BloombergGPT Highlights the Power of Smaller AI Models in Finance
How BloombergGPT Champions Smaller, Focused AI Models
The Gist
Bloomberg launched BloombergGPT, a 50-billion parameter, large-language model designed for financial applications.
The model is built on a knowledge base Bloomberg has gathered over the past 15 years and is provided to customers through its Terminal product.
Researchers from Bloomberg and Johns Hopkins argue that smaller models work better for domain-specific applications like finance.
BloombergGPT incorporates chatbot functionality from ChatGPT and provides greater accuracy than comparable models with more parameters.
Smaller models improve result accuracy, train faster, and demand fewer computing resources compared to one-size-fits-all models like ChatGPT.
Bloomberg dedicated nearly 1.3 million hours to training BloombergGPT on Nvidia's A100 GPUs in Amazon's AWS cloud.
More Detail
Bloomberg is refining artificial intelligence applications in smarter ways, without the ethical or security concerns that trouble ChatGPT.
The company recently launched BloombergGPT, a 50-billion parameter, large-language model designed for financial applications. Bloomberg built the model on a knowledge base it has gathered over the past 15 years and offers it to customers through its Terminal product.
While BloombergGPT's scope is smaller than GPT-3, which relies on 175-billion parameters (GPT-4 is 170-trillion), researchers from Bloomberg and Johns Hopkins argue in an academic paper that smaller models work better for domain-specific applications like finance.
Some IT executives also advocate for smaller models with a few billion parameters, particularly for scientific applications. Smaller models improve result accuracy, train faster than one-size-fits-all models like GPT-3, and demand fewer computing resources.
Bloomberg dedicated nearly 1.3 million hours to training BloombergGPT on Nvidia's A100 GPUs in Amazon's AWS cloud.
54% of Bloomberg’s data set came from their internal database, equivalent to 363 billion internal documents dating back to 2007. The remaining 345 billion documents came from public sources such as press releases, Bloomberg news articles, public filings, and even Wikipedia.
Bloomberg will not release its BloombergGPT model for evaluation, following OpenAI's example. OpenAI open-sourced GPT-3 but charges for access to the closed-source GPT-4, announced last month. Bloomberg's business model depends on proprietary algorithms that provide intelligence to traders and analysts, and releasing BloombergGPT could expose core assets.
Bloomberg plans to enhance the model by incorporating more data into the system and addressing any issues that arise.