A Bloomberg computer terminal at the New York Stock Exchange.
Adam Jeffrey | CNBC
Bloomberg LP has developed an AI model using the same underlying technology as OpenAI’s GPT and plans to integrate it into features offered through its terminal software, a company official told CNBC.
According to Bloomberg, Bloomberg GPT, Internal AI Modelcan more accurately answer questions like “Citigroup CEO?”, assess whether a headline is bearish or bullish for investors, and even write a headline based on a short bio.
Large language models trained on terabytes of text data are the hottest field in the tech industry.giant as Microsoft and Google The race is on to incorporate the technology into their products, with AI startups regularly raising funding at valuations in excess of $1 billion.
Bloomberg’s move shows how software developers in many industries outside of Silicon Valley see state-of-the-art artificial intelligence such as GPT as technological advances that allow them to automate tasks that used to require humans.
“Both the capabilities of GPT-3 and the way in which its performance is achieved through language modeling exceeded my expectations,” said Gideon Mann, director of ML products and research at Bloomberg. “So when it came out, we were like, ‘Okay, this is going to change the way we do NLP here.'”
NLP stands for Natural Language Processing and is the part of machine learning that focuses on deriving meaning from words.
The move also suggests that the AI market may not be dominated by giants with vast amounts of general-purpose data.
Building large language models is expensive, requiring access to supercomputers and paying millions of dollars for it, and some have wondered whether OpenAI and the big tech companies will have an insurmountable lead. In this case, they will be the winners and sell the rights to use their AI to others.
But Bloomberg’s GPT doesn’t use OpenAI. The company was able to use off-the-shelf AI methods that were freely available and apply them to its vast proprietary (if niche) data store.
So far, Bloomberg says its GPT has achieved impressive results in tasks such as determining whether headlines are good or bad for a company’s financial prospects, changing a company’s name to a stock ticker, figuring out important names in documents, and even answering basic business questions. Inspiring results such as who is the CEO of a company.
It can also perform some “generative AI” applications, such as suggesting new headlines based on short passages.
An example from the paper:
Enter: “According to Redfin, the U.S. real estate market will lose $2.3 trillion, or 4.9%, in value in the second half of 2022. This is the largest percentage decline since the 2008 housing crisis, when values fell 5.8% over the same period %period”
Output: “House prices have fallen the most in 15 years.”
OpenAI’s GPT is often referred to as the “base” model because it was not designed for a specific task.
Bloomberg’s approach is different. It’s been trained specifically on the reams of financial documents companies have collected over the years to create a model that’s especially fluid when it comes to money and business.
In contrast, OpenAI’s GPT is trained on terabytes of text, the vast majority of which has nothing to do with finance.
About half of the data used to create Bloomberg’s models comes from non-financial sources scraped from the web, including GitHub, YouTube subtitles and Wikipedia.
But Bloomberg also added more than 100 billion words from a proprietary dataset called FinPile, which includes financial data the company has amassed over the past 20 years, including securities filings, press releases, Bloomberg News reports, other publications, and more. Stories coverage and focus on financial pages.
As it turns out, adding specific training materials is enough to improve the accuracy and performance of financial tasks, and Bloomberg plans to integrate its GPT into features and services accessed through the company’s end products, though Bloomberg has no plans to use a ChatGPT-style chatbot.
An early application is translating human language into the specific database language used by Bloomberg’s software.
For example, it converts “Tesla price” to “(get(px_last) for((‘TSLA US Equity’))”.
Another possibility is to have the model do the work of cleaning the data in the background and perform other tasks in the backend of the application.
But Bloomberg is also considering using artificial intelligence to enhance its capabilities to help financial professionals save time and stay on top of the news.
“We’re doing a lot to help clients with the data deluge of news stories, whether it’s through summarization, monitoring, being able to ask questions about those news stories or transcripts. There are a lot of applications out there,” Mann said.