, what historically has been us, right, the human, and the things that we excel at? And I think that stack honestly looks much more like biology, you know, it's something where you have a few principles. I mean, so there's the, you know, there's obviously the LLM ops, you know, we can talk about that maybe later in terms of the details of it. I don't, I just mean much more fundamentally.

If you think about how AI systems or how intelligence is built and how it operates, what's, what are the principles of it? It feels more like biology in the sense that there's a few very simple principles. Uh, like, you know, evolution in the case of biology, but, um, in the case of, you know, ML it's basically how do you apply data and compute, uh, to achieve sort of some objective or train some system, the model itself can be very simple. Right. And then this tremendous complexity emerges.

Tristan Zajonc (07:17.902)
So rather than in the engineering stack, you kind of build up these building blocks. Each thing is very deliberate. And it's an engineered system in that sense. In the world of intelligence, it's much more of an evolved or a learned system, where you actually have a very simple set of primitives. I think we're still working on what those primitives are. And then you evolve or retrain the system to behave in the way that you want.

And of course, you're going to build tools to help you do that. There's going to be mechanical tools. The LLM ops tools are examples of tools to help you do that. But really, the core is that learning process. And I think that does have a different nature than the traditional engineering sector. So what's the next result? Where does this end up in the future? I think you basically have these two sort of systems. One is your software applications. Obviously, AI is going to help you build those things.

You're going to have this intelligence layer, which is much more about evolved complexity and how do you evolve the system or train the system, have the system learn what you want, behave in the way you want. And then there's going to be this incredible dynamic interplay between the two, right, where the intelligence system is calling the tools and leveraging the tools to accomplish tasks. And of course, the tools themselves are going to be infused with intelligence. They're going to be calling into that intelligent artifact that comes out the other end of your intelligence stack and doing things. But I don't think they, I think they are sort of

fundamentally different just like the way biology and physics are different, even if they're kind of built on each other.

Prateek Joshi (08:48.507)
Right, that's actually, I love that, that way of thinking, how it's kind of functions the way biology does. And maybe it's a good stopping point for listeners who don't know. Can you quickly explain what continual does?

Tristan Zajonc (09:07.086)
Sure. So at Continual, we are huge believers that AI can augment humans and empower humans to achieve more in sort of everything that they do. And one of the things we thought very hard about is how is that going to happen? And what is that going to look like to your kind of introduction question? What is that interface? And while there are going to be many interfaces,

I think one of the dominant interfaces, particularly in software interfaces, is this idea of an AI copilot, an AI assistant that is embedded inside of an application powered by large language models and these generative AI that deeply understands an application domain, the application data, the application APIs, and is there as an expert to assist the user to automate and accelerate workflows to enrich experiences.

And so what Continual does is we're building really the fastest way to build an AI copilot into an application that empowers your users to achieve more and kind of do things that were previously impossible to do. And how we do that, we have what we call an AI copilot platform for applications. So we're laser focused on not this general idea of an AI assistant, but this idea of a domain AI copilot, an application-specific AI copilot embedded into the application, deeply understands that application.

and empowers those application users to sort of do what was previously impossible for them to do.

Prateek Joshi (10:38.559)
amazing. And when you look at all of the applications, is that a specific sector or cluster forming where the customers tend to use it a little bit more than say others?

Tristan Zajonc (10:55.246)
I think we're still in the very early innings of what is possible with AI copilots and assistants. And so, you know, right now, you know, if you just look at what these large language models are incredibly good at, they're incredibly good at digesting information and understanding information and then synthesizing it and kind of question answering type use cases. And so, you know, today we see

Maybe they're also good at sort of like, you know, drafting first content, right? So you need to write a job posting. Okay, it can certainly draft that, you know, just instantaneously give you a first pass at a job posting. But the other types of questions that it's really good at are these, these, these just how do I do X, right? Inside if the large language model, you know, is exposed to the necessary information, you know, has been sort of knowledge base or knowledge index with respect to what the application can do. It can obviously ask, you know, answer questions.

about the application. I think, honestly, that provides a lot of value to every application, every business to business applications in particular, whether that's customer facing business to business applications or more internal business applications within larger enterprises. And so there's this sort of base level. So first level, look, these things are gonna be incredibly useful just as kind of a base, hey, there's somebody who's sitting there who understands this application, who's kind of an application expert who can help you.

And then it starts to, you know, that's just the question answering. I think then it starts to get into, well, how can it really assist you doing your work? And today we're seeing that within certain, I would say creative tasks or knowledge generation tasks, where you're doing, for instance, writing heavy tasks. And that could be very broad. It could be writing emails. It could be writing job posts. It could actually be writing, you know, content, marketing, et cetera. There's, you know, it really, really radically accelerates the workflow of those types of people. If you're a lawyer, right? All you're doing is writing.

These LLMs are amazing at that. At creative tasks like generating images, with these multimodal applications or generative things like Dali, Mid Journey, Firefly, Do We Firefly, what's obviously coming to Figma and tools like that, these creative tools. I would say that's just starting in terms of being fully embedded into all the creative tools that are out there, but that's going to just completely revolutionize the workflow, 3D modeling tools. And then it's going to be engineering tools, CAD software. All this stuff is going to be just

Tristan Zajonc (13:10.99)
radically expanded as the large language models become capable of actually doing that. And then over time, I think we're at the beginnings maybe of the rise of more agent-based AI systems. And that's going to really start to make it possible for these AI systems to actuate applications in ways that are tremendously useful. So today, they can take actions on behalf of users and applications. You can do that on continual. You can connect it to your APIs. It can help users.

But when these models become more powerful, I think the user value is going to very significantly increase. So I really think we're in the first chapters of it. Sometimes I hear people saying, oh, is this like a chatbot? And occasionally, maybe, maybe there's little elements of that. But if you kind of squint, I really think there's no limit in terms of what is potentially possible. And maybe there will be a dominant conversational interface, but there will be many other interfaces as well. So.

I think essentially all applications are going to be transformed by this idea about co-pilot.

Prateek Joshi (14:14.771)
There are many product delivery models when it comes to getting AI delivered to your customers. Sometimes it's a cloud hosted model. Sometimes it's on-prem. Sometimes it's a model hub where the platform takes the guarantee that, hey, everybody on the supplier side will vet those models. And on the demand side, you can just use them because we underwrite it. Or maybe there are other companies that provide fully packaged enterprise solutions. So when it comes to

all these different delivery models. How will this market shape up for AI vendors? And more importantly, there's room for all of them. Who's gonna need what in terms of the customers? Who'll end up going with what?

Tristan Zajonc (15:00.878)
So I'm bullish on sort of two approaches, and maybe not bullish on some of the other ones. So the first one is cloud. I mean, this is, I think, it's obvious. Cloud-based APIs for these intelligent services, the ChatGPT OpenAI APIs is obviously the dominant one today, but there's going to be more that are extremely compelling. And I mean, it is just so the fact that it is so simple for any developer.

with no AI background to build remarkable software on the basis of these APIs with no infrastructure overhead, no engineering complexity. And honestly, at quite low cost, and at cost that is going to drop tremendously due to the immense ability of these large cloud providers to optimize the inference stack all the way down to the silicon. I think the cloud delivery model.

is just going to be an incredibly compelling delivery model for customers. And so I think that is going to be a dominant one. I think there's another one, which is there are customers who need full control of the model. They're really trying to push the envelope in terms of what is capable with the model, and they want to do something where the cloud, foundation model from the cloud provider doesn't have the flexibility that they need. So mid-journey needs to have its own models. You can't build mid-journey without that runway ML, which is a c**t.

a video, text to video type application, you really need your own models, you really want your own control. And so I think there's going to be the need to have these open source models that you control, and then there's gonna be a need to be able to just serve those, your models that you control, you're gonna wanna serve them. And so there could need to be, and I don't think necessarily everybody wants to serve them themselves, and so there's gonna be inference providers that basically allow you to take your model, any scale endpoints, together.ai,

Obviously, all the major cloud platforms are going to have inference platforms that allow you to take your own models. Just finally here, I mean, to say what I'm not bullish on, I'm not a big fan of model hubs. So, I just feel like, I mean, I'm a big fan of them for like, hey, let's put everything out there and we can see it and we can collaborate. It's great. It's like GitHub. But in terms of value that it delivers to end customers in terms of how useful are the models, not super useful, right? I mean, these models just go there and there's only, you only really care about the few...

Tristan Zajonc (17:20.27)
best ones for each area. So text generation, what's the best model today? Maybe at what latency level? So what's the size of the model? And a few different domains. So these model marketplace, model gardens, I just sort of think they don't do too much. And sort of the on-prem packaging story, I mean, I was at Cloudera, which was an on-prem. My previous startup was bought by Cloudera, which was primarily an on-prem company at the time. And there's a market there, but I

It's the scalable one. I wouldn't focus on it too much, although it will certainly exist in the future.

Prateek Joshi (17:59.099)
Right. Actually, it's a good segue into my next question about model routing. So in the sense that let's say, okay, cloud is the dominant business model and it makes sense. So as a customer, I have access to a few different models because I just have to, sometimes I have to generate an image, sometimes it's text, sometimes it's video. So it makes sense to have access to a few models. And also, let's say for the sake of this question, I'm only dealing with text.

Tristan Zajonc (18:08.462)
Mmm.

Prateek Joshi (18:28.975)
And I have certain constraints like accuracy, speed, and I just have some cost constraints. So based on that, the input query should be routed to the right model for me. Sometimes it's Lama 7b, sometimes it's 65b. So when you think of a model routing product, how do you think that product

might look like. And more importantly, let's say if I'm a big company, if I'm open AI, if I'm some other company that provides a bunch of text models, should they be allowed to select the models for me? Or should that be independent because there's a conflict of interest? Because if I'm asked to select a model, I'll choose the highest margin one for you in the sense that you give me a query, I'll pick a model that makes the highest amount of money for me. So, how do you think that product or that the work of model selection

will look like.

Tristan Zajonc (19:26.126)
I'm a big believer in the idea that there should be a model router in your stack. So at the very core, actually, of Continual, which is kind of focused on this copilot use case, there is a model router. We basically have a dynamic model router. We select between different models. For instance, today, the most common models we select between are GPT 3.5, which is cheaper and faster, lower latency. So lower latency is also a better user experience. So actually, the quality from a user perspective goes up and the cost goes down. So it's a double win-win.

And then for many of our types of questions, and then GPT-4, for instance, is better at certain types of questions, particularly these technical questions that some of these copilots need if they're in more technical products. And then I think as, well, two things. One, as other large language models come out that are exciting, I think we'll see what Google does with Gemini, for instance, or where Anthropic goes, or what AWS

leases at AWS Invent, they're saying they're going to release something, two trillion parameters or something, it's going to be very compelling to potentially switch between them. And honestly, just to have that optionality to switch between them. And so I think a great architecture for your application is to have a model router to try to abstract the degree that you can, especially these common interface things like text, chat interfaces, to try to abstract that and dynamically route between the models. And then I think...

You know, the final one is that you also want to route between models because often in an application, particularly as you become more advanced, there's a particular area, like you have an asset format that's unique to your application. For instance, in Figma, that is the design file of Figma, right? In a CAD software, that's the CAD design software. In, you know, if you're building a Verilog program or something, you know, it's a particular domain specific language that's important for your users, right? Or it's going to be a legal document.

It's something that's very specific with the way legal documents are done. And you might have, in that case, you might want to really take control over that model for that specific use case and route to that model when appropriate. And I think there's ways to actually build very intelligent routers that they can do a lot of this very automatically. To optimize this set of constraints, which in my mind are really cost, which is one, cost management, which is important.

Tristan Zajonc (21:47.662)
Latency, which is tremendously important from a user experience perspective, and then quality or performance of the outcome. And so, yeah, I think model routing is great. I think it's also great for the ecosystem. I think if we can build a world where applications can use potentially multiple models, that's going to lead to a richer ecosystem. I am a little bit concerned that we may end up in a monopolistic situation if we don't have something like this, which I don't think is going to be great for innovation.

It's not going to be great for startups and it's not going to be great for innovation. So, you know, we'll see where it plays out, but I think it's a, it's a great way to hedge, you know, hedge against different things and to get the best of kind of all possible worlds as in an incredibly dynamic environment.

Prateek Joshi (22:31.315)
Right, that's amazing. I love what you said about the need for model routing and also what it means for the ecosystem. And also when you think about traditional ML ops versus LLM ops, there are some tasks that didn't really exist in the previous ML ops world that now you have to do because it's just a new construct and you have to deal with new operational realities of LLMs.

So maybe can you just compare and contrast just a couple of things that you now have to do for LLMs that you didn't really have to do before for like traditional ML ops.

Tristan Zajonc (23:09.486)
Yeah, so this is close to my heart because I, you know, prior to Continual, I basically spent 10 years of my life building MLOps tools for enterprises with one startup and then at Cloudera. And so I've thought a lot about this. And I think of, you know, MLOps is very close to data engineering. Like I think, you know, if I kind of take away the branding and I think of what at the core of MLOps, I think honestly data pipe.

Prateek Joshi (23:17.718)
Right, right.

Tristan Zajonc (23:35.822)
I think what, you know, my recommendation was, you know, often we'll start with, you know, choose a data orchestrator, you know, an airflow, a DAGST or a prefect. I mean, there's a bunch of these things, but at the core, you kind of just need to write code, do things that is essentially data engineering, but fine. You're calling some model training, model prediction calls, you know, explainability calls in these pipelines. And then sure, you know, log them to weights and biases or some experiment tracker or something else, which is, you know, you know, and then, okay, fine. Once you're in production, you know,

do some monitoring, but it's honestly, that monitoring is often a data engineering job as well, unless if you don't care about real time. So I think of MLOps really like data engineering. And I do think data engineering is kind of a different discipline than software. I mean, it's software engineering, of course, at the core, but it is sort of a different set of tools than software engineers, like application software engineers, like the application development stack, right? You're using tools like Spark and the big data stack and Hadoop and all that world.

So in data, and that's really where ML apps comes from. I think then, LLM Ops, it feels to me much more like application development. And in parts, that's because of the fact that so many of these frontier models are offered via APIs. And so for many people who are building applications, are essentially just building applications and they're leveraging these APIs, which do new, fundamentally new things.

Um, and so you, you know, the persona who's kind of using the, these large language models is more of an application developer than an ML engineer or a, you know, a data engineer. And, um, so that changes, you know, everything. I mean, all of a sudden typescript or something becomes like, you know, actually, you know, Python is still there for sure. But, you know, in many cases, actually these, all of these applications are building and built, built, you know, built more with a kind of a JavaScript based stack. Um, now there is then.

the tooling on top, the kind of this LLM ops category of products that is emerging. And I would say those tools, if you look at, especially the commercial, not just the libraries, but the kind of the platforms that are evolving, they're really observability tools, right? In some ways. And so that's actually very similar to application observability, right? They're building tracing solutions. Well, application developers have these amazing tracing systems, Honey, Sunny, I mean, a whole bunch of them, Honey Ive, and there's Honey Comb now, which is doing the equivalent for LLMs and...

Tristan Zajonc (25:57.294)
Langsmith from the Lang chain folks. And so I actually think these LLM ops tools are looking, look much more like the observability stack, the performance monitoring stack, right? You have the new relics of the world, you have application performance monitoring. So that's kind of also very similar to, you know, debug ability. I mean, there's a very rich sentry air tracking, right? And so just the, you know, yes, these are diff, they're not application.

Um, monitoring tools, like there is a unique role for LLM ops tools, but I would say it feels much more to me, analogous to application performance, like production application platforms. Then it does to the data engineering stack, which is really the backbone of MLOps.

Prateek Joshi (26:39.431)
That's amazing. That's a fantastically, it's such a nuanced point that you're making here. And I want to take a second to just double click on that. And traditionally like monitoring data pipeline, it's, I never would say it's kind of hardcore engineering. You got to know so much and so much data is coming in. You've got to keep an eye on it. But when it comes to monitoring applications, do you think that big companies like Datadog will just say, you know what, we can do that because we've been doing this and it's so close to.

like monitoring applications that they're doing so well. So is it so that the previous like APM performance monitoring companies, will they be able to take this and run with it or is there room for startups to make them work and build a big LLM ops?

Tristan Zajonc (27:30.862)
I think there's room for startups. And so I mean, you certainly can monitor LLM-based applications to a certain degree with tools like Datadog. And Datadog will obviously try to do something here. But currently, if I look at even some of the early LLM ops observability tools, LangSmith, HoneyHive,

HelioKone, HumanLoop, I'm probably missing a bunch, but they seem to offer, to me they offer, they're very tailored towards the types of problems you have when you want to monitor LLMs in production. I do think they're missing some things, but which I can talk about, but I think there's plenty of room to do things in the startup space in this area.

Prateek Joshi (28:15.643)
Right, amazing. All right, let's talk about open source. There's been so much debate, more recently about what's happening in open source and also the effort around regulating AI. So one, a part two part question, part A, what's your view on open source AI today? And also what's your view on the effort to regulate AI? That's been happening.

Tristan Zajonc (28:44.11)
Um, yeah, so two, two parts there are somewhat related, but two parts. So, so the first one, um, I'm a big believer, you know, I think it's going to be a yin and the yang. So, you know, obviously the proprietary models behind, behind cloud APIs, you know, delivered as a service is going to be an enormous business. It just is great for customers. It's easy for companies to deliver and, you know, support and maintain. Uh, and we've seen that with SaaS applications and I, and we're seeing that with open AI with these API tools.

I see no reason why that's going to change. And I see no fundamental reason why open source is going to disrupt that, except of course, put a typical market pressure to continue to innovate and to not charge crazy prices, which is fantastic. It protects us against monopolistic behavior. It's one, also just competition between companies will do that. So I think these cloud non-open source models will be a big thing.

I think open source, I think this intelligence wave is so important that we need open source models just like we need Linux. We need models that we can fully control to unlock all the innovation that is out there potentially in the system. If we're in a world where innovation just comes from one or two companies, right, which is we end up in this sort of oligopolist situation where because of some sort of the market dynamics of the size of the model, the scale of the data.

that you really only can deliver value if you're one of these large companies. Innovation will suffer. Competitive prices won't probably drop as fast. And it won't be a good thing. The good news is I think we're in a world where there's lots of reasons why I think great models will be accessible in open source. There's nothing fundamental that doesn't. The frontier models may always be

done by in a proprietary way, just given some of the costs involved with data and compute. But amazing capabilities are going to become incredibly cheap, both incredibly cheap to train, and then incredibly cheap to reuse. So I'm a big fan of open source from that perspective. One thing that, from my perspective, I would add here is I actually don't think open source from a cost perspective is that compelling. So I mean, it can be. If you have a very specific application, you go, and you definitely is compelling if that application is narrow. But if you're a general LLM,

Tristan Zajonc (31:04.014)
there is going to be tremendous cost pressure on the, you know, that you're going to have due to normal market conditions, it's going to drive down cost. So actually, I'm much more, I mean, cost is important, but I actually am much more excited about open source from the control perspective, what it unlocks from an ability to innovate. And, you know, you see that, I mean, mid journey is one that being the runway, as I mentioned, is another, but you just see the, and I think we're just going to see more and more sort of like vertically integrated companies that do their own models and build different products, right?

It doesn't replace all the other companies that are going to build on more standard models, but it's going to unlock a whole different avenue of innovation. And then it's going to force the big folks to react and to up their game in various ways. And I think that's just incredibly healthy. So I mean, yeah, I guess we can get into the second question, which is regulation. That's a longer one.

You know, I am a, I'm an incredible optimist in terms of what AI can do. I think there's a lot of FUD and, and, uh, sort of misunderstanding about AI risk. Um, uh, and, you know, I think I went much more believer in, you know, there's potential room for regulation in, you know, applications. Uh, so once these models are deployed, if they're doing bad things, um, uh, where bad things are really creating harms to people, not just like things you maybe don't personally like.

There might be room for some regulation, but I don't really see that happening, honestly. And in the areas where it might be happening today, maybe disinformation or worries about that, that's a very fraught area to regulate. And so I'd say we've got to be very careful. There's a very fine line there with respect to our rights as individuals to speak and things like that. So I think I do actually, if I just look technically at these models, I'm more in the camp of Jan LeCun.

there's ways for us to control these models. It's actually, we have almost dictatorial control over these models actually, if you look at how they behave. The alignment problem I think is going to be solved technically. And I think we're gonna have to be a little bit, I mean, I think it's just gonna be solved. I think it's gonna be solved. I think there's gonna be some profound society shifts that this causes and things that we need to think about. But.

Tristan Zajonc (33:26.958)
I think this, you know, the fear that's going on right now for in some camps, I think is unwarranted. I think, you know, it'd be a tremendous loss if we lost out on the innovation, the potential of AI, the amazing ability of AI to solve so many problems, help us as individuals, you know, if we sort of, you know, get too involved too early. And there's gonna be time for us to sort of see real world harms and then react appropriately, which, you know, we absolutely should do if we see those harms.

Prateek Joshi (33:54.595)
Right. Amazing. I love the point you made about Linux and the funny thing, because we have Linux and the entire world right now is built upon cloud and a vast majority of the cloud infrastructure, even now, like it's, it started off using Linux and people like many, many servers still are built using Linux. Obviously there are flavors of that. So I think even though at the time, like 20, 25 years ago, Linux was

because hey, open source needs to exist, but it can lead to huge, huge economic outcomes too. It can give rise to like a whole bunch of ecosystems that you know, it's hard to fathom in the moment. So, all right, that's fantastic. With that, we're at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready?

Tristan Zajonc (34:42.382)
Sure, let's do it.

Prateek Joshi (34:43.747)
Alright, question number one. What's your favorite book?

Tristan Zajonc (34:47.886)
I mean, this one is such a hard one. I mean, I'll just choose something totally off the wall, which is I actually read a book called Suburban Nation many years ago, and it still sticks with me. I drive around and I love books that are in areas that are actually not areas that I think about day to day. I mean, I nonstop, I'm reading ML stuff and statistics stuff, but I love areas that just sort of expose me to another world. This is a book about sort of the urban design, the way we live in the world.

And you read it and you're just like, wow. And you go onto a street and you just look at the way the trees are laid out, the way the cars are parked, and it totally changes the way you think. And that was a book that even 20 years later, basically, it sort of stuck with me. It's just a total, I don't even know if it's a good book. It's one that just has influenced me.

Prateek Joshi (35:36.547)
I really enjoy reading those books that take you completely outside of the thing you do on a day-to-day basis and it makes you think and maybe apply some of those concepts to your work. So I really enjoy those books. All right, next question. What has been an important but overlooked AI trend in the last 12 months?

Tristan Zajonc (35:56.558)
Uh, I mean, I honestly think everything that's not autoregressive transformers. So, I mean, these GPT models just work so well that everybody is focused on them. I mean, you see every single open source model that comes out is essentially a kind of a clone of these models and it's, it's rational. I mean, they work, they work immensely well. They're kind of proven. Um, uh, but I mean, I think it's sort of, you know, it sucks, sucks the, some of the oxygen out of the room. Maybe that's a little bit because also maybe some of the innovation that's going on is not being published, you know, so it could be that there actually is more.

It's not totally overlooked, all the other things, but unfortunately a lot of the large research labs are not publishing some of this stuff.

But I get really excited by things like adaptive computation, like Google has this adipate paper, memory and continual learning. So how in general are we going to build systems that have more continual learning ability that maybe has memory embedded into it, so it's not just a training regime that causes these systems to learn, but there's really a continual learning dynamic. Maybe some sort of search and search. You know.

sort of the, how do you get something that's more like self-play and, you know, sort of like the alpha zero? You know, I think people are definitely working on that, but I wish more people were working on it. I think just a lot of researchers out there that are not stuck in these large labs are just being like, how do we catch up to what the large labs are doing and they kind of look for proven ideas and then just implement them? But I would, you know, I want some exciting, I think some exciting breakthroughs are gonna come that are not just these autoregressive, you know, transformers, which are gonna be, you know, huge part of

I think AI architectures, but, and can take us a lot farther. But I think just it's unfortunately sucking a lot of the air out of the. The public room and hopefully some people in the open source world that are publishing still will, will pursue these other ideas.

Prateek Joshi (37:44.807)
What's the one thing about AI co-pilots that most people don't get?

Tristan Zajonc (37:51.502)
that they're just going to get way better. And I think you can be a little bit dismissive sometimes of them when you think, how am I going to use it? But if you just think about, if you have any domain, if you're a manager and you have an amazing software engineer or an amazing designer or a amazing marketer or an amazing lawyer, I mean, what do you do? You basically tell them your problem and then they go fix it right for you. And that's how it works.

in the real world, then I don't see any reason why ultimately that's not how it works with your co-pilot, even if that's not possible today.

Prateek Joshi (38:27.899)
What separates great AI products from the good ones?

Tristan Zajonc (38:33.038)
Um, well, this is similar to my last answer, which is I think quality, uh, I think AI is a tipping point product. Um, there's a quality threshold and when it tips past that it's amazing. And if it doesn't tip past that it's horrible and annoying. And, uh, I think we saw that with, you know, chat GPT a little bit. And I think actually there's a next, you know, some next version, which is. Whenever, you know, a chat GPT like system becomes, you know, better maybe than the, you know, one person. Like.

Prateek Joshi (38:48.756)
Ha ha

Tristan Zajonc (39:00.782)
all but 1% or all but maybe even 0.1%. We all in our world, I think we're, the people that we love working with in these narrow domains are world-class experts, right? It's that amazing writer or that amazing film director or music creator or engineer, right? And so we're used to actually interacting with things in our world that are actually created by remarkable people, remarkable exceptional people. And I think AI really hasn't gotten to that level. And once it does,

Oh, it's going to be, it's going to open, it's going to just tip, you know, and it could even be very similar experience. It could just be chat, GBT, but it, now it writes, you know, amazing, you know, insightful, you know, con articles rather than today. It kind of writes these fluid, but incredibly dull, you know, articles, if you're honest with yourself, right. Very kind of 50% sort of style articles, you know, when it, when, if you ask it about, you know, to write you an article, it's on, you know, the future of MLOps or something, right. It's going to kind of give you pretty vanilla stuff.

Prateek Joshi (39:59.615)
As a founder, what have you changed your mind on recently?

Tristan Zajonc (40:05.166)
Well, I mean, I do flip-flop a lot on this open source versus frontier model question and sort of how are the market dynamics going to entirely play out. And I think I have become more bullish on open source and vertical integration. I think it's going to become more accessible, like remarkable models are going to become more accessible and that, you know, full control does allow you to.

tip past these points, right? And that there's, I think a lesson that I've seen with a few startups is if you just relentlessly focus on quality, you kind of can actually a small, you know, a thousand paper cuts, if you solve them, can lead to something that's very different, sort of more as more as more as different and or better as different. And having full control, which open source allows, does allow that. It's not the right path for every team for sure.

But if you have a, you know, if you're capable and you have a remarkable team and the necessary funding, I think I have become more excited about that as a possibility.

Prateek Joshi (41:06.003)
What's your biggest AI prediction for the next 12 months?

Tristan Zajonc (41:11.982)
Um...

I don't know that we're going to see some wild breakthrough over the next 12 months. Um, I think what I'm watching for carefully as a startup is whether open AI loses its monopoly. Um, and by that, I mean, you know, open AI, if I think Ron is just does have the best models, you know, that are, that, you know, you might as well use for, for, if, if, if you're, you know, if your application fits that they're, they're paradigm, um, they're doing a, they're doing a fantastic job. Um, but you know, there's some, obviously some other players, you know, Google is one.

Um, where, you know, are they going to do something with the Gemini series? That's remarkable. And, you know, all of a sudden opens up, you know, a whole new sort of, uh, set of opportunities for, for startups to leverage and change the sort of the market dynamics, um, uh, or is that not going to happen? And so, um, you know, maybe my prediction might be a, you know, a opening, I will lose its monopoly over in the next 24 months. Like, well, there'll be, there'll be these models that are remarkable. Um,

from other vendors, whether that's open source, maybe it's llama three. I mean, that's the other one, you know, llama three or some or Mistral, you know, three or something like that. Uh, or maybe it's the Google Gemini series. Um, or maybe it's what AWS does, although I, you know, I doubt.

Prateek Joshi (42:25.319)
All right, final question. What's your number one advice to founders starting out today?

Tristan Zajonc (42:34.126)
I mean, there's so much good advice out there for founders and in terms of, Hey, it's going to be hard and follow your passion and be persistent and, uh, you know, don't follow the trend. Although that's hard now without the AI world, uh, you know, in the AI world where it's just so exciting to kind of be in this, in this, in this dynamic environment of what's going on in AI. I mean, one, one maybe that comes out of my own experience, um, you know, is like really try to do right by everybody.

your employees, your investors, your customers. It's just amazing the payoff that has and life is long. Startups is a relatively small ecosystem of people. You usually do a couple, unless you happen to, you usually do a couple, honestly, especially if it doesn't go well, right? So if it doesn't go well or if it doesn't go that well, usually you probably wanna do another one. And so how you treated all of those people

It will, they will just, you know, it's the right, it will make you feel good and make you feel, be able to sleep at night if things don't go well. And it will also, you know, if things go well, you know, it's, it will build, be this tremendous platform on which to grow. I mean, I work now with, you know, a lot of people, honestly, that I worked with in my previous startup, my co-founder, actually my old co-founder just rejoined us actually this week. And a lot of the investors that I had in my first startup has joined again. And, you know,

It's just been, I actually been shocking actually how much that mod matters and how, uh, uh, yeah. So I would just say, you know, really, really try to do that. Um, you know, even if it costs you a little bit, uh, along the way.

Prateek Joshi (44:14.451)
Fantastic. Tristan, this has been such a fantastic discussion about so many different topics. So thank you so much for coming on to the show and sharing your insights and knowledge.

Tristan Zajonc (44:26.382)
My pleasure. It was really fun. Thank you.

Prateek Joshi (44:30.504)
Perfect. Hold on.