Access control, model governance
There are several measures organizations can take to ensure they curtail the risks involved in accessing and deploying foundation models while still reaping their advantages. These are some of most readily available ones:
♦ Trusted vendors: One way to avoid the data privacy shortcomings associated with publicly accessible foundation models is to do so within the environment provided by a trusted vendor, which serves as a buffer between end users and those models. For example, users might access these models through a vendor’s cloud. “Instead of sending information to your privately provided ChatGPT or GPT-4, we are going to bring GPT-4 level capabilities into the confines of our secure, FedRAMP, SOC 2-compliant environment in the future,” Galal noted. Some vendors currently offer these capabilities.
♦ Open source models: There are entire communities of open source resources, advocates, and foundation model builders—complete with expedients such as GPUs on demand—that are swiftly democratizing the deployment of these models. This approach enables organizations to “construct your own LLM and run it privately,” Martin mentioned. “You use one of the open source LLMs and get your own hardware, so there’s no one in the loop recording your data. But, you need the expertise to do this.”
♦ Model management: Companies can also avail themselves of model governance and management tools. Accord- ing to SAS’s Browning, competitive offerings “provide transparent details about data and models in a scorecard format, including governance and business process auditing information, model limitations, indicators for private and sensitive variables, and a clean readout of variable importance for machine learning predictions.”
An extraordinary moment
Foundation models are as proficient at understanding natural language as they are at generating text and images. The most pervasive applications involve search, conversational interfaces with enterprise systems, and automation of everything from intricate business processes to query generation. “This is an extraordinary moment,” Martin summarized. “It’ s a little frightening because it’ s hard to know where the end of this stuff is.” Generative AI and foundation models will have a profound effect on digital transformation and require a reimagination of the user experience.