The release of ‘9fin AI’, GPT for LevFin
- Sean Löfgren
9fin AI enables users to find and process the information they need quicker. Speed is essential in financial markets and debt is no exception.”
9fin has been leveraging AI since the company was founded back in 2016, but we’ve recently stepped-up our game.
‘9fin AI’ allows subscribers to ask questions about companies within debt capital markets and receive high quality answers in real time.
Listen to our latest episode of Cloud 9fin, where we sat down with 9fin’s Sean Löfgren and Hannes Kindbom to talk us through the exciting new project, their thoughts on future opportunities and their stance on potential risks linked to generative AI.
Listen on Apple Podcasts
Listen on Spotify
Sammy: Hello and welcome to Cloud 9fin, a podcast on all things leveraged finance. We follow corporate debt, from issuance to redemption, credits from performing to distress, and everything in between. I'm Sammy Cole, your host in London, and this week I'm welcoming head of data science Sean Löfgren. Thank you for joining me Sean.
Sean: Thanks for having me on the podcast.
Sammy: And also data scientist Hannes Kindbom.
Hannes: Thank you.
Sammy: Thank you so much for joining me both. And the reason I've got you both on is because of the release of 9fin AI a few weeks ago. And it seems like the perfect time to have a chat about how we're leveraging AI at 9fin, and also to take a sneak peek into what the future holds. So in actual fact, 9fin has been leveraging AI since the company was founded back in 2016. But we've recently stepped up our game. So to start, Sean, could you please give me a brief overview of how the recently released 9fin AI works?
Sean: Yeah, sure. 9fin AI allows our clients to ask questions about companies within the 9fin universe, and that means receiving high quality, fully referenced answers in real time. For example, how could EG Group deleverage or, maybe, what does Pure Gym do? This ChatGPT-style interface is fully integrated into our existing global search functionality.
Sammy: Amazing. Thanks for the intro. So 9fin AI is still in beta. What do you see being integrated down the road, Hannes?
Hannes: Good question. I think one thing we definitely want to do is to give it access to more 9fin data. Because we started with a subset of our companies and data to ensure precision, since we know that's extremely important for us and our clients and hallucination is a known issue with LLMs [large language models]. And so we need to be careful there. And we'd rather say sorry, I can't answer that right now than saying something incorrect that could make the user lose a lot of time, or money. And secondly, we also want to make it smarter so that it can answer more complex questions such as comparing different companies' performance, etc. And part of this is giving it access to more data. But we also want to improve the way we retrieve content. And I can talk a bit more about retrieval. Because I think that's such an essential component of these systems, since it allows LLMs to be aware of more recent information not included in their training data. And yeah, I can probably talk about this for hours. But to keep it short, I see a lot of interesting approaches to it out there. And I believe there is not a one size fits all at this point. And so we're continuously experimenting with various approaches to see what works best for our use case. And some examples of fast evolving projects that you can build on or just use as inspirations come from the great work at LangChain and LlamaIndex, for example.
Sean: Yeah, and if I can jump in here, I mean, two extra points, better models, I guess we want to be at the forefront here. So I mean, they're releasing new foundational models every month. You got Llama 2, you've got the 180 billion parameter Falcon model, and more recently cloud 2.1 and GPT-4 turbo from OpenAI on their dev day recently. And last but not least, this is not just a data science problem. We're heavily invested in UX and overall system design and engineering. And we know all our users have high expectations here, especially with ChatGPT setting the bar. We want our users to have a very seamless interaction, make it a bit addictive to a degree.
Sammy: Interesting. It sounds like there are so many different things to think about. And also exciting things to come as well. Did you find any roadblocks when you were working on the 9fin AI project in particular?
Sean: There's like two main roadblocks that we've encountered so far. And the first thing is the fact that this space is moving so fast. So keeping up with that space is our biggest problem at the moment. But also we have so much data in varying formats, so much unstructured data, so much structured data, long documents, we've got charts, plots, tables, and everything in between. And you know, pre-processing all of this in a way that we can search and retrieve the most you know, we want to give you the best we want to give you a correct answer. We can dive into that later. But our main challenge is being able to provide you with all this data so you can search. So it's a challenge both in terms of like data engineering, you know, building all these pipelines, there's a lot of infrastructure, and there's a lot of search optimisation that we have to work on here. So those were our main roadblocks.
Sammy: Yeah, that makes sense. I mean, even when we're writing QuickTakes, we've got a lot of information from long super long offering memorandums. And then also, for example, in sustainability reports, there are all these charts and graphics as well. So I can see why that would be a huge challenge. And Hannes, what have you learned from working on this project?
Hannes: Yeah, a few things. I think, first of all, close collaboration in cross functional teams with designers, domain experts, products, people, engineers and scientists is very powerful when it comes to shipping research and engineering heavy user facing solutions. Because I think you need knowledge from each area. So having all of those competencies on the same team is very effective. And also prototyping has been very useful when iterating on new approaches. And especially if you can use or test it easily. And it's very useful when you have little, or no data available. And it avoids spending time on the wrong things.
Sammy: Definitely. I've really enjoyed testing out the product myself. So Sean, what's the team currently working on at the moment?
Sean: Oh, yes, what are we working on currently? I mean, we're working on improving 9fin AI as we believe it can significantly improve discoverability on the platform and enable our users to find and process information quicker. Speed is essential in the financial markets, you know, faster speed, they know more leverage and that is no exception here. All this while also maintaining the current models in production, we have the infrastructure and live monitoring in place with alerts so that we get notified about any anomalies that happen that need our attention.
Sammy: And what's the feedback been like so far on 9fin AI?
Hannes: I think it's definitely created a lot of excitement in the company and among clients. And we've had record high weekly active users, which is super fun. And that's even though it's still a new feature and clients are coming back to us saying it's already saving them lots of time and it can already answer pretty complex questions, which is really rewarding for us to hear. But we still see there is so much more potential in this feature. So we're constantly, you know, considering feedback, we encourage more feedback on how to improve it and what kind of questions people want supported in the future, etc.
Sammy: We've got a very exciting new product to share with you: 9fin's CLO structured credit offering with exclusive up to date reporting, world leading expert analysis and a modern data platform. Our new product vertical will enable you to make better decisions faster, simplify your workflows and ultimately win business. Head to 9fin.com/insights to try it out for free. Now let's get back to this week's episode.
Back to the podcast
Sammy: And of course, 9fin AI isn't the only way that 9fin been leveraging AI in our processes. Sean, could you give us an overview of how we've been using AI over the last few years?
Sean: Sure thing. Since the early days of 9fin and we've been using state of the art machine learning to provide our clients data with speed and precision. And this goes from like computer vision tasks all the way to natural language processing tasks and everything in between. Like I mentioned before, we have so much unstructured data. So we've been tackling problems around detection and classification to extraction and interpretation problems. And we have plenty more still to tackle here. And thanks to generative AI the last two years we've been able to supercharge our work in the presentational space, pretty much putting a natural language interface on top of all of our products.
Sammy: And so now on to your general thoughts on generative AI. Hannes, is there anything that worries you in particular or any developments that are particularly exciting to you and the rest of the team?
Hannes: Yeah, so I think the capabilities of these models become increasingly general, which is super exciting. I think the technology has also been much more democratised with ChatGPT. Because lots of people know how to use the chat interface, so it becomes more accessible. And this has fueled lots of ideas and creativity in the space. And I think these models have a very high value to effort ratio. It doesn't necessarily mean that the absolute value and performance is high in all use cases, but you can get something pretty useful with quite low effort. And I think these models also facilitate applications that were previously very complex or time consuming to do. And prototyping especially is now much faster with quite powerful few or zero shot learning. Chat GPT was originally text only, but I think we'll see models become more multimodal, like for example, OpenAI's clip or the more recent GPT for vision that can take both text and images as input. I'm particularly interested in seeing the progress on solving the outdated knowledge and context window limit problems of the LLMs. I see that as one of the main bottlenecks right now and lots of investments are going into solving that problem.
Sammy Yeah, I mean, that's even something that I've noticed myself when using ChatGPT and other tools. Do you see any other major risks?
Hannes: Yeah, I'm glad you bring up the risks because I think while generative AI can give humanity huge benefits, I think it also comes with serious risks we should consider. And I sometimes worry about these risks being ignored in favour of people trying to outcompete each other with the fear of missing out, or lagging behind. And I definitely think authorities, nonprofits and governments will and should set the rules and enforce AI safety first. And anyone with power really has a responsibility here in this space. And to concretise, I can give some examples of what I mean by AI risks. First of all, I think these models are very hard to interpret.
So you don't really know necessarily how it came to a certain conclusion. And there's a risk for discrimination. Maybe it made a certain decision based on gender or race or other factors. It's also now easier than before to scale up fraudulent activities and use the models for misinformation or terror activities. And I also see a risk like a big tech monopoly here. And there's also data privacy issues with sending data over APIs. And yeah, especially this concentration of power to a few companies behind these models. So I think we need to think about how we make sure it benefits everyone. And lastly, giving instructions for complex tasks is not easy. And as these models become increasingly powerful, it can have devastating consequences. Just one classic example is if you ask the model to solve climate change, and this model then decides to eliminate all the humans, because humans are probably the main reason for climate change. We need to make sure that the AI understands our goals that we set, and that it adopts the goal and also keeps the goal over time. And these are very hard problems to figure out. I also want to mention that I think expectations on AI right now are sky high. And I think far from everything can be solved with AI yet. And we should make space for learning and understanding the technology. And sometimes a human is actually more appropriate, or a much simpler technical solution. For example, if I'm reading, let's say I'm reading an earnings call transcript, I probably don't want to have to ask an AI to click play to hear the recording. Maybe I just want a simple button.
Sammy: Yeah, that's so true. I mean, we have to think about how AI can be used in conjunction with things that already exist and the wider picture as well for sure. And Sean, what kind of future do you think AI will have at 9fin?
Sean: Yeah, I mean, we will continue to invest heavily in AI and technology just purely to make our client lives easier. And the feedback we get from them will influence the direction our product takes. Not every generative AI use case is even visible for our clients. So I mean, we already have so many ML power systems running behind the scenes.
Hannes: Yeah and to add to that, I think it's always hard to give time estimates on these things. But I think generative AI is still just a tool for increased efficiency like Python or Excel are. And I believe we will have a human in the loop for a while. Just look at, for example, self-driving cars or even in factories where we still have humans supervising the robots.
Sammy: I love that example. Thank you so much for joining me today. It's been super, super interesting.
Sean: Thank you, Sammy.
Hannes: Yeah, thank you so much.
Sammy: Okay, well, that's all we've got time for this week. Thanks for tuning in. And please let us know if you have any feedback. We love hearing from our listeners and you can reach us anytime by emailing email@example.com. Check in next week to hear the latest on US markets. And we'll be back the week after that. See you then.
If you enjoyed the episode and want to find out more about 9fin’s AI products, reach out to firstname.lastname@example.org and speak to a member of the team now!