How AI is reshaping Wall Street

Note: Transcripts are automatically generated from episode audio, and are not fully corrected for spelling, grammar, and formatting.

MAITHRA RAGHU: I think, paradoxically, as the AIs get better, the thing no one’s talking about is that we’ll need more and more people in the loop so that we can make sure they all have the right frameworks for judgment and all of that.

RANA EL KALIOUBY: It’s kind of a big mental and emotional shift to move from research to entrepreneurship, to leave big tech and go start from scratch. What was that transition like for you?

RAGHU: I always say the motivation for starting a company is really, really important. I think it has to be quite pure in some ways. You have to leave not caring too much about the exact outcome or where it goes.

EL KALIOUBY: If you could wave a magic wand and have a personal agent that can do anything for you, what would you want it to do?

RAGHU: My goodness. Probably clone myself and be me in all of these places.

EL KALIOUBY: Maithra Raghu is the co-founder and CEO of Samaya AI. They’re building AI systems for the financial sector.

Think hedge funds and investment firms like Morgan Stanley. Their AI agents tackle complex analysis that financial professionals do every day, which raises some big questions even if you don’t work in finance. What happens when an AI agent can do your work? And when billions of dollars are on the line, how do we ensure that these AI systems are accurate?

Plus, how close are we really to seeing personal agents at scale? Today we’re digging into those questions and a lot more, so let’s get into it.

I’m Rana El Kaliouby, and this is Pioneers of AI, a podcast taking you behind the scenes of the AI revolution.

[THEME MUSIC]

EL KALIOUBY: Hi, Maithra. Welcome to Pioneers of AI. I’m so excited to have you on the show.

RAGHU: So excited to be here, Rana.

Copy LinkHow a global upbringing and math training shaped an AI founder

EL KALIOUBY: First things first, we both went to Cambridge. I guess you did your undergraduate degree there in mathematics at Trinity College. How was that experience?

RAGHU: It was extremely intense, is the truth of it. Trinity attracts all these folks who have competed in these international math competitions, these Math Olympiads, which are now in the news again because AI is making progress on them.

That’s been especially exciting for me, seeing parts of my past and parts of my present come together in this way. My story of getting into Trinity is that I was really into math in high school.

EL KALIOUBY: Did you grow up in the U.S.? You grew up all around, right?

RAGHU: I grew up all around.

The bulk of my life was split between the U.S. and the U.K. When I was in London, I discovered these math competitions pretty early on.

I got really excited about these Olympiads — the reasoning, the complexity, the abstraction involved in them, the problem-solving. If you compete internationally in an Olympiad, that just takes you straight to Trinity College in Cambridge. It pulls everybody across Europe who’s competed in these Olympiads.

Some people go on to pursue mathematics research of various kinds. Of course, many people end up in London working in financial services, especially in hedge funds and lots of different types of investment firms.

Copy LinkWhat industry research taught her about building AI in the real world

EL KALIOUBY: Before you started Samaya AI, you were a researcher at Google Brain. I’m curious about your experience there, and also what it was like to do research in industry versus research in academia.

RAGHU: Oh my goodness, yeah. I started collaborating and working with folks at Google Brain all the way back in 2015. To put that in perspective for where the AI space is right now, this was before OpenAI even existed. There was Google Brain, Facebook had just started the Facebook AI Research effort with Yann LeCun joining them, and that was basically it.

One exciting piece about being in industry at that time, and maybe a difference from academia, is that even back then you were seeing the formation of these slightly larger teams that were going after specific capabilities, specific infrastructure, or specific foundations for AI. That was maybe a little bit different from the academic approach.

That’s what really pulled me into Google Brain.

I thought, hey, look, there are all these people. They have all of this secret know-how, these intuitions about how to build and train some of these systems.

Now that I mention it, maybe that’s one thing that has actually stayed the same between back then and now. Working with this group of people and getting insights that aren’t easily written down anywhere, but come from lived experience, is something that has stayed the same in the AI field.

Copy LinkWhy better AI will require more humans in the loop

EL KALIOUBY: I actually think people don’t talk about that. It’s almost a bit of an oxymoron that in this age of AI, where we’re thinking about automation and everything is productivity-focused, there is still a lot of intuition in how you train these models. Talk about that a little bit more.

RAGHU: Absolutely. I think people are underestimating how much human perspective and human input are required to really enable our glorious AI future. It’s going to come up in all kinds of different ways.

One thing that’s been top of mind for me recently as I’ve seen some of these AI agents develop is that pretty soon we’re going to have AI agents running through the night. In some sense, nighttime might not be dead time in the same way that it is right now. You can imagine that, as a human, you’re going to do a handoff process with some AI agents that go do things for you and then come back to you the next morning saying, “Hey, here are all the things I found.”

Actually, this future we’re heading toward is not that different from the experience of being a machine learning researcher over the past decade, because as a machine learning researcher, you wouldn’t want to leave those GPUs unused overnight. You’d be setting stuff up for your models to train, and then the next day you’d look at the results you got.

Now that’s just going to be more accessible to everybody. We don’t have to think about the models anymore. People are going to have these agents, and these agents are going to go do things for them overnight. But coming back to your question, for those agents to do well, just like with the machine learning researcher, that handoff has to be really, really good.

You had to go in there and get all those details right, make sure the model was well set to train. Similarly, people are going to have to make sure their agents are well enabled to succeed, and that’s something we’re really underestimating. In a world of AI agents, I don’t think you get that productivity and positive impact unless you also have people guiding them on how to succeed.

EL KALIOUBY: Yeah, I remember those days. You don’t want to run the training iteration during the day, because then you’re just sitting idle, right?

RAGHU: Yes.

EL KALIOUBY: So overnight runs were super smart. But to your point about setting it up for success, you don’t want to go home, be asleep, and then an hour in the code runs into something and just stalls or stops, right?

RAGHU: Yes, exactly. It’s going to be similar for the agents if they have that amount of time. We’re going to see the AIs get better at decision-making, planning, reasoning, and all of those things. But putting that within the right framework, the right context for that decision-making, is still going to be heavily driven by humans.

I think, paradoxically, as the AIs get better, the thing no one’s talking about is that we’ll need more and more humans, more and more people in the loop so that we can make sure they all have the right frameworks for judgment and all of that. There’s going to be an element of that that’s not easily scalable, and there is going to be a heavy human touch in those pieces.

That’s really the future I see us heading toward.

Copy LinkWhy leaving Google Brain made sense when language models matured

EL KALIOUBY: Hold that thought because I want to come back to it. I think it’s a very important one and a very relevant one, and I think it’s top of mind for a lot of people. But you left Google Brain to co-found Samaya AI, and as I think of my own experience leaving research and starting a company, it’s kind of a big mental and emotional shift to move from research to entrepreneurship, to leave big tech and go start from scratch.

What was that transition like for you? And also, what was the impetus for the transition?

RAGHU: I always say the motivation for starting a company is really, really important. I think it has to be quite pure in some ways. You have to leave not caring too much about the exact outcome or where it goes, but driven by this desire to bring something to the world, and hopefully by the positive impact you see in bringing that thing to life.

That was definitely the driving force for me. I’d had 10 years as an AI researcher at that point. I had this beautiful body of work. I’d been fortunate to collaborate with a number of the leading figures in the field, some of whom went on to become some of Samaya’s first angel investors.

It was a very satisfying period, and I think the thing that really drove me was exactly that desire to go zero to one, to bring something to life that didn’t exist yet. What I saw was that, over this past decade of AI research, AI had not really been ready to be used in the real world the way it is today.

People had tried in various ways, but it was either too brittle, or you had to make it very focused on one specific thing, or it worked better within a larger ecosystem of some kind. It wasn’t in a place where it could really stand on its own. Finally, with the emergence of large language models, which we saw early — I think the world saw it in late 2022 when ChatGPT came out and everyone started paying attention, but sitting where I was in Google Brain, you started seeing the pieces come together in 2019 and 2020 or so.

From that point on, I could see the potential of finally having this AI that was general-purpose enough, generalizable enough, that you could put it out into the real world and have it drive impact for people.

EL KALIOUBY: Very cool. What’s behind the name?

RAGHU: Samaya actually means time, or moment in time. I really like the name. It has a history in Sanskrit, which is an ancient language in India, and I’m Indian. Most of all, the goal with starting Samaya — the impact that we wanted to see in the world — is giving people back their time and driving them to that moment in time that is a moment of insight as they go about their day-to-day. In our case, that’s a lot of these investment decision-making use cases.

EL KALIOUBY: In a minute, why ChatGPT won’t cut it when it comes to making big investment decisions. Plus, we go behind the scenes of Samaya and see what it takes to build an AI system that can model real-world messiness.

Copy LinkWhy benchmark wins do not reflect messy financial decision making

EL KALIOUBY: Samaya has garnered a lot of attention, so I’ll name some of your awesome investors. That includes former Google CEO Eric Schmidt, Yann LeCun, and Mark Cuban, who we’ve had on the show. You also announced recently a new investment from NVentures — that’s NVIDIA’s venture arm — as well as Databricks Ventures. So, congratulations. That’s a huge accomplishment. How did that all come about?

RAGHU: The unifying theme between the investors was our clear north star of seeing both these AI capabilities and our vision for how we would take that and really translate it into meaningful value in the investment decision-making space.

How that influenced what we wanted to focus on in terms of our AI work, and the use cases we were able to support, was something that I think was very inspiring for investors. From the inception of the company, we saw both the development of these fundamental AI capabilities, but we also saw the work it would take to take some of those capabilities and translate them into practice.

You see all these benchmarks, but these benchmarks are often quite stylized evaluation settings, and progress on benchmarks does not translate to progress in the real world.

EL KALIOUBY: Can you give us an example? Because I think the average person is looking at how AI is passing all these benchmarks and not realizing that oftentimes it does not translate to the messy real world.

RAGHU: Let me give you an example from Samaya. We work on investment decision-making. There are some benchmarks out there to try to test the capabilities of AI in the financial domain. What these benchmarks look like is the following:

There’s a very specific question that has a very specific numerical answer, and you can look at that question and the numerical answer the AI gives you and decide whether it’s right or wrong. That’s the benchmark.

That is not at all representative of the real world. In the real world, somebody is asking some very complicated, ill-formed question. There are many variations of what “right” would look like. Some things are more right than others. It’s not just about a specific number, yes or no.

What we did — and this is a project that we call Criteria Eval — is we actually created an evaluation rubric that you would use to evaluate different types of answers, and then grade based on how many different things they touched in the rubric. That’s one example of the messiness seeping in, and there are many more examples like that.

EL KALIOUBY: Can you give us an example of what this looks like at one of your customers or users?

RAGHU: Take the financial market. It’s a market because people have opposite views of what might happen. Now imagine two users with something like ChatGPT, and two users give ChatGPT a relatively similar prompt.

Then ChatGPT is going to give these two users a relatively similar response. But what we need to do is personalize the AI so much that it’s able to give opposite responses to different people based on the context, frameworks, and views that the people are providing to the AIs.

Copy LinkHow AI agents can support investment theses without replacing judgment

EL KALIOUBY: What’s an example of an instruction? Are you asking the agent to go do research on a specific investment, or are you asking it to actually go execute on it? How agentic — what does agentic mean in this context?

RAGHU: I’d say it’s adjacent to the decision-making. It’s not doing the end execution all by itself just yet, and I think having people in the loop is still very important.

One example is evaluating a thesis. One of our users, an investor, might come in with a particular thesis, a view of the world, of the market.

That thesis might also be tied into specific firmwide context and frameworks, as well as their exact portfolio. They might want the agent to go do research on their thesis, understand the implications for their portfolio, cross-reference that with the broader firm context, and come back to them with key takeaways.

That could include things they should change a position on, or potentially new opportunities they could pursue that are in line with everything they’ve shared.

EL KALIOUBY: The analogy I’m creating in my mind is that, in the same way that vibe coding democratized access to building apps and websites and whatnot, are you, in a way, democratizing access to financial information, databases, or manipulating and visualizing data?

RAGHU: Vibe coding didn’t democratize access to codebases; it democratized access to coding itself. You could say something similar here. I wouldn’t say Samaya democratizes access to information necessarily.

We certainly make it easier. We put it all together within a firm’s context. But I think some of the reasoning and analysis on that information does become more accessible. This has also been a guiding principle for us: we really want to see AI that up-levels humans, that lets us do things that are innovative.

It’s not just a productivity game. You can approach things completely differently, and that’s what we’re seeing a little bit, too. Now that people can put all of these content sources and data in — things that weren’t even possible before — maybe some structured data, along with broader views of the market and very specialized perspectives, then that becomes something innovative. That becomes something you couldn’t do before.

Copy LinkConnecting geopolitics macro trends and company level decisions

EL KALIOUBY: Yeah, I love that point of view. One of the considerations when you’re making investments is also the geopolitics of the world, right?

How do you incorporate this into your models?

RAGHU: That’s actually one place where I think Samaya has been especially valuable to all of our users and clients because, like I said before, when you look at financial services, often you have this choice between zooming out and zooming in. But that’s still not perfect for decision-making.

It’s a mechanism that we as humans have put together because we need some way to navigate what we have to do. Samaya really lets you connect that thread between the zoomed-out and zoomed-in views. That’s so important, especially in a situation where there’s so much geopolitical change.

There’s also this wave of AI and how disruptive it’s being. There are these broad themes going on, and they’re affecting everything at the very zoomed-in level, too. You need to be able to understand the connection between both of those.

One thing we’ve been working on internally — it’s a project we have called Causal World Models, and we posted some things about it in a research preview — is getting the AI to do some of this cause-and-effect reasoning, taking people from macro to micro, tracing through millions of sources, but doing that in a way that’s very sensitive to the cause-and-effect pieces, fully attributable, and giving people that ability to connect the macro and the micro.

EL KALIOUBY: Can you give an example? Say I’m a hedge fund manager and I’m trying to make an investment decision. What kind of questions can I ask Samaya’s AI, and how does that causal world model come into play?

RAGHU: There are so many open questions right now on AI and software, for example. You might ask this broad question: We have this huge disruptive wave of AI. We have the SaaS apocalypse on the other side. Help me work through and understand where the places are that we see huge disruption and huge change.

Where are places where things are just much more reactionary in terms of the market? Where are maybe new categories that are coming up that we should pay attention to?

That’s an incredibly hard question because what you’re doing is taking this high-level theme on AI and translating it down to the very specific entities being impacted, then categorizing those in different ways and doing that cause-and-effect reasoning.

You need to be able to show your work when you come back. You need to be able to say, “This is the prediction, but this is also why, and here’s my chain of reasoning, my cause-and-effect reasoning, that connects me all the way back up to that macro theme.”

Copy LinkWhy specialized AI systems beat one size fits all models in finance

EL KALIOUBY: It sounds like your approach is very focused on these smaller models as opposed to a general-purpose model that can do everything, and you’ve built these models specifically for the financial services sector. Can you talk about that decision? Why not use a general-purpose model?

RAGHU: Absolutely. First, we’re all about systems over models. If you look at every AI advance that’s happened out there in the real world, it’s always systems of some kind. I always like to bring up self-driving cars as a meaningful reference point.

Computer vision predates large language models, and self-driving started even earlier. Now we have them out in the real world. I actually took a Waymo in today. But that’s a system — not a single model by itself. It’s all these components. So, always systems over models.

As part of that system, when you’re trying to go from the library to the office — when you’re trying to embed in people’s day-to-day work — there are all these other AI components that need to be built out to do that translation very effectively.

That’s where some of these smaller language models that we’ve trained specifically for the domain, and for various different types of tasks, come in. For certain types of accuracy and precision issues that we see larger models trip up on again and again, and that are very domain-specific, it’s very powerful to have all of that.

EL KALIOUBY: I love the approach of thinking about it as a system of technologies and components that all need to work together effectively and productively.

EL KALIOUBY: If you’re listening to all of this and thinking, wow, Samaya’s AI system does a lot of the same things as a junior financial analyst, you’re not wrong. But after a break, Maithra shares why she thinks AI is actually good for job growth.

[AD BREAK]

Copy LinkHow AI could reshape jobs while creating new roles for human oversight

EL KALIOUBY: I want to zoom out a bit and talk about what this means for jobs. I’ll make it quite personal in two ways. One is we’ve been using this chief of staff AI agent, and we have a couple of junior analysts on our team. I’m already noticing how this is causing us to rethink what our junior team members ought to be focused on, right?

The second personal anecdote is that my son is a junior in high school, and he is quite interested in, or exploring, potentially studying economics/finance in college. But I don’t know — will these jobs exist anymore? Will it look very different? I’m curious about your thoughts on how this is changing the jobs landscape.

RAGHU: I’ll share some maybe bad news and good news, in that order. I think there’s more good news than people talk about, so I’d like to spend a bit more time on that.

A bit of bad news is that it’s true, this is going to be disruptive. It is going to have us rethink a lot. Some roles are going to be transformed quite substantially. We’re going to have to rethink what we do. Some tasks may end up fully automated. So there is a wave of disruption that’s coming toward us, and we have to acknowledge that.

But the good news, again, that I see fewer people thinking about is this: I deeply believe that, especially as these agents get more capable, you are going to have to have more humans working with them just to deal with how much output they’re going to be producing. The AI doesn’t need to sleep. The AI is constantly going to be working. That’s great. But then you need humans in the loop to put in perspectives, frameworks, the broader universe in which whatever decision-making is happening, all of that.

With long-horizon agents running overnight, who is going to be shepherding them? There’s a whole set of jobs around these high-agency shepherds, maybe, that we don’t even see yet. And there’s going to be a lot of that to come.

Secondly, I’ll say that sometimes when you have this wave of technological disruption, you see the places where it’s clear productivity gains or clear cost optimization much earlier than you see the things that are deeply innovative and are going to up-level all of us — make things possible that we wouldn’t have even done before.

EL KALIOUBY: Do you have favorite examples?

RAGHU: I’ll take one from the investment landscape. Financial services is a huge industry, of course. Within that industry, Samaya is laser-focused on investment decision-making, which we think is at the heart of the industry. The company’s mission is taking people from information to conviction.

We’re super focused on that decision-making piece, partially because we think that’s a place where you can drive innovation, where you can really enable things that weren’t possible before. A super simple example is publics and privates.

There’s a whole ecosystem around public companies, and there’s an ecosystem around private companies. Those ecosystems have changed and transformed in various ways. They now influence each other way more. Often, different strategies were used for each of them.

Is there a way to bring some of these together and, in that process, create new opportunities for people to invest in mixes of these? That’s a simple one, but it would be very meaningful.

It would be meaningful for more people to have access on the private side, for professional investors to gain confidence that they have a really good understanding of both of these areas and the factors that influence them, and for new types of investment mechanisms and products to suddenly become available to people that could touch both of these.

Copy LinkWhen personal agents become real and what they still cannot do

EL KALIOUBY: Very cool. You also talk about personal agents. When do you think those will become mainstream?

RAGHU: I think this year. I think this year is going to be the year of real personal agents — maybe with some things happening on the model side, some stuff where you’re seeing more involved translation, especially on some of these enterprise use cases. But if I trace back the history of agents, to put it in context a little bit:

The term agent became very popular in 2024 or so. I don’t think we had real agents until about late last year. From 2024 through 2025, what we had were more workflows, not agents. The difference between a workflow and an agent is that a workflow is very hard-coded.

It might be multiple steps, but it’s a very specific set of steps. There isn’t actually any agency happening in that execution process. Late last year — I think Claude Code was an early beginning of this — and since then, the capabilities of these systems have developed.

Now we have a little bit more real agency. You give the AI the relevant information and inputs, and it is able to go away and make decisions as it goes through that execution loop.

EL KALIOUBY: It knows the desired end result, and it can backtrack what it needs to do, right?

RAGHU: Exactly. It’s real agency for the first time.

EL KALIOUBY: If you could wave a magic wand and have a personal agent that can do anything for you, what would you want it to do?

RAGHU: My goodness. Probably clone myself and be me in all of these places — come in as me, assess the situation the way that I would, and then be able to take the appropriate actions or produce the appropriate outputs.

EL KALIOUBY: Have you experimented with creating a digital twin of yourself? I’ve tried doing a video digital twin, and also feeding a model with all of my blog posts and my book and my interviews. It’s not there yet at all. I would not send my digital twin to speak on my behalf anywhere.

I don’t trust it at all. But I’m curious if you have experimented with some of these.

RAGHU: I have done some experimentation. Probably closer to what you’ve done — I’ve tried to connect it to things I’ve written. I like audio a lot, so I like speaking to it, because I think if you can really speak to it instead of typing, you can be much more descriptive.

You can give it more information about where you’re coming from. That’s been really powerful. But I’ve noticed two things. If the thing I’m trying to do is scoped enough and it has very good inputs — if I tell it, “This is something I’m trying to put together. These are a few different draft versions. Here’s what I think is missing. Help me put this together” — if I give very specific inputs, it can do a reasonable job.

But when I don’t give it precise inputs and I try to get it to produce an output, it produces something that is not like me, and I need to correct it in a bunch of ways.

What I’ve learned is that for very scoped things, it’s fine. But for something more open-ended, you can’t hill-climb from it.

EL KALIOUBY: OK, so I’d like to do a quick rapid-fire. I’m going to throw out an assumption about AI, and you tell me if it’s myth or reality. Humans in the loop slow down progress.

RAGHU: Myth.

EL KALIOUBY: OK.

RAGHU: The feedback is really helpful. How is the AI going to get better if humans can’t give any feedback, steer it, or correct it in any way?

EL KALIOUBY: That’s great. I think I agree with that one. Large models always win.

RAGHU: Myth. I think there are some things they’re very good at, so I don’t want to take that away. But in other places that we’ve lived and seen as we’ve built out some of these agents, I think the smaller, focused models can really, really drive value.

EL KALIOUBY: Chat interfaces are dated.

RAGHU: Reality. I think that’s coming.

EL KALIOUBY: Interesting. Cool. And then for our younger audience: economics/finance majors are obsolete.

RAGHU: Myth. There’s a lot of human judgment and human taste that will still be a key piece of that field in the future.

EL KALIOUBY: That’s awesome. Last question: What does it mean to thrive in the age of AI?

RAGHU: Energy and experimentation. We’re in a time of disruptive change. That’s true. So let’s face that with energy. Let’s also face it with experimentation. Go out there, see what’s possible, see what we can create, and see what’s out there.

I think that will 100 percent leave behind some of those old inhibitions. Embrace the new, embrace it with energy, don’t be afraid to experiment, and I think it will lead you to exciting things.

EL KALIOUBY: I love that. That is awesome. What a great way to end our conversation. Maithra, thank you for joining us on the show. This was great.

RAGHU: Thank you so much, Rana. This was fantastic.

EL KALIOUBY: It was great talking to Maithra about how to build an AI system that can make sense of our messy world.

An AI agent for the financial industry is incredibly complex. Think about all of the contextual information that Samaya’s agents need to digest and then pull together to make a recommendation. But at the end of the day, there are judgment calls that need to be made, and this is where humans will stay in the loop.

I’ve said it before on the show: AI is disrupting the labor market, but it won’t destroy it. Yes, some jobs will become obsolete because they will be automated, but there will be a whole new class of jobs to manage AI output. We’re already seeing those kinds of jobs grow.

That’s it for this week — thank you for joining us. We’ll be back in your feeds with a new episode next week.