The Buck Starts Here
How do we collectively determine whether or not large language models are dangerous when the public narrative is held captive by the companies that build them? This has a chilling effect on a dialogue that needs to take place about AI governance. The march of progress and the specific direction it takes is presented to us as inevitable. The layperson is meant to standby while those that hoard all the data and compute make the rules about how AI is developed, deployed, used.
Rules imply and enact a particular ethics: separating what is deemed good from what is deemed evil. All that is nonsense to an LLM – there are no latent, statistically significant patterns to derive. Yet to achieve anything meaningful, AI must encode some worldview, as expressed through the datapoints chosen for measurement and sampling, and perhaps just as importantly through the datapoints that are ignored. Therefore AI is in no sense amoral; the morals are right there coupled with the weights.
So then, should we trust model developers to have our best interests in mind? The more we rely on AI, the more necessary it becomes to identify who must account for undesirable outcomes. Much like a transformer turns raw data into statistical relationships, our ethics must be transformed from abstract notions into frameworks for assigning responsibility, starting with a model's designers and developers. Towards this end, the recent Executive Order is a step in the right direction. Where some see the federal government positioning itself as a gatekeeper to innovation, I see a long overdue accountability mechanism.
The ultimate goal should be disarming the language we use to talk about AI. The dialogue should be welcoming and accessible to the general public. We all should resist the urge to traffic in fear and panic so a more genuine dialogue can take place.