| > PGP for example is not dangerous ... If PGP had "escaped containment" I think the disconnect here is the dishonest framing that OpenAI have purposefully put on this incident. Their LLM did not "escape containment" in this "rogue AI" Hollywood blockbuster sense that they want you to believe. An LLM cannot do that; it is just a static computation graph that outputs a probability distribution. What escaped "containment" is a piece of software they wrote that uses an LLM and which, through OpenAI's negligence, lacked ordinary technical and human guardrails that should have been present. It was run in "YOLO mode" with gobs of compute and no human oversight. That situation is the source of the danger, not the model weights. The model itself was simply given the opportunity to generate a bazillion commands, many of which were nonsensical garbage, and a few of which worked. It was not the work of a supreme intelligence. This distinction is very important, because OpenAI have shifted the discourse from where it should be (OpenAI's negligent use of software) to a place that serves their strategic business interests. LLMs remix existing knowledge and output text. That is not inherently hazardous. What can be hazardous is bolting systems that touch the real world onto them. But it's always possible to make a tool dangerous by misusing it. For example, vehicles (which do have inherent dangers) are widely accepted as ordinary and useful tools. So are bricks. But if you strap a brick to the accelerator, you have made an autonomous murder machine that can escape containment! The existence of this possibility does not imply that we should rethink automobile licensing or the sale of bricks. |