Revolution and evolution: An interview with Andrei Serjantov

Can you tell us more about yourself, your background, and your role at BNP Paribas?

I have a background in tech. I've got a computer science degree and PhD, and I've had many roles in the QUANT sector. I have previously worked with BNP Paribas as the Head of Flow Credit Quantitative Research and then Head of Electronic Credit Trading. In 2022, I became CFO of a crypto startup which was a rather different affair, but it made me very familiar with how startups function; the fast-paced environment, the people, etc. It was about one and a half year ago that I came back to BNP Paribas to take this role, which is a broad role across global markets. I actually joined only a few months after the famous November release of ChatGPT.

Can you tell us more about your collaboration with Mistral AI? Why did you choose to partner with them, what made them stand out?

Around about September last year we started engaging with Mistral AI. They were only a few months old then. We spotted them as a very, very promising company that we really wanted to have a look at. It was their 7B model, a small model that was very performant, that really caught our attention. We’ve recently announced a partnership with them with an engagement across the BNP Paribas Group. It’s a multi-year agreement across the group, and is the product of quite a number of months of discussions.

We are also investors in Mistral AI. You can see a couple of rounds; one in December, one in June. It’s a multifaceted partnership. Needless to say, we can talk about where their models are being used. One key example is the model that they've released in May.

It's again one of those landmark models which is very performant and very quick because it's a small model that fits very nicely into the code assistant framework where our developers can leverage it to help them check code, write unit tests, write documentation, all things that developers do. This model runs on-prem within BNP Paribas, so there's no code going out, no issue with code leakage. It all stays in house.

It's top of the range performance. Quick speed matters a lot when you're coding.

By choosing Mistral AI, was there the desire from BNP Paribas to develop the French AI ecosystem?

It does help that they are in Paris from an engagement perspective, but it certainly was not the deciding factor. The deciding factor was that their strategy fits in very nicely with our needs. We're a regulated institution, so safety is of paramount importance. Being able to run models on brand is a very, very big selling point for us. It also means that they are smaller models, cheaper to run models of the same quality and more secure.

It seems that BNP Paribas is at the forefront of technological innovation. Have you noticed banks worldwide being inspired by your example?

What I can say is that we have quite a specific way of working.

We think that optimising for agility in a space which is growing and moving very fast is very important. A lot of people ask “What are you doing with Gen AI? Gen AI is here, LLMs are here.” But if you look at it under the hood, there's been a significant amount of progress. The first models were really focused on English and a few languages, and now you see LLMs that are very capable across not just “regular” languages, but also programming languages. For example, use cases like translation are becoming very natural, which helps us as a global bank.

It's an evolving space. If you look at the last six months, the other piece of progress is really in coding benchmarks, right? The first models were ok at it, but in the last six months, it's gone up a lot. If you have an agile setup, what that means you can do is you can take a use case, test a model or a setup, and if it doesn't work, you just discard it very quickly because these models are very capable and testing such a use case is not very hard.

There are also a lot of open source models out there. Yes, "open source" is a buzzword, but it has fuelled this massive innovation and massive progress within LLMs and so again it's the agile setup that helps us fit into this ecosystem very nicely. The challenge of course is that we're a regulated institution. So we need to weigh up this desire to progress fast. We're incredibly careful on risk management as well. Why? As a regulated institution, safety is obviously paramount.

The real question is how.

Firstly, educating and making sure that we have expertise, making sure that our processes are incredibly thorough yet aligned to this technology in place. Those two combined means that we do have agility, safety, multiple guard rails, etc. It’s the risk management culture brought over by the sales and trading side to the LLMs space that enables us to do that.

When did BNP Paribas know that AI what the future?

If I may, I'll give you a personal anecdote. These models have been evolving. If you look at natural language processing, there have been models over the years. I used to ask these models questions such as “When was the last war between the Aztecs and the dinosaurs?” And these models were very happy to talk about it, and they would make up a whole story. And then ChatGPT was released. When asked the same question, it was instantly saying that there was never a war between the Aztecs and the dinosaurs. For me, that was a step change where we saw these models suddenly getting very good.

To me, it’s not a revolution, but an evolution. These models are evolving are are capable of handling more and more things. Initially, they would speak a handful of languages and not very well. Nowadays you have models capable of speaking many languages. That illustrates progress. It's not about having one big model for everything; it's about understanding which use cases these technologies are capable of and then progressing on those use cases.

Is there anyone who inspires you?

I draw a lot of motivation and enthusiasm from the people I work with and talk with on a regular basis. I have a set of people back from my PhD days who have built incredible cutting-edge technologies. Similarly, BNP Paribas is full of amazing French engineers who are absolutely amazing at what they do. By adding Mistral AI to that mix, you get a great ecosystem.

I am inspired by this, rather than a visionary figure of some kind.

Andrei Serjantov is the Head of Digital Global Markets at BNP Paribas.