AI companies say rivals are distilling their models and why it's so hard to stop (www.scientificamerican.com)

AI companies, notably Anthropic and OpenAI, are increasingly concerned about the unauthorized distillation of their proprietary models, a process akin to extracting distilled alcohol from a larger batch. This practice allows rival developers, particularly those from China, to create smaller, imitation models that mimic the performance of original ones at a significantly reduced cost. In recent reports, both companies shared their experiences in thwarting campaigns aimed at distilling their AI technologies, highlighting a pressing issue in the competitive AI landscape where high valuations are at stake. The significance of this development lies in its implications for innovation and competition within the AI/ML community. Distillation not only helps competitors bypass the hefty investment required to build advanced models but also raises concerns about intellectual property theft and model capabilities. Techniques to distill models typically involve prompting the original AI extensively to gather response data, a method that poses a complex security challenge for companies striving to protect their assets. Efforts to curb this practice will need to balance improving security measures without sacrificing the usability and performance of their models in the public domain, as they face an ever-evolving threat from more sophisticated distillation methods.