OpenAI’s ChatGPT’s extraordinary capabilities necessitate the use of massive language models. Text samples numbering in the billions, if not trillions, are used to teach these models. The goal of ChatGPT is to have such a deep understanding of language that it can quickly and accurately guess the next word in a conversation. This requires a substantial investment in time for education, computing power, and development expertise.
Yet, perhaps these models will eventually become more narrowly targeted than the “boil the ocean” approach taken by OpenAI and others. What if there was a model specifically educated to comprehend the lingo, methodology, and processes of a given industry or even a given company? If people were required to answer questions using only the words and phrases from a smaller pool, we could obtain fewer wholly fabricated responses.
Each company’s data may be its most valuable asset in an AI-driven future. A company’s language model is comprised of its customers’ data and the entire body of material produced by the company, and its lexicon is unique to the type of business it is (e.g., insurance, healthcare, automotive, and legal). That might not be a huge model in the sense of a truly massive language model, but it would be the perfect model for your purposes.
To prepare the corporate dataset in a form that can be consumed by these smaller, large language models, a series of tools to gather, combine, and continuously update it will be necessary (sLLMs).
It could be difficult to construct such models. They will likely use existing LLMs from places like open source or a commercial corporation, then refine them using data specific to the industry or company in a safer setting.
Startups have a lot to gain from this, and there are already many established businesses that have jumped on the bandwagon.