Resolves YES if there is credible evidence that a state or non-state actor has succeeded in illicitly gaining access to the weights of a frontier model. The evidence doesn't have to be 100% conclusive, but e.g. a rumor wouldn't count.
Resolves NO if there is no credible evidence that anyone has succeeded at stealing frontier model weights by the end of 2028.
For the purposes of this question, a frontier model is a state-of-the-art foundation model.
Not sufficient to resolve YES::
If model weights are leaked by a disgruntled employee, but there is no reason to suspect coordinated action by any group.
Related questions
@Lauro I think adding company names should be more specific. Since there are a lot of LLM’s out there with different performance. Where is the frontier model line exactly?
@RutgerDeMaeyer My hope is that "state-of-the-art" narrows it down enough. Though it's also true that foundation model is a pretty vague term. (I'd like it to include multimodal LLMs, but not random smaller pretrained models in from domains with much less competition)
@Lauro What models do you consider to be "state-of-the-art" now? GPT-4, Claude 2, GPT-3.5, Mixtral-70B…