In the Senate testimony and afterwards, Sam was pressed on what capabilities should trigger licensing. He mentioned biosecurity and cybersecurity as two that came to mind. Another, ‘manipulation’, was mentioned but I’m going to give him a pass on that one, because ‘manipulation’ is omnipresent everywhere and that particular horse left the barn a long time ago.
I suspect that Sam imagines that OpenAI will be one of the first to develop an LLM with dangerous capability in terms of the ones listed above, and that gives his team a seat at the table in terms of crafting regulatory capture.
OpenAI being the first to do these things is an improbable outcome. It’s very likely institutions and organizations with actual expertise in these areas will be the first to release capable models. They will have the required knowledge to fine tune models to have these dangerous capabilities. And they will be based on OS LLM backbones, because that’s what they’ll have access to.
I present two papers as evidence. The first, is a microsoft/berkeley paper using Llama to out-perform GPT4 on ApiBench, a benchmark for testing a models capability of providing accurate suggests for API calls to huggingface/pytorch apis.
The second paper is out of Meta/Carnegie Mellon. It talks about how they were able to achieve significant results by manually curating a fine tuning training set. Reading the paper, you start to appreciate the expertise required to do this.