July 15thMicrosoft CEO Satya Nadella The blog was published earlier today, and he warned thatenterpriseuse AI In the process, the model providers relied on may eventually have enough business knowledge to become potential competitors for the enterprise。
Nadella states that the company actually "pays twice" when it uses AI: one for model service token and the other for the know-how that was handed over to make the model work。
The more businesses want the model to understand their business, the more they need to enter more in-house information, and the hints, the smart body call tools, and the amendments made by employees to the model errors can also continuously expose the way the business operates and the criteria for decision-making。
Models learn from "use marks " , including tips entered by people, tools used by intelligent bodies, and, in particular, amendments made by people when models go wrong. Each amendment will eventually sink into intra-firm knowledge。
In Nadella, it appears that such knowledge is a business experience that is difficult for competitors to purchase directly, but may be progressively acquired by model providers in the long-term use of AI by enterprises. Model providers are thus not only aware of how customers operate, but may also use this information to enter the same business areas and compete with their original customers。
In his view, model providers could use public data training models, while limiting the ability of clients to study models through "distillation", which was not equivalent。
In response, he suggested that enterprises retain ownership of data such as tips, feedback, assessment results and business context, and create an exclusive learning environment at the cloud end, while avoiding over-reliance on a single supplier through layering access to different models。