People get mad when AIs do or say politically incorrect things. What if it's hard to prevent them from drawing such conclusions, and the easiest way to fix this is to teach them to hide what they think? That seems a scary skill to start teaching AIs.
-
I believe the criticism is that it shouldn't have learnt that way. If you train a face recognition AI with one ethnicity only, it will show bad results for other ethnicities. The fix is to use a more diverse training set, and not to just throw a PC result once in a while.
A case like that is easy to fix. But the more sophisticated AI gets, the harder it will be to get exactly the outputs you want by tweaking the inputs. The AI will start to draw conclusions from all sorts of apparently unrelated inputs.
-
-
Hum sounds like a subset of a more generic problem: what if, given a sufficiently complex AI, regardless of input tweaking, we're not pleased with the quality of the outputs? Be it because they're not pleasant, or accurate, or nuanced, or...
-
BTW I don't have an answer to that problem. It sounds like there's either something wrong with the AI itself, or with the inputs we give it, or with our own interpretation of reality
- Show replies
New conversation -
-
-
For GPT-3 (the reason for all this fuzz), the objective was to learn the statistics of language from the entire internet. If you begin priming it with controversial Q&As, and then ask something innocent, is going to answer something controversial, as it is likely to follow [1/2].
-
However you can do it in the opposite way too. Instead of seeing it as the model thinking is more similar to a highly specific Google search. Which is what all those cool Twitter demos are hiding. [2/2]
End of conversation
New conversation -
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Even more badly, it might understand systemic racism as the way things are done and that in turn ramps it up. Because now you also have computer systemic racism.
- Show replies
New conversation -
-
-
This Tweet is unavailable.
-
Good luck with that.
End of conversation
-
-
-
Humans do the same thing. Epistemology is the supposed solution. Ultimately AI will need to have some sort of training on logic, epistemology, ethics, and metaphysics to get "true" outputs from generic data. Otherwise it will make errant connections like a conspiracy theorist.
-
The issue here is not 'errant connections', but true statements that humans don't want to hear.
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.