GenAI explosion

Today, Barlow applies his techie background to harnessing generative artificial intelligence (GenAI), which has taken AI’s potential for the business to a new level.

Bloomberg started using machine learning about 16 years ago, according to Barlow, when it applied the technology to process news articles to give customers targeted information. Since then, Bloomberg has been applying AI to its software, where it believes it can deliver a better experience to users.

Today, GenAI has “upped the capabilities”, says Barlow. “Things that we felt were not feasible before are now.”

• Read more about Bloomberg’s AI strategy in this interview with Amanda Stent, Bloomberg’s head of AI strategy and research.

It was about three years ago when a team of researchers and the chief technology officer (CTO) at Bloomberg published a paper after building their own large language model (LLM) “to gain deep expertise in the technology”.

“That allowed us to think about how we’re going to apply [GenAI] to our product line,” says Barlow.

Its first suite of GenAI products was focused on “single documents”. For example, in January 2024, the company released its first external-facing GenAI product, an earnings call transcript summarisation tool. Barlow says there later came a number of follow-ups, all focused on single documents.

GenAI is also summarising news articles. “For anything longer than a certain number of words, we give [users] a quick summary, and if they are really interested, they can read the full story,” he says.

Bloomberg has simultaneously worked on ways to apply GenAI to multiple documents.

“We know that customers have a very low tolerance for errors, so we want to make sure we can give them very accurate summarisation across multiple documents, and we are exploring different ways of doing this, like dense vector search systems,” says Barlow.

These systems can use statistical models to understand related words. For example, if you search for AI, it will also understand that it needs to look for references to machine learning.

That was the state of the art in early 2024, when people were looking at these databases, Barlow tells Computer Weekly.

“I went to Virginia Tech expecting to be a mechanical engineer. My roommate was a computer scientist. I helped him and others with their Fortran programming assignments and realised that programming was where my aptitude and interest were. Instead of studying mechanical engineering, I got a computer engineering degree”

Wayne Barlow, Bloomberg

He says Bloomberg worked closely with customers on projects around multi-document capabilities. “With the multi-document development, we worked in very close partnership with our customers. We went out and got some feedback, and we made sure that we had the performance, but what happened is this really big change in agentic AI.”

One project Barlow’s team is working on, currently in its late Beta phase, will allow customers to ask GenAI-specific questions, rather than a simple semantic search system or an index system, with synonyms.

“It is an agentic system that will go out and try to find all the documents across all of these different collections that will best answer the question and then give a summary,” he says.

Pace of change not seen since mobile

Barlow can’t remember a time when the pace of change was as fast in IT, although he says the arrival of the web, smartphones and everything that came with them were similarly significant explosions.

“I’m not sure if anything was quite at the pace of AI today,” he says. “What I think has made this different is there is a lot of competition between the providers. There’s enough competition, but not too much, so you’re seeing a lot of providers of these LLMs, and they’ll leapfrog one another’s performance.”

Barlow says there are not just expectations around the software and the LLMs themselves, but the hardware that drives this capability, with advances in chips and graphics.

He adds that everyone is betting on the technology becoming cheaper. “When I say cheaper, I don’t necessarily mean just dollar amounts – I also mean power usage.”

Barlow’s career might have turned out differently if he’d had a different university roommate. “I went to Virginia Tech expecting to be a mechanical engineer. My roommate was a computer scientist. I helped him and others with their Fortran programming assignments and realised that programming was where my aptitude and interest were. Instead of studying mechanical engineering, I got a computer engineering degree during that time,” Barlow tells Computer Weekly.

Barlow also took an internship at the Naval Research Laboratory in Washington DC while at Virginia Tech, where he worked on infrared missile simulation. He says it “sounds more fun than it actually was”.