On one hand LLMs make Tesla look bad but I wouldn’t be surprised if when put into the real world these models will have the same problem. Always 5 years away from fixing all the edge cases
So you have some stats proving that Tesla cars are more dangerous or crash more often than other cars? With so many Teslas on the road, there must be an obvious spike if AI is making them more dangerous than the baseline of all cars.
That sounds totally irrelevant to the discussion of whether Tesla AI has caused more deaths than other AI. Even if human-driven cars crashed 100% of the time, and Tesla AI crashed into a tree one time, this would not refute my original claim, which was that Tesla AI has directly killed more people than any other AI.
If you want to use a statistical definition of how many people Tesla AI has killed by comparing it to a baseline of non-AI cars, then you also need to do this for every other non-Tesla AI that you might assert has killed people. For example, for the medical misdiagnoses referenced in the sibling discussion, you would need to ask whether ChatGPT has misdiagnosed more cancers than doctors.
I'd be surprised if there's enough evidence to determine in either direction, for that.
How many used an AI for medical advice? A medical AI which said "no cancer" when there was? Does Therac-25 count as GOFAI? How many have taken an LLM at face value about some topic, and like those stories about GPS directions gone wrong, done something daft that we've not necessarily yet heard about (or if we have, the headline was "Florida Man does X" rather than "AI tells man to do X, and he does"?)
It's like how we don't know how many people had a crash shortly after failing to notice FSD had switched itself off.
I think at this point we can reasonably assume nobody has used AI for cancer diagnosis. Give it a year or two and maybe that's valid. Or maybe it even is true now if some high number of people have asked ChatGPT for tips about symptoms and decided not to go to the doctor based on its advice.
But either way, it's inarguable that Tesla AI has directly killed more people than any other publicly known AI. I suppose you could consider guided missiles, but those are not really AI (although you could make a similar argument that neither is FSD).
That article doesn't mention a service that was open to users (not to mention the point of the article is that AI had fewer false positives and false negatives than doctors, which would invalidate the premise of this argument). But if you want to apply the same logic to ChatGPT, then even if it's true that misdiagnoses are leading to skipped doctor's visits, it's still unlikely that anyone has died yet from that lack of preventive care. ChatGPT launched a few months ago, and it's unlikely anyone in a late stage of cancer would have prevented their death if they went to the doctor instead of asking ChatGPT over the past few months. So for anyone affected by a ChatGPT misdiagnosis, it will take some time for the cancer to kill them. And note that it is the cancer that will kill them, not the AI.
On the other hand, a "self-driving car" driving into a tree and killing its occupants seems an obviously more direct case of death by AI than a user asking an AI if it has cancer and the AI saying no. And if you want to make the argument that Tesla drivers are supposed to have their hands on the wheel, then you have to also make the argument that ChatGPT users aren't supposed to use it for medical advice.
> That article doesn't mention a service that was open to users
So? Most AI isn't. It's not all consumer products.
> not to mention the point of the article is that AI had fewer false positives and false negatives than doctors, which would invalidate the premise of this argument