Le 07/06/2023 à 23:28, Simon a écrit :
> altoid via Dng <dng@???> wrote:
>
>> As AI is trained by humans and its data sets ie: information comes
>> from what has been produced by humans.
> For now …
>
>> Thus, it goes to reason that *stupidity* and *racial bias*, both
>> heavily present in human history (among many other negative traits),
>> is part of AI.
>>
>> We have to assume that AI is trained with incomplete data sets,
>> either by choice or simply because *everything* is simply *not*
>> available.
> Just the other day I came across an interesting piece https://www.lightbluetouchpaper.org/2023/06/06/will-gpt-models-choke-on-their-own-exhaust/
> Summary: The models are currently trained on human written text. As time goes on, inevitably some of the input material will be AI generated. And the result will be that over time the models get worse and worse as each iteration incorporates the rubbish from the previous ones.
>
> "n our latest paper, we show that using model-generated content in training causes irreversible defects. The tails of the original content distribution disappear. Within a few generations, text becomes garbage, as Gaussian distributions converge and may even become delta functions. We call this effect model collapse.”
>
    This reminds me of a movie I watched long ago but don't remember
the name. A guy (adult) was cloned exactly like in Unix fork(): the two
clones shared the same past and only their future was different, well,
their present as the movie went on. But the clones were not all made
from the original guy, but from previous clones, and, everytime a small
defect appeared. After some levels of cloning, the result was badly
deficient. I enjoyed very much this movie.

    Actually, AI can become like the first of the class, able to repeat
all it is has learned, nothing else. That's a chance for us, because it
provides an opportunity for real humans to develop skills to detect it.

--     Didier