Next in Tech | Ep. 257: CEO Series: Agents and context with Douwe Kiela

View Full Transcript

Eric Hanselman: Welcome to Next in Tech, an S&P Global Podcast where the world of emerging tech lives. I'm your host, Eric Hanselman, Chief Analyst for Industry Research at S&P Global. And today, we have the next in our CEO series, where we get some insights into how businesses are putting AI to work. With me is Douwe Kiela, CEO of Contextual AI. Welcome to the podcast. 
 

Douwe Kiela: Yes. Thanks for having me. 
 

Eric Hanselman: It's great to have you here. Normally, the podcast really digs into some of the mechanics of AI. We wanted to raise the discussion a little bit, and especially with where contextual is and a lot of the work that you've been doing around real implementations around agentic capabilities. It seemed a great opportunity to get your perspectives. Before we get too far into it, can you give us a little background into Contextual and really what you're doing? 
 

Douwe Kiela: Yes, for sure. So my background is in AI research originally. So we started this company almost 3 years ago now based on the research that we had done before. So I was Head of Research at a company called Hugging Face. And before that, spent 5 years at Facebook AI Research in the early days where one of the things I had been working on with my team was retrieval augmented generation. So that's RAG's what everybody uses these days to get AI to work on top of your data. So we recognized after ChatGPT happened that the world was never going to be the same again. And that after the intelligence problem, which is how do you get a language model to be really good and really smart, the next problem would be the context problem. How do you make sure you give that smart thing the right context on your business in order for it to be able to deliver value and solve real problems. So we started the company with that focus. That's why we're called contextual AI. And yes, it's been a lot of fun sort of seeing the world discover that AI is extremely powerful, but sometimes also extremely stupid in surprising ways. And that's why the context problem is really the big problem to solve these days. 
 

Eric Hanselman: Well, it seems so often we talk about quality of input versus quality of output. And it is much the same in AI that you can get an incredibly smart model to do some relatively dim things if it's not managed well. 
 

Douwe Kiela: Yes. It's the same with a human, too, right? -- context, right, then you're going to be yes, not great at your job maybe. 
 

Eric Hanselman: Well, so in the broader sense, what are the kind of problem areas that contextual is trying to address? 
 

Douwe Kiela: Yes. So I think you just made an interesting point. So one of the things we hear very often is this kind of catch phrase on garbage in, garbage out. And so people really like saying that. I think most people who say that don't really understand what it means. But I think where we are going with AI is -- or where we want to go is garbage in information out, right? So real-world data is incredibly noisy, and that's true for every enterprise. And there is no way you are going to fix that. So people are trying, but I think it's futile. People should just make their AI good enough to deal with the garbage and... 
 

Eric Hanselman: We come from this age of data lakes, right, in which the general idea was all you have to do in order to garner great insights is put all of your data into a data lake. And of course, data lakes wound up accumulating all the kinds of things that regular lakes accumulate, right, as well as a whole bunch of fish and some wonderful vistas and things that are great. There are a lot of bicycles, old tires, shopping carts, and other kinds of things. And those are -- they're a real part of that lake, even though you wouldn't wish them to be. 
 

Douwe Kiela: Yes. Yes. And data lake still, I think if you look at most real enterprise implementations are very heavily focused on structured data, so databases. But most of the company's information is actually unstructured. So it's e-mails, chat messages, their documents, people write. It's all of the sort of activities of employees, the institutional knowledge that is never really properly processed because it's not structured data. And so that's really the promise of AI. If you can do both structured and unstructured data using a smart system, then you can get much more valuable insights out of the system. 
 

Eric Hanselman: Makes sense. In terms of problem spaces, we deal a lot with energy and a lot of what we've been looking at recently. I'm curious about some of the challenges that you see. One of the historical issues is that data sets are very large. The results need to be very specific. It's not similar to, I don't know, maybe financial services, certainly health care and other problems. but it has its own particular aspects. I'm curious if those are areas that you've been looking at. And what do you see as some of the challenges and maybe ways beyond some of the hurdles that many have been facing. 
 

Douwe Kiela: Yes. I think that's a super interesting mark. So as a company, we are focused on difficult problems. So hard engineering disciplines, for example, where you have very high-value problems that you're trying to solve, and a standard language model doesn't cut it. So you can't just take 1 or 2 documents and drop them into ChatGPT, and it will magically solve it for you. These types of problems exist. But I think if you look at the energy or semiconductors or financial services or just the areas that require deeper expertise, there you just need better solutions, especially with the volume of data that you needed to work on and the complexity of the data that it needs to work on. So energy for us is a really interesting sector where we see some really nice use cases around essentially just agentic search. So you could think of something like tracking market moving events across prices, news, outages, different policies, monitoring competitors, looking at emerging trends. So you could think of an example query would be something like show me recent regulatory or geopolitical developments affected sort of offshore wind in Asia or something like that. So very, very specific search queries where you need to understand the data and the context very deeply. -- or we see a lot going on in root cause analysis, so explaining why a commodity price spikes or drops or even in deep research. So if you're doing like M&A diligence, for example, or writing investment memos, yes, doing that is tricky. And if you want to do that with AI, then yes, you need to have very rich context, and that's what we do best. 
 

Eric Hanselman: Well, one of the biggest issues, I think, that we see in terms of enterprises is it's not only looking at what some of the more complex challenges are. In many cases, they're even still wrestling with sorting out effective use cases. It's something that we saw really in the early days of chat and really initial LLM implementations. But now as we start moving towards agentic capabilities, this really is a step in new directions. And in terms of being able to get the application definition right, wrestling with use cases really seems to be one of those areas that really requires much more thought and much more understanding about what organizations are really trying to accomplish. 
 

Douwe Kiela: Yes, absolutely. I think a lot of companies have not been thinking about this properly. So I think initially, everybody was doing some loose experimentation, and it was very easy to bring in kind of ChatGPT or whatever your company's name is GPT and maybe buy a license for glean or something so that you have internal search and then you have like your AI check mark, it's like, I did my job. I brought in AI. But that really is just the beginning, right? That's for very generic productivity problems like writing e-mails or summarizing things or doing basic internal enterprise search. I think the real promise of AI is around these much harder problems where you have subject matter experts doing incredibly difficult things. And if you can have agents automate parts of that, then you can have these very valuable people work on even more valuable things. So identifying the right use cases to tackle there, I think, is something that companies are starting to do now that they've seen that agents are real and that they are actually going to really change how the world works. So yes, I think it's very interesting time. So one of the big problems I've seen with a lot of large enterprises is that they are much too top-down about it. So they think that they -- like whoever the leaders are in this company, they think that they understand all the use cases. And so they will just tell people like these are the 5 or 10 that we're going to focus on because I don't know, they're the biggest cost center, like call center or something like that, like easy things, again, to apply AI to. I think it's much more powerful if you give people a platform like our platform and you tell them you can build any agent you want and you can capture your subject matter expertise in that agent and build it using natural language. So you would literally just describe the process that you would go through yourself. And then you build an agent automatically that automates that process. And if you do that, you can be much more bottom-up and you really put it in the hands of people who directly benefit from the value that that agent brings. So I think that's a much better paradigm to think through use cases. 
 

Eric Hanselman: It seems like we -- as we sort of move along that ladder of maturity for organizations, it's a matter of real understanding what you can tackle. And it's one of those things that we see in our end user data. There are top-down imperatives that are identifying that we will do AI, and we will do it in so many interesting and new ways. But fundamentally, being able to put AI in the hands of those who understand where bottlenecks are, understand some of the challenges, understand some of the toil that exists in existing business processes can be particularly powerful. Much has been said about democratizing the use of AI and agents and all the rest. We start getting into those questions, though, One, how do we do that with a set of end users that may not have a deeply developed set of skills, may not have as much of an understanding. How do you start to expose some of those ideas of what is possible? What can you get access to and what are some of those capabilities? What do you see as ways to be able to surface some of that so that you really can enable that much broader swath of potential end users and start to turn on some of those better delivered values? 
 

Douwe Kiela: Yes, great question. So I think it starts with the data, right? You need to hook up the data or make it easy for people to hook up the data that they want to work on top of. And then you need to use language because language models are pretty good at language for people to describe what they want the agent to do. So we have some really nice examples on our demo website, demo.contextual.ai where you can think of like what does a jet propulsion engineer do for a living. So that's a very complex process. They need to go through test procedures, document them, and do a lot of analysis. And so you can just describe line for line what that process looks like, give that to the system and it uses that information to build an agent that now automates that process. And so I think that's really the right way to do it, right? A lot of developers, and I don't blame them for it, I'm a developer myself. They like building things using sort of in the developer frameworks like a link chain or LAMA index and you kind of write some Python code that automates some of that process. But I think it's much more powerful if you give business users the ability to build custom solutions using natural language on top of the data that they care about. and they can iterate very quickly on the outcomes of this agent. And then they're the end user of the thing that they've built, right? So they understand their problem better than anybody else. And I think that's really one of the big unlocks that we should try to offer to folks who are not software engineers, but maybe hardware engineers, right, or like mechanical engineers or other folks who are still very expensive, highly specialized who don't want to write stuff in that chain. 
 

Eric Hanselman: Well, and be able to simplify that process. I guess in terms of building a data foundation, how do you go about figuring out what you've got for data sources? I mean this winds up being -- again, another thing that we see in a lot of our study data is that organizations back couple of years ago, organizations were kept saying, "Hey, we don't have enough data and then realized that, well, actually, they did have data, but the data was either not where they needed it to be or they weren't sure about the quality of the data in that it was unstructured. They didn't know how to classify it. They didn't know whether or not it was up to date, the temporal nature of the data. They didn't understand the details of what they were potentially exposing. And so they had that moment of sort of really pausing and really wrestling with the data side of this. What do you see as ways to be able to tackle some of that and ways to be able to work through some of the challenges of building what is -- we've gotten into software development pipeline building and, in most organizations, have got pretty good, but data pipelines are much more challenging, especially when so much of an organization's data may not be well understood. 
 

Douwe Kiela: Yes. I think it's back to the point we were discussing earlier where garbage in, garbage out should be replaced by garbage in information out. So the information is hidden somewhere in that garbage. It is just impossible to clean that up. So companies that think, oh, I first need to make my data AI ready, that's just wrong and it's very misguided and you are going to lose. You're going to be what's the equivalent of the last sort of tech cycles. You're going to be the blockbuster of the AI cycle. Really, you need to move. And if you look at the capabilities of modern AI systems, sifting through large quantities of complex data is becoming much better very, very quickly. So you -- this is the context problem, right? So we have intelligence. If you give intelligence, the right data, it can do amazing things. And now the next problem is how do we surface that data through context to the language model. And so this is happening right now. The way we do it is we basically have agents working together. So one agent ingests the data and essentially preprocesses it, looks at all of the different parts of it. So when it sees a document, it says, "Oh, this is a figure, maybe I need to like extract the relevant pieces of information from this figure so that I can retrieve it. It knows what date the document is from. So it automatically knows which one is the most recent. You can tell it to rank data in particular ways. So if you have specific authorities in the company, if you want to upweight the voice of the CEO over the voice of an intern, you can do all of that using AI. So there's this ingestion agent that is processing all of the data as it comes in. And then there are agents on top that are query agents. So essentially, you ask them questions, or you tell them to do something, and they work on top of the data that has been ingested by this other agent. And so every use case has its own sort of pareto optimal, so its own sort of optimal way of ingesting data and annotating data and providing metadata for that specific use case and then dealing with that information during query time so that you give the right answer. So this is happening right now. And yes, people should wake up to that reality, I think. 
 

Eric Hanselman: Well, because it is -- as I was saying, it is that stumbling part. But the ability to be able to apply agentic preprocessing sounds like, one, you do -- you get rid of the hard problem, which is the classification piece. which winds up being the primary concern. 
 

Douwe Kiela: Exactly. The classification that you would need is also task dependent, right? So that's why you don't just want to store it in a data link with some of your classifiers running on top. That's very archaic. You want to have agents decide what information is relevant for other agents to do their job. 
 

Eric Hanselman: Well, and again, to your point, the actual application and use case winds up being that thing that is going to break out of what has been that historical challenge with structure because in structured sources, the big stumbling block was always that you wind up having to have a priority decision about how you're going to actually put this thing together. And that may work at the outset, but of course, over time, that may change. It might not even be accurate at the outset, but you still got to work through the process for each new use of that data. And the one thing you're pointing out is that if you've actually got the ability to classify in ways that are task-specific, you start to open up that level of flexibility, you start to open up new opportunities. You mentioned the process of building context. It's something we talked about a few episodes ago, which is that shift. I think many organizations were just starting to get comfortable with the idea of prompt engineering and really understanding what AI query language really was looking like and how you needed to effectively interact with models. Shifting gears there, you -- when you start thinking about agents, -- we're now starting to look at persistence. We're now starting to look at automating the process that humans were doing before, which is managing the data interchange between individual module queries and actually now start to think about how you build context in ways that are now going to sustain across these various iterations. What's important about context engineering? And what are those things that organizations really need to understand in order to build a solid base? 
 

Douwe Kiela: Yes. I think when language models came out, everybody started doing prompt engineering, as you said, and then they recognize that if you want to use an underlying ground truth data source, then you are going to need retrieval in that data and then give those pieces as a part of the prompt. So that's basically RAG. And I think recently, people have just especially in the new Agentic world, people have started to realize that the context is a dynamic object. You are not going to give the entire enterprise's data to the language model as context for every single question you're going to ask it or every single test you're going to ask it to complete. So what you need to do is you need to have a way to do retrieval in that large data, then give it to the language model, then maybe the language model because it can now reason says, actually, that's not what I had in mind. I made a mistake with my initial query. I need to do it like this or maybe I need to combine sort of the retrieval results from unstructured and structured data and do some reranking based on recency, right? So people started to recognize that we were doing much more than just basic RAG, and we were doing much more than prompt engineering. We were engineering the entire context of that system, so not just the language model, but the entire system that you build around that language model to complete the task. So that is what people now call context engineering. So it includes front engineering. It includes RAG. It includes memory, it includes state and tool calling, things like MCP. All of those things come together to solve this context problem of how do you give the language model the information it needs to be useful. 
 

Eric Hanselman: You've made, I think, what's a really important point in this, which is another big part of context engineering is managing scaling because you can't throw -- well, you can certainly throw all of the data that you have at every single problem you have, use billion parameter models, and we've seen some of the challenges that creates. But one of the other aspects in terms of being able to build in scalability is understanding what that context framework needs to be and adapting it to the problem at hand. And that seems that's going to be just as dynamic as the underlying questions that are being asked of that application to begin with. 
 

Douwe Kiela: Absolutely. So scale is also enterprise gradness, security, role-based access controls, entitlements, all of the things that people take for granted. What we see every single large enterprise that we talk to, they have lots of pilots going on, but it's very hard to get things to really be production grade because you need accuracy, you need all the safeguards in place. You need scale, as you said, and then you need it to have the right security model. All of those things need to come together. And that's why doing it all yourself is going to slow you down. 
 

Eric Hanselman: I think we could do an entire episode just on the security aspects. But it seems like -- I mean, with my security hat on, how do you work to build confidence in the classification capabilities from a security perspective? The big concern for organizations is always, of course, that we're -- historically, we couldn't get to all the data. So that was sort of a protection in and of itself. Now we've got this great hoover mechanism that is going to basically suck up all the data and make it accessible. But classification, of course, becomes all that much more critical. How do you have confidence in terms of building classification capabilities that are going to satisfy enterprises that, that data that you're working with is the data that they really want to be able to lay hands on and work along and reduce those risks of concerns about exposure. 
 

Douwe Kiela: Yes. I think that's the question that's top of mind for every CISO in the country right now in the world. I think actually -- and I'm not a security expert, I should add. So -- but the people I talk to about this where I take their opinion very seriously, what they're telling me is that actually, all the tools are already in place to have a very solid security posture around AI. So we have very good tooling for data management and sort of access controls and all of the things that you need. So the thing you just want to avoid is leakage. And so there's leakage sort of in between applications. It could also be leakage outside of your VPC. So with us being focused on enterprise, one of the things that we do is we can deploy on VPC to make sure that no data ever leaves your security boundary at all. And so these are the types of decisions, I think, that CISOs are constantly wrestling with is how do I give my team access to this brand-new powerful technology, but in a way where I am confident that there are not going to be any negative consequences. 
 

Eric Hanselman: Yes. It is a big concern. And well, actually, as we start thinking about concerns, I want to do a more topical question. I was wondering, of course, we've seen a lot about what's going on with the open cloud community, the Ve book, all these other kinds of interesting new shifts. And of course, a lot of the concerns that are being raised are like, oh, no, all the data is going to be exposed. It hasn't been helped by at least some of the challenges that the Open claw community has faced in their platform security initially. But I'm curious what your take is and what your thought is about where this is headed. 
 

Douwe Kiela: Yes. It's been fun for me to see everything happening around that. So I think open claw or Cloud bot as it was called, is actually very interesting. It's just how do I make things work locally, right? So if I give local tools and then I have a language model that I send information to and sort of let the agent manage my local computer, like what happens? So that's a very nice thought experiment. And as it turned out, that can do pretty interesting things, and that's why it went viral to begin with. But then, yes, I think that piece of software was maybe vide coded itself. I don't know, maybe Claude wrote most of that, but there were something... 
 

Eric Hanselman: What could go wrong, right? 
 

Douwe Kiela: So yes, that was unfortunate. But I think that just happens when somebody just post a cool open-source project. I don't think it was designed to be enterprise grade or anything. It was designed to be a hobby project. It just happened to go viral. And I think people should have been a lot more careful giving something like that, their credentials, including sort of saw that Karpathy had this like Twitter token exposed and things like that. That's not very security conscious of him. But -- so I think the moebook side of the story, I'm just surprised how those types of things keep going viral. I think it shows a lot about people not really understanding AI at all. So there's a lot of entrepreneorphization happening around AI always. That has always been the case. But this reaction to Mldbook was pretty laughable. I think the posts on Mebook that people were talking about where they were like, oh, my human is like abusing me or something like that, those are written by humans 100%. So... 
 

Eric Hanselman: Yes. And we've now actually got what seems to be -- because access was not actually really validated, yes, all of the robot overloads are going to go take over the world and destroy humanity. 
 

Douwe Kiela: As a human, you can just prompt your language model like post some slop around humans being annoying and then it would just post it. So I thought it was silly that it went viral. What was interesting, though, is it reminded me of a time -- a long time ago when I was just starting in AI research in, I think, 2017, some of my colleagues at FAR published a paper where they were working on emergent communication in multi-agent systems. So I've also done a lot of work on this area where you have agents. This is before language models where these agents can sort of talk to each other. And what they found was that if you let these agents talk to each other, talk to each other really just means like sampling from a neural network is the technical term. So there's nothing sort of human-like in that communication. But what happened was when you trained that system, then it would obviously diverge from natural language because there's no reason for that system to produce natural language. So when it was asking for like 3 bananas to the other agent, it would just say III or something like that. And then the other system would know, okay, it's asking for 3 bananas. So it's a very uninteresting finding in a way. Until the daily mail, I think, discovered this archive article, and wrote this big piece around like Facebook AI researchers have discovered that AI discovers its own like secret language. And that was the end of the world sort of narrative. And so the same thing is happening with Multbook, where just people who have no idea what they're talking about are sort of overinterpreting, I think maybe for some of them even deliberately sort of the impact of this piece of technology. So it's like nice headline, but I just -- I wish people could be a bit more critical when they look at headlines or maybe the slightly less reputable journalists out there. 
 

Eric Hanselman: Well, it certainly has grabbed the popular emotion because it feeds into so many different concerns and yet you've got a basis that I think we can all reasonably agree is maybe an interesting toy, but a lot further from the real benefits and the capabilities that we can build with well-architected systems. But... 
 

Douwe Kiela: Yes. Yes, it's -- I mean it is possible that AIs can talk to each other, right? And they, I mean they're trained on language, so they should talk to each other in language that we can understand. And so that would be a constraint that I would put on that system. But for the rest, I'm not too worried about them worrying about their humans being maybe a few years from now, not yet. 
 

Eric Hanselman: We'll have to see what happens then. Well, this has been great. I appreciate all your insights. Thank you very much. 
 

Douwe Kiela: My pleasure. 
 

Eric Hanselman: That is it for this episode of Next in Tech. Thanks to our audience for staying with us. And thanks to our production team, including Sophie Carr, Feranmi Adeoshun, and Kyra Smith on our marketing and events teams. Please keep in mind the statements made by persons who are not S&P Global employees represent their own views and are not necessarily the views of S&P Global. I hope you'll join us for our next episode because there is always something, Next in Tech!