#328 The Challenges of Enterprise Agentic AI with Manasi Vartak, Chief AI Architect at Cloudera

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At datacamp, we offer everything you or your team need to adapt and thrive with AI. Whether it's business users looking to get the Most out of ChatGPT and Copilot, or developers and data scientists looking to fine tune models. You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today. Start learning@datacamp.com I think with all the experimentation we see folks leading with technology versus leading with a business problem. And often that's a recipe for not being able to follow through. So it's experimentation for the sake of experimentation versus I'm trying to reduce the time required to do this kind of processing, or I'm trying to expand our TAM to target this market. So I think that's one of the challenges. How do we find the use cases that have business backing that are actually the right use cases for gen AI? I think data is a part of AI governance, but AI governance is a much bigger sort of umbrella. It does start with data. Where did the data come from? So if you think about the lawsuits that currently Anthropic and OpenAI are facing, they have to do with what data did you use to train the model. That's a data governance and lineage issue. Like if you could trace it down and say this model was trained on these 10 data sources, it's an easy question. So the challenge is the data goes through so many steps of pre processing and post processing that by the time the model gets to it, it's hard to figure out where it came from. For AI governance, we need to solve those kinds of fundamental issues around data governance.

Welcome to a special edition of Data Frame recorded from the Cloudera Evolve Conference. This is Richie. AI agents have been hyped all year as the next big thing. Given the level of enthusiasm, you'd think they'd be everywhere by now. However, in an enterprise situation, there are still many hurdles to overcome. Governance issues are top of mind and privacy in particular is a major concern. Today we're going to explore private AI and other solutions to making agentic AI work in the enterprise. Today's guest is Manasse Vartak, the Chief AI officer at Cloudera. She leads product and engineering teams who build tooling and infrastructure for AI engineers. Previously, Manasse was CEO and founder at the enterprise machine learning infrastructure platform Virta, which was acquired by Cloudera last year. Let's find out how to get agents into the enterprise. Hi, Manasseh. Welcome to Data Framed.

Thanks so much for having me, Richie.

We've been having a lot of fun here at Cloudera Revolved. Can you talk me through what have been your highlights so far?

Several. New York is probably the most exciting city for us because Cloudera serves a lot of financial institutions. And, you know, New York is kind of the epicenter of all of that. Lots of great discussions around how AI is being used in financial services, also insurance, oil and gas too. And it's been really exciting to see how deep the AI adoption is going. It's not just the technical users. We got the business users. So we have a whole session packed with AI this afternoon. So I'm really looking forward to it.

Yeah, lots of exciting things going on. Certainly in the finance world, I think there's a lot of companies that are getting very deep into AI, so perhaps we can touch on that later on. So we've had three years of the generative AI revolution. Feels like all year everyone's been talking about AI agents. So I've got a conundrum for you. Why aren't enterprises just filled with staff full of bots yet?

I'm gonna maybe put that in perspective a little bit. We say it's three years, but it's only three years. If it were a cat, it would still be a kitten. You know what I mean? It's like it's baby right now. So when I think about something, say, digital transformation, it's been around for 15, 20 years and we're still in that journey. And so when I put the gen AI revolution in that perspective is like, we're still in the early days and there is a survey making the rounds and you know, hey, 95% of POCs aren't going to production. I kind of look at it as we're in the experimentation stage. We're still trying to figure out where does this technology work, where are the rough edges that we need to still kind of, you know, figure out how to navigate? But given how early we are, I think we're seeing some amazing results all across the board. So whether you're thinking about coding agents who can now pretty much code fairly well all the way to. We're solving math problems and we're doing legal contract review. And so I think we're seeing very specific use cases getting great results. And I think a lot of us use GPT to kind of finesse all those emails. We're like, okay, please make this so that I don't regret it tomorrow. So I Think it's having a lot of great impact already and we're just at the early days.

I think it's a valid point that maybe three years isn't that long in terms of enterprise cycles. Like with some enterprises stumbling mainframes and soapy businesses are still built on top of Excel. So things take a long time to change. Can you talk me through what are the main challenges you see with enterprise AI adoption?

There's a few different categories. I think one, I might separate them as business and then technology and there's also change management piece to it from the business side. There's I think with all the experimentation we see folks leading with technology versus leading with a business problem and often that's a recipe for not being able to follow through. So it's experimentation for the sake of experimentation versus I'm trying to reduce the time required to do this kind of processing or I'm trying to expand our TAM to target this market. So I think that's one of the challenges. How do we find the use cases that have business backing that are actually the right use cases for gen AI? So that's one technology wise. I think there's a couple of key areas. One is LLMs already know the public intranet. There's nothing that you know you can teach them there. The enterprise data is where is what's going to make these LLMs useful for enterprises. And there the challenge is how do I connect this LLM to my enterprise data in a way that preserves privacy and is safe so that you know that if there's a question about, oh tell me the salary of my colleague, it's going to say no, no, you can't ask that kind of question or you only give the right users access to the right data. So I think it comes down to how do you make sure that the LLMs have the access to the right data and it's permissioned in the right way. So I'll highlight that one on the technology side. And LLMs are not like predictive software. If you ask it the same kind of question two times, it might give you a different answer. And so how do evaluate these AI based systems or agents is still very much an open sort of open question, even research. So that's where AI evals. If you've heard about it, it's like, oh, how do I evaluate my systems to make sure that it's making the right decisions? Still big problem. And I think that's one of the areas that's limiting the adoption because you need to build trust in the system. The last one I'll mention is change management. Because imagine the first time that employees started to use Google at work. It's a muscle you need to develop is like how do I ask Google the right questions to get me the right answers? How do I process the results correctly? Similarly, how do I prompt, how do I structure the prompt in a way that it gives me the outputs in a way that I expect? So I think that's a skill that employees currently need to learn. So unless there is concerted effort to train employees to use AI, that's where we're finding that that's also a bottleneck.

But certainly the first one you mentioned, where it's very easy to just dive into AI is going to solve everything. Let's go with the technology first and not think about what business problem you're trying to solve. That seems to be a very common occurrence at the moment. So you brought up some good points there around some of the challenges around privacy with AI. What do you think those main challenges are?

There's a few different layers to this one is, is your data structured and properly, is it architected in the right way? This might be that in certain GEOs your data needs to be on prem. Perhaps in certain GEOs it can be on the cloud. Do you have the right data ingest mechanisms? Do you have the right processing? So the basic data architecture and how well it's set up is going to enable your AI to leverage that data much more effectively. So that's like, that's just a foundation before you can, you know, go to the next level of AI. Another key piece is metadata and governance and that's where I would put the permissioning piece in there. What kinds of data sources do we have? Where did they come from? Do we have PII in there? If we are going to send that data to a model, do we have ways to mask the PII so that we don't leak some data inadvertently to the third party? So at Cloudera we think about what we call private AI, that means it is your data, it is your models on your hardware. It never goes off to a third party. And we found that to be a really good. When you have super sensitive data that you're working with that cannot say leave your VPC or premises and that gives the most sort of stringent of enterprises comfort that AI is you're not going to leak your data to the AI. So those are some of the things that we're seeing. The governance piece, the data architecture and then the last one I'll also mention is AI governance, which is almost an umbrella term because you need to do your data governance. Then you add on your AI governance, which might be things like is the use case for this application of AI? What about the ethics of that? Are we producing results that are biased? Did this go through the risk management process? So all of these pieces built on top of each other. And I think enterprises are slowly marching towards the future where that stack is well defined.

Okay, so even more problems. I feel like we're generating more problems than we're solving. But one thing that I've never considered before, you said that you might need different setups for different geographies. Can you talk me through why that is?

Oh, 100%. And we see this a lot. So Cloudera works if you pick any vertical, I think the top 5 of 5 of the top 10 or 8 of the top 10 in certain verticals are customers of Cloudera. And these are necessarily multinational companies. So you might have a bank which has a presence in the us, the uk, in Asia, and then maybe they also have a presence in emea. And so the Middle east, certain areas like apac, there are government regulations because of which say financial data cannot live in the cloud. And so it does need to live on prem. If you go to a market like the Middle east, all the processing that happens on customer data needs to happen there versus being shipped off to the US So soon you have an architecture where your data resides in different geographies. That's where your customers are. And so you want to get your AI models to your data in each of those, those geographies. And that's required by a lot of times by regulation, but also latency because if you are sitting in India and you're shipping your data to the US to get a prediction, it's going to be a lot slower. So it's also practical considerations there. But we're seeing that a lot and we think it's a growing trend.

Okay, so it's really about dealing with different laws around the world, but also just about making sure that your application performs well wherever the users are in the world.

User experience. Yeah.

Okay, nice. So data privacy has been around for a while and I'm going to say it's a solved problem. But do you think that the challenges around AI privacy are the same as those in data privacy, or are there some new challenges there?

I think data is a part of AI governance, but AI governance is a much bigger sort of umbrella. It does start with data. Where did the data come from what was the lineage of that data. So if you think about the lawsuits that currently Anthropic and OpenAI are facing, they have to do with what data did you use to train Model. That's a data governance and lineage issue. Like if you could trace it down and say this model was trained on these 10 data sources. It's an easy question, I think. So the challenge is the data goes through so many steps of pre processing and post processing that by the time the model gets to it, it's hard to figure out where it came from. And so for AI governance, we need to solve those kinds of fundamental issues around data governance and then layer on things like for this particular use case, suppose, let's say HR use case, should we be using AI chatbots? So it's an ethical consideration. If we stick to the HR use case, is it giving the same results to men and women, to people of color? So the ethics and the responsible AI pieces also go in there.

And then Claudia has been talking a.

Lot about because LLMs, if you're not doing AI, you send your data to a third party. The data privacy there is even more important. Is like, like what data is anthropic or OpenAI getting from my most sensitive estates and then having governance on there. So it very much there's a data piece, there's the model piece, there's also risk management typically. So financial institutions have a fairly well defined risk management practice and so the AI governance kind of needs to pull that along with it as well.

Okay, so it sounds like you really need good sort of data governance in place if you're going to have good AI governance or it's a sort of a cornerstone of IT and then extra things on top of that. Yes, and we're just still adding to the list of challenges.

Now we know why they say 95% aren't quite making it to production.

Absolutely. Let's talk a bit about the technology side of things. So for private AI, what does the architecture look like?

So maybe to recap, like private AI, when we think about it at Cloudera, it has three components. It has your models, which means typically open source models or models that you can fully control because then you can fine tune them for your needs. You can also distill them, like make them smaller as needed. So your models, then it is your data again to going back to the data piece and then you're running it in an environment that you control. So this might be your data center where you control the hardware fully. It might be your cloud VPC as well. And then when you have those three and you're running in a secure environment, we call that private AI. That's where you know exactly what the boundaries are for data sharing or potential data leakage that might come up.

Okay, so there's a few different layers in the stack there. So you're talking about having your own data center as well, like as I guess the, the servers and then you've got the software on top and so you're controlling everything there. Do you need to do that full stack all yourself or are there other options where you're doing part of it yourself?

Ah, you mean like in a customer scenario?

I think so, yeah. So do you have to. I mean, I guess not many organizations are going to want to build the whole data center themselves. Can you do some bits completely private and others others more public services or buy in services?

I think so. A lot of we work with enterprises that often do have their own data centers. So if you think about the biggest banks, all of them do have their own data centers. If you only have very light workloads, should you be standing up a data center? I think the answer is no. It's very expensive and also human labor intensive. You don't want to be doing that. But we actually did an interesting case study of a bank in Asia Pacific where we in collaboration with them we modeled out like when does it make sense to have inference in particular. So like running the mod, the models running on prem and for them the sweet spot was when you go beyond say 100 models. That's when it makes sense for you to be running it in your data centers for it to be cheaper. If you are at a smaller scale, the cloud might be a more flexible and faster option for you to get started with. In that case you don't need to manage the full stack. You're just spin up me, spin me up machines on AWS or GCP and there you go.

Okay, so the scale is like really important there. So small scale cloud provider is going to be I guess cheaper because quicker to set up and you don't need to build your own data center. But for large scale, once you're running things yourself, you're not just paying a fortune to all the cloud providers.

Yeah. The caveat though is if you are a DOD or like in an area where you cannot use the cloud for some reason, then you're going to have to be on prem. Even if your scale is smaller then you might be working with data that's super sensitive.

So are there any things you can't do. If you need to make use of private AI, I think you.

So if you're in the private AI setting, the main thing is you don't want to send your data out of your secure environment. And so you. These would be use cases or verticals where you cannot just send your data to a third party. So that's where you would not want to. If your use case is that sensitive, then you want to stick with private AI versus sending it off to someone else.

Okay, so this is like no sending things to like the OpenAI, anthropic, Google, whatever APIs. It's like you run your own large language models, I guess yourself.

Yes, yes, we do that. So while Clara offers private AI, we also know that a lot of customers are going to be on the cloud and not all use cases are that sensitive. Right. Maybe you are rewriting a blog post. That's okay. Send it to whatever your favorite third party LLM provider is. That's all right. But if you're doing that with your illegal contracts then, or customer information, you need to be a lot more careful. So hopefully that's cool.

Do you have a sense of which models people are using in those more sensitive cases?

A lot of open source stuff. So OpenAI recently also open sourced their GPT model. It's not their latest and greatest ones, but it's a pretty good model. Google has their Gemma models out there. There is of course Deep Sea that created a lot of waves. IBM has their Granite models. So open source is really catching up. Proprietary models. And if you're going the open source route, I think right now you have a lot of different options that you can choose from.

Okay, plenty of choices there. So just I guess try a few and see which one fits your use case.

Yeah, and if I can build on that for a minute, I think benchmarks are great to a certain extent, but public benchmarks can be gamed. So you can train your model to be really good at a particular benchmark, but that doesn't mean that the model is actually the best one for your use case. And so what we recommend to folks is do an apples to apples comparison on your use case. It generally doesn't matter what the public benchmarks are saying because the benchmarks might not reflect your reality.

That's really interesting that that's every press release for every new model. It's like, wow, we've just, you know, beaten some math challenge or scored really high in the coding challenges. But actually, yeah, you, you want to test against your own use cases. Because it's, it might be nonsense in those cases.

Yes, yes. And it's just, that's a good and bad about public benchmarks. Right. They are public and so it's possible to game them. If it's your use case, then the likelihood of your data being in their sort of training set is lower. And so that's a better example.

Let's talk a bit about processes. Are your workflows going to be different if you make use of private AI?

I would say process generally with AI is a little bit bit different, whether it's private or you're doing your third party cloud stuff. Maybe I'll pick an example of building an AI capable product. Suppose you're building a product that is a travel assistant. So I'm going from SF to New York and this assistant agent is going to go off and it's going to book flights for me. When you're writing deterministic code like when we did, like Rails app or whatever app that you're building, you could predict exactly what the outputs were going to be. Or building test cases around it. With AI, it might give you only United flights one day because God knows why, or it might give you United and Delta and JetBlue the next day. So what you then need to do is to write test cases, evals or AI evaluations that make sure that your results are always have the characteristics you want. Like I have all the airlines represented, these are starting at the right airport and they have no layover. So this is a different kind of how do you write test cases? But it's super critical to building a useful AI application. And the really kind of fascinating thing for builders right now is the model is, you know, let's call it 70% good today. In a year it's going to be 80. So do you build for the 80 or do you build for the 70? Because the boundary always keeps shifting. And if you want to build a business that is long term, you can't just be targeting the 70, you need to be targeting the 80 and the 90 and solving the problems that will exist even as the models get better. So I think if I've started a company in the past, if you're a founder right now, you really need to be thinking hard about, okay, am I just solving the problem for today or will this be a problem three years from now? And I think that's just a very different mental model.

Absolutely. When the underlying tech infrastructure is changing so fast, it's very easy to solve a problem. It's like, oh, actually this has already been solved by everyone, no longer relevant. And then, yeah, that's, I guess that's the point where you pivot your startup. You mentioned that it's quite difficult to evaluate the, the results of AI when it's, particularly when it's not deterministic. You get different results every time. Is there some sort of benchmark you need to target then for like, oh, it has to be right 95% of the time, or like, how do you decide how good it needs to be and how do you make sure that happens?

So with, usually how it goes is you start off a particular project, you have a set of evaluations or test cases, and it doesn't need to be thousands. Andrew Ng, you know, one of the godfathers of deep learning is like, you need to start with a handful, which is what we found too. And handful means at least start with 10 or 20 while you're getting off the ground. So you can get a feel for where is the model and AI failing versus not, not soon. You're going to, you're going to give it better instructions and then it's going to get better on that test case. And then you're like, oh, it's working well here. I need to add these different kinds of test cases in because now that's a new kind of say, data that you're seeing. So then your test set expands and then it expands even further. So ultimately you will probably have thousands of evaluations, but you don't need to start, you don't need a thousand to get started. Those usually build very organically and over time.

Okay, that's kind of good in theory. Like if your user can ask any question, you've got sort of infinite test area, so you can just keep writing tests forever and ever and ever and never get there. So I like the idea that you start with like 10 or 20. Just see does it work on the basic use cases and gradually build up from there. Okay, so do you have a sense of like, what skills you need in order to be good at building with AI?

It's a great time to be working in AI. The skill sets kind of, you know, they, you need to be on your toes to be working in this field. I see a couple of trends. So if you are a build, if you're building with AI, let's call it building versus using. And I can also touch on using a little bit. If you're building with AI. I found that the quick prototypers, great skill prototyping because you just need to get something out there. See if it works or not. And then iterate. And your coding assistants also help you a lot with it. Right? Vibe coding is, the whole thing is like you can say, all right, spin me up a webpage that is say the travel assistant example. Give me a UI that lets me pick the start city and the destination city. You can do that in a couple of minutes without writing a single line of code. And then the copilots are going to do that for you, which is great. And then you keep iterating from there. So I think prototyping is really key. User experience is getting really, really critical here because we as humans are trying to figure out how do I use, how do I effectively weave this intelligence into my workflow. OpenAI has their canvas UI to like for editing. Anthropic has their own ones too. We're trying to figure out where do we, where is natural language the right user experience and when do we need checkboxes for things? I'll give you a fun example. So I use GPT to keep track of my grocery shopping list. I'm like, so tell me what I need to buy from Costco. Great. It tells me things, but then it doesn't let me check it off there because I'm like, no, no, don't tell me. I just, just show me the list and I'll check it off. So these are like small affordances that we're still trying to figure out when do we use natural language, when do we use existing GUI capabilities? And so if you're in hci, human computer interaction, you're on UI ux. I think that's amazing. And those skills are gonna be so critical. So those two. And then maybe I'll just throw out one on using AI. All of us need to be using AI and so getting comfortable and actually kind of good at prompting so that you can tell it, oh, I want abc, don't give me X, Y, Z. I think that's going to be a base level expectations regardless of job function. And that's where we're seeing a lot of the change management that needs to be done in companies is educating the broader workforce on, here's how you use a Gemini model in your workflow, here's how you use a GPT model and we're seeing a lot of benefits from that.

Great point on user interface there. Because I think there was a point after ChatGPT came out, everyone was like, oh, this is the final interface. Everything's just gonna be a chatbot from now on and actually being able to like Just point and click on stuff is still really quite useful a lot of the time. Yeah, I like just having checkboxes and radio buttons and all the rest of it. Yeah. So you mentioned the idea of prototyping being really important. I imagine at some point you want to put things into production. Does it need to be someone different putting things into production compared to prototyp? They feel like different skill sets.

I was gonna add in the fourth one there when I was listing out the three and I was like, you're gonna. Maybe I'll so 100%. I think you need folks who are going to build critical production systems or even productionize the prototyping. I think about it as a distribution. So if you think about a normal distribution, your X axis is like, how deep are you in technology right now? You might assume it's a normal distribution. Bunch of people are in the middle. There are a few people who are not super technical, and then there's some people who are deep tech. With AI, I feel like the middle is getting flattened out. You either have deeply technical people who are going to need to understand the innards of these systems, to build them at scale and to make sure the AI is working correctly. And then you have SMEs or like business experts who can get by with quickly building workflows, prototypes. And so I think we're going to see that bifurcation a lot. So it becomes a bimodal, like super technical or super business, because super business are going to have the AI capabilities anyway. So it's going to be fascinating how the how skills and expected skills change.

That's very cool. Somehow, even though we're a data show, we don't often talk about bimodal distributions on the show.

I love that.

That just cropped up naturally. That's wonderful stuff. So you mentioned that there's going to be of an interaction between subject matter experts and people who are building with AI. So this strikes me as very similar to the way, I guess historically we've had the same interaction between subject matter experts and data teams. I'm curious, like things like machine learning, AI have traditionally been a data team thing. They're now moving a lot more towards software development teams. Do you see the relationship between software and data changing as well?

I do, I do. And we were having this conversation at the conference actually on like, what happens to this, the business analyst Persona? Do they need to become more technical or do they need to become more business? And I think I do lean back to the same sort of trend that I'm seeing where I have friends who are chemists, never in the line of code in their lives. They can build workflows now using OpenAI or you know, Cloudera launched this thing called Agent Studio, which is natural language build your agents. And we've seen folks who have never coded do this. And so the SME who historically used to come to the data team to get information now can do that a lot or should be able to do that a lot on their own, hopefully freeing up the data team to work on deeper technical challenges about how do we better architect things, how do we optimize how we're running our queries? And so I think we will see more specialization in those areas. I think hopefully much more, many more opportunities that way.

Okay, so the role of the data team then is going to really just I guess enabling the rest of the business to use data more easily rather than actually going and solving the analytical problems themselves.

Yeah, I think like we get to a more self service model which is where I think the movement has been as like data visualization is more self service, analytics is more self service. This helps us get to that feature. Because usually I feel like most data teams are like no, no, I got too much stuff to do, please ask me another quarter. This way they can focus fewer things that are going to be fundamental to the data architecture.

So I know Cloudera has a lot of assistance for helping with data tasks. Can you talk me through the difference between agents and assistants and co pilots and all the rest of it?

So I think the future that we're going towards is the data platform has AI embedded in it everywhere that it makes sense to do so. For example, later today we're doing a demo off a data classification agent that we've built and that goes to the metadata collection, the governance collection. This activity might have been on the data team. Plus SME is now you can build an agent that's going to automate it, making the data team's life a lot easier. There are also ways to going off of the lineage bits is how can we look at the data in the tables, figure out how do we document this data? Because usually you know the typical thing is you have five columns with the same column name and then you have no idea what they are. So those are places where AI can start providing some insights organically. So this might come from this data was written to from this particular script so you can add some semantic meaning to it. And Octopi, which was a recent acquisition by Cloudera, their whole product was around building data Lineage, so they can extract lineage from scripts and you can start enriching your data sources without human intervention, which I think is awesome. Another thing I'll throw out there is also you can optimize your data systems using AI. Now, there was, there has been previous research on how do you build indexes as you go, so you don't need to predefine them. I think we're going to continue to see those kinds of innovations every. I think if someone's trying to solve a hard problem, see if ML can do it, it doesn't have to be Genai. See if ML can do it, see if AI can do it. And I think the answer is going to be maybe 50% you can. And that's a great win.

Okay. I do like the idea of just see if technology is going to solve your problem for you. And some of the use cases you mentioned, like metadata collection, documenting data sets, these are not fun tasks for humans. So I think, yeah, they're sort of prime candidates for being outsourced to technology. So you talked a bit about data lineage. Do you want to explain a bit about like, what is data lineage and why is it important?

100%. So if you think about, let's start from the end product, maybe it's a visualization. You know, it's a bar chart somewhere. If you trace it back, the bar chart came from some table that was maybe joined. Two tables that were joined. Now those tables might have come from a data cleaning process from yet another table, maybe that came in from a streaming system. Maybe there was another kind of function that was applied in the streaming system. So what lineage does is it starts from where the data first entered your system and then traces it through all the transformations to the end product that you're tracking. And so that gives you a view of who made what changes and answers questions like, oh, that's odd, I'm seeing this odd result. Where might that have cropped up? Or if you're in the governance setting, I have how did this PII thing make its way to the visualization when we shouldn't? So it's kind of critical to both making your governance and metadata work, but also operational things. My pipeline is broken. Okay, Your lineage tells you what are the places you should go check?

Okay. Yeah. So I think it's a very common problem when like some executive looks at the dashboard and like the numbers are wrong here and why the wrong on the data team, but not really sure here. And it turns out it was like last week's data or something. Like that. So yeah, I can certainly see how being able to track where the data's come from, it's going to give you those quick answers for when users start complaining about dodgy data.

You know, building off of that. If you had an AI assistant, you were like, hey, tell me why this doesn't look. I expect this to be 50 and it's 100. If you have the lineage in principle, automation could go back and compare to. This is what it looked like last week, this is what it's looking like this week. I think this might be the issue. We're a little ways away from that, but I think we are with more product building we can get there.

Okay, so you mentioned the idea of an agent there. And when do you want like a full AI agent versus an assistant versus a copilot? I mean there's all sort of different bits of terminology and they blur a bit. But talk us through, when do you want which.

So agendic AI is really loosely defined right now. And again, going back to Andrew Ng, he actually coined the phrase because he found that a lot of the community was squabbling over is this an agent or is this not an agent? And he's like, okay, that's not useful. Let's just call it agendic AI and let's move on. And I kind of see it that way is there are different degrees of automation building in built into our processes. Sometimes the automation is just one and done. As I ask something, a question, it gives me an answer, okay, great. Sometimes it needs to do deeper thinking or reasoning. It might need to use tools. And that's been a big unlock, right? MCP is like, how do we standardize how LLMs call tools? So whether. So that's on the other extreme where think about the international math Olympia problem problems, any very deep reasoning. And so that's the other extreme of agents which actually do reasoning. But it's a spectrum and depending on, depending on how well defined your problem is and how hard it is, you might be better served with a deterministic workflow. So think about deep research. For instance, you. It's fairly deterministic. I need to get, I'm trying to do research on say AI evaluation. It's going to go off. Look at 10, 20 something sources on the web. It's going to collect that data, it's going to summarize it, it's going to get it back. That is a fairly, I would say linear. But then if you want to get to a autonomous agent for coding, like what the Devin was supposed to be that is much more iterative and a lot harder. And so that's where depending on what you're trying to solve, you might be well served with a really simple solution. And that's okay. You don't need to go with the, you don't need to get a sledgehammer if you're just trying to put like a nail in the wall. So making sure you have the right complexity of solution for your problem I think is how I think about it at least.

Okay. Yeah. So there's definitely a big variation between just doing some sort of simple like software workflow and there's a bit of LLM magic sprinkled in there versus I'm trying to like. Well, you mentioned devit trying to create an entire software engineer here using AI. Okay, so are there any tests you think should still be done by humans then shouldn't be outsourced to AI at all?

I think it's critical to have human oversight over all of these processes. And it doesn't need to be real time necessarily. But while we build, continue to build trust within these systems, you do want a human reviewing the results. So for example, suppose you were picking off the chemistry pipeline example. You might get a molecule, you figure out its configurations, you do some testing and you see the results. You still want a chemist to go and look at either surprising results or some subset of results to see. Okay, it's doing what I expected it to because strange things will crop up where maybe the units are off in one of the data file that you uploaded and then everything looks odd. So those are the kinds of things where having a human in the loop I think is critical. A lot of these systems are not set it and forget it yet. So even if it's an audit after the fact, I think it's pretty crucial and as with all of automation is like, what if we get this wrong? What happens? And if the cost of getting it wrong is really high, you better have a human like handholding the thing. If the cost is like, hey, yeah, I might need to redo this analysis quickly, that's okay, maybe they can do it offline.

Okay. Yes. We really got to think about what are the consequences of things going wrong. And I guess is the human actually going to be good at solving the problem? So just to wrap up, can you talk with you what you're most excited about in the world of AI every.

Day there's so many things going on. I do think that agents are, are here to stay. There's a lot of work to be done there. Another thing is how do we go from Vibe coding to production? What we kind of touched upon is like, are they different people? And how do we help more Vibe coders also transition to writing real production applications? I'm excited about that. We have seen general purpose models with GPT, but then having a model for healthcare, having a model for drug discovery, I think that is to be another game changer. And so I'm excited about all of these. And again, getting nailing the human interface, like, how do we build the AI so that it makes the human's life easier versus there's like a lot of slop is what it's called, like generated by AI. How do we not increase that? How do we still retain human attention for the areas that are critical? I think those are broadly the things that I'm interested excited about.

That AI slop, definitely a problem. It's being flooded everywhere on social media. I'm not sure we're going to be able to stem the tide. But do you have any advice on how to deal with AI slop?

Couple of things. One is the authority of the person writing is going to matter a lot more because if I trust someone historically or over time they've built that trust, then I'm going to trust their content more. And with AI, when everyone can use GPT, I know that this person has really interesting ideas and so I'm going to hear what they have to say. So the authority and those human relationships actually I think matter more in this era. And there was this really interesting model by One of the VCs out there actually from Theory Ventures, and his thesis was, in the future, you just ask an assistant a question, the assistant goes and almost gets quotes. You know, if you think about an airline or an agent trying to book something, historically they were going to get quotes from airlines and then surface them. So the feature that he was talking about which resonated with me is like, if I want to get a question answered is like, how do I do AI evals? It should run a bidding, so to speak, on like, who are the experts? Go get me the expert opinions and just get me the results. And that's a really different mindset. And the question is, what happens to the World Wide Web and the Internet? Like what happens to the blogs out there? So it's a very different worldview and it's going to be fascinating to see how that evolves.

Like a marketplace for quotes and ideas. That's kind of interesting.

Expert marketplace, right. That's what we're kind of been trying to go towards and this just. It's on steroids, so to speak.

All right, wonderful. Yeah. An exciting idea for the future. And finally, I always want more people to follow. So whose work are you most excited about right now?

Now I'm plugged into research more than others because I came from academia. There's two folks who are doing great work on the human computer sort of interaction piece. Currently both at Berkeley. One is Shreya Shankar. She's done great work on Doc Wrangler. How do you get a human in a loop and help them educate the AI on how to clean data? So that's awesome. And Aditya Meswaran is her collaborator there. He's the PI. They've both done awesome work and I think they'll continue to too.

That sounds very cool. Just figuring out like where do you actually want the human where? Don't you? Yeah. Very topical question for our time. All right, wonderful. Thank you so much for your time, Manasseh.

This was great. Thank you so much for having me.