Bloomberg AI strategy head: ‘All 350 of our AI experts are hammers looking for nails’

‘We only use AI if necessary to address client needs’

Dr. Amanda Stent, Head of AI Strategy & Research in the Office of the CTO at Bloomberg, explains Bloomberg’s minimalist approach to AI

Bloomberg is global powerhouse, feeding up to the minute financial news and data onto the screens of financial professionals around the world. Bloomberg also boasts a news division and channel offering reporting on global financial and business trends, markets, and industries.

Dr. Amanda Stent is Head of AI Strategy & Research in the Office of the CTO at Bloomberg, and an expert on the enterprise applications of AI and machine learning, having worked in research roles at companies like Yahoo, AT & T Labs and in academia. AI has been used for years at Bloomberg, the focus only recently turning more to generative AI application.

“We have more than 350 AI experts at the company,” says Stent. “I would call all those people hammers looking for nails. It's a question of finding the right nails and the right nails are our client’s real problems.”

Bloomberg seems to be avoiding the path that many other organisations seem to be headed down, which is being desperate to use GenAI for any reason at all.

“I would say that our approach is to figure out what our clients’ needs are, and then how we can address those needs, using AI if necessary,” says Stent. “AI introduces complexities that traditional algorithmic methods do not introduce, such as a lack of determinism, and in the case of generative AI, the hallucination problem.”

Absence of determinism relates to the fact that if you ask your favourite AI assistant the same question twice it may well give you two different answers. What GenAI is very good at is the discovery, analysis and synthesising of vast quantities and types of data, and this is mainly how Bloomberg has deployed it.

“News theme summarisation went live in 2021, “says Stent. “We cluster news stories into themes evolving events, and then we construct a little snippet or a headline for each cluster, theme or event. That was very early. Earlier this year, we released earnings call summaries.

“Earnings calls can be long, and the piece of information that you might be interested in could take hours to get to. We have a product that does thematic summaries. Our subject matter experts identified a set of themes, and we construct bullet point summaries on each theme.

Stent is emphatic about the importance of the subject matter experts – the human in the loop.

“As a technologist, I could construct a summary of an earnings call but the one that I would produce would be fairly generic. With the involvement of our subject matter experts, our product teams, our data teams and our sales force, we can create products that are really targeted at financial services, so the themes, the clusters, the topics that matter for our clients to make a product that's really differentiated.

“We have started rolling out AI-Powered News Summaries [the service was formally announced in January] so that helps people who don’t have time to read the whole news story to get the top two or three points. As with all our generative AI products, there is transparency to the source. We always want to make sure that the user can go back and see the source and check the facts.”

Checking the facts remains, in Stent’s view, a responsibility of the user of GenAI summaries.

“When we think about things like earnings, transcript, summarisation or news clustering, the speed at which we’re working means we couldn't ask a human to check every point, but we want the user to be able to check, because those are the people who may be making decisions on that information, so they should have the ability and indeed the responsibility to check.”

Sustainable AI

Some people (mainly those seeking investment) claim that GenAI will be able to solve some of humanity’s more pressing challenges. However, right now GenAI, or more accurately the energy and water consumed by the compute activity necessary to train models and answer queries, is adding to the challenge of climate change.

Stent doesn’t pretend that the sustainability of AI models is at the top of the priority list for Bloomberg or its clients, but says it is a concern and argues that the time constraints under which the financial markets operate necessitate efficiency, which is, by definition, the most sustainable option.

“The most efficient models tend to be the ones that you can run on less compute which means they're inherently more sustainable. Our goal is to have the most performant, efficient, robust model that works for a particular problem. That often means running smaller models because they're faster. It may also mean running models sometimes in public cloud which is more sustainable than running it in our own environment.”

Is GenAI Clippy 2.0?

How does Stent think the GenAI marketplace is likely to develop over the next few years? Will FOMO continue to drive the huge inflows of investment into a product where the promised gains always seem to be deliverable at an unspecified point in the future?

“This year in the US, 41% venture capital money is going into generative AI, and that is an insane amount. This technology has to make money. I think one possible outcome is that this turns out to be Clippy 2.0 and all the money goes away. That would be a bad outcome.

“Another set of potential outcomes is that the models are useful, but we end up in some kind of international geopolitical information warfare where the models from some countries are trained using a particular set of data and a particular set of guidelines, and the models from other countries are trained in in different ways to reflect different countries’ notions of truth and social acceptability.

[This interview was conducted before the launch of DeepSeek V3 and its suppression of information on subjects such as the Tiananmen Square Protests which underlined Stent’s point perfectly.]

“A third set of potential outcomes that I worry about is when you give these models access to run their own code, when you give these models access to perform actions on computers, then because AI is nondeterministic, it's trickier to manage the cybersecurity of it. I am quite concerned that we're going to see a dramatic increase in cybersecurity issues.”

Stent has these risks front and centre of Bloomberg AI strategy.

“We're constantly checking the models that we're using for risks. We don't want these models offering financial or investment advice. We don't want certain kinds of bias in the output and we conduct regular pen testing and red teaming and checking that the models are run in secure environments and with appropriate risk governance to mitigate against some of those risks.“

Bloomberg has also endorsed the Open-Source AI definition from the Open-Source AI Initiative (OSI) and Stent remains optimistic about the trajectory of AI believing that at its best, GenAI can augment human intelligence – not replace it.

“Bloomberg has endorsed the Open-Source AI definition because we think model creators should say something about the nature of the data that's in the model, not necessarily release all the data, but tell you what's in the model and some of the design choices that were made when the model was created.

“The models on their own are fundamentally not that powerful. They just predict the next token. But connected to tools, they can be legitimately useful for many people. We used to talk in terms about computers doing things we would consider smart if humans did them. Now we’re talking about machines that can read a million documents at once, and there's no intelligent human who can do that.

“We’re talking about augmenting humans with different types of intelligence to make a combination of both better than either type of intelligence alone.”