Thomson Reuters, Where Is Your Large Language Model?
April 3, 2023
Note: This essay is the work of a real and still-alive dinobaby. No smart software involved, just a dumb humanoid.
I have to give the lovable Bloomberg a pat on the back. Not only did the company explain its large language model for finance, the end notes to the research paper are fascinating. One cited document has 124 authors. Why am I mentioning the end notes? The essay is 65 pages in length, and the notes consume 25 pages. Even more interesting is that the “research” apparently involved nVidia and everyone’s favorite online bookstore, Amazon and its Web services. No Google. No Microsoft. No Facebook. Just Bloomberg and the tenure-track researcher’s best friend: The end notes.
The article with a big end … note that is presents this title: “BloombergGPT: A Large Language Model for Finance.” I would have titled the document with its chunky equations “A Big Headache for Thomson Reuters,” but I know most people are not “into” the terminal rivalry, the analytics rivalry and the Thomson Reuters’ Fancy Dancing with Palantir Technologies, nor the “friendly” competition in which the two firms have engaged for decades.
Smart software score appears to be: Bloomberg 1, Thomson Reuters, zippo. (Am I incorrect? Of course, but this beefy effort, the mind boggling end notes, and the presence of Johns Hopkins make it clear that Thomson Reuters has some marketing to do. What Microsoft Bing has done to the Google may be exactly what Bloomberg wants to do to Thomson Reuters: Make money on the next big thing and marginalize a competitor. Bloomberg obviously wants more than the ageing terminal business and the fame achieved on free TV’s Bloomberg TV channels.
What is the Bloomberg LLM or large language model? Here’s what the paper asserts. Please, keep in mind that essays stuffed with mathy stuff and researchy data are often non-reproducible. Heck, even the president of Stanford University took short cuts. Plus more than half of the research results my team has tried to reproduce ends up in Nowheresville, which is not far from my home in rural Kentucky:
we present BloombergGPT, a 50 billion parameter language model that is trained on a wide range of financial data. We construct a 363 billion token dataset based on Bloomberg’s extensive data sources, perhaps the largest domain-specific dataset yet, augmented with 345 billion tokens from general purpose datasets. We validate BloombergGPT on standard LLM benchmarks, open financial benchmarks, and a suite of internal benchmarks that most accurately reflect our intended usage. Our mixed dataset training leads to a model that outperforms existing models on financial tasks by significant margins without sacrificing performance on general LLM benchmarks.
My interpretations of this quotation is:
- Lots of data
- Big model
- Informed financial decisions.
“Informed financial decisions” means to me that a crazed broker will give this Bloomberg thing a whirl in the hope of getting a huge bonus, a corner office which is never visited, and fame at the New York Athletic Club.
Will this happen? Who knows.
What I do know is that Thomson Reuters’ executives in London, New York, and Toronto are doing some humanoid-centric deep thinking about Bloomberg. And that may be what Bloomberg really wants because Bloomberg may be ahead. Imagine that Bloomberg ahead of the “trust” outfit.
Stephen E Arnold, April 3, 2023
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
