“I think we will get the hallucination problem to a much, much better place,” Altman said. “I think it will take us a year and a half, two years. Something like that. But at that point we won’t still talk about these. There’s a balance between creativity and perfect accuracy, and the model will need to learn when you want one or the other.”
What is Altman is talking about? An #LLM has no notion of truth or accuracy, so you can't just dial up some "truth coefficient."
@chrisoffner3d the “hallucination” or rather confabulation is an integral feature of current LLM systems and I don’t think there is a clear path to get rid of that. Not saying it’s impossible, but I feel like systems that cope with that would have to look substantially different having a memory structure they can infer from. To me it seems like he’s just trying to cater to his shareholders..
@kentindell@chrisoffner3d In fact, I see quite a bit of similarities between what Chris has mentioned in the post cited below and the extremely dangerous assumptions that #Tesla has been using to underpin their #FSDBeta program - namely, the so-called "generalized self-driving" (no defined Operational Design Domain).
The Tesla #Autopilot Team has always embraced a very primitive safety strategy that they try to sell as "validation" - if I am being generous.
@kentindell@chrisoffner3d The fact is that a automated driving system that does not require a human driver fallback under any conditions over an unbounded ODD would require a engineered system such that it would be capable of validating itself, continuously.
#Tesla has never remotely offered up a safety case to achieve that.
No one has.
Instead, they just sloppily scope it as a "software problem" and toss it out there, as you are aware.
Add comment