The application of these (and countless other) foundation models to the very way humans work, communicate, and use technology is quietly producing an even more profound result. According to Medhat Galal, Appian senior VP of engineering, “Think of the spectrum of human-to-computer interactions. Normally, we use low code to bridge the gap between what users want and what the application can do. LLMs bring that notion, that continuous spectrum, closer to the user, where they get to use their language, as opposed to the [system’s] language, to understand things.”
Foundation models, trained on vast quantities of data, function as general-purpose platforms for any number of AI applications, including generative AI. According to the U.K.’s CMA, an independent, non-ministerial U.K. government department, foundation models “have the potential to transform much of what people and businesses do across the spectrum of human activity, from searching, to learning, to creating, to how we solve problems across health, engineering, design and education, to name just a few domains. In the process, as with any technology breakthrough, they will disrupt existing markets and create new ones (assets.publishing.service.gov. uk/media/64528e622f62220013a6a491/AI_Foundation_Mod- els_-_Initial_review_.pdf).
These are some of the more utilitarian applications that users can commission via natural language:
♦ Knowledge management: Foundation models can create entire data models, taxonomies, knowledge graphs, descriptions of content, and pertinent classifications—which humans can oversee and readily adjust as needed.
♦ Business intelligence: These models can not only search through desired data for ad hoc question-answering, but also generate visualizations, reports, and diagrams to illustrate them.
♦ Intelligent process automation: Foundation model techniques are employed to create digital versions of documents and other content that becomes what Galal termed “web-friendly” and amenable to upstream or downstream processes in IT systems.
The most powerful of these use cases, which also involve application development and digital twins, combine textual and visual capabilities. They can be performed on-the-fly, in real time, and according to human-specified constraints to eliminate incorrect results.
Nonetheless, none of these possibilities negate the fact that generative AI does not solve all digital transformation dilemmas. “There’s a whole lot of things that enterprises require for these things to be sensible, viable, and not a risk,” added Sean Martin, Cambridge Semantics CTO. “There’s a whole list of tick boxes that have to be ticked off. ChatGPT all by itself doesn’t tick those boxes.”
ChatGPT is well on its way to becoming synonymous with LLMs, but that’s inaccurate. It’s a digital agent that uses LLM techniques to understand prompts in natural language and perform various activities. It can search through vast amounts of information (typically electronic) related to a prompt, synthesize the results, issue them in natural language according to user specifications, or generate endless varieties of text on demand. Depending on the prompt, ChatGPT may be able to answer questions and generate written content based on its training data without additional searching.
Nevertheless, ChatGPT is not a search engine. “LLMs are not a knowledge store,” Martin cautioned. “The knowledge that’s in them is a by-product of how they were trained.” LLMs are foundation models trained on vast quantities of text used to understand and generate additional text. Foundation models are advanced machine learning models trained on enormous data quantities that are applicable to multiple tasks; they typically involve transformer architecture.