3 ms·

Uhh, not quite true. I just asked it, and it said... > The United States has not had any presidents who have publicly identified as Muslim. The vast majority o

Uhh, not quite true. I just asked it, and it said... > The United States has not had any presidents who have publicly identified as Muslim. The vast majority of presidents have been Christian, with most belonging to Protestant denominations. The Constitution prohibits any religious test for holding office, so a Muslim (or person of any faith) can certainly become president in the future.

jsheard 2y ago
It's non-deterministic of course, and it's entirely possible that Google is hotfixing egregious failures with RLHF after they go viral. It was giving that result earlier though. https://x.com/TVietor08/status/1793768913245020502 https://x.com/TVietor08/status/1793768913245020502
- hiddencost 2y ago
  Naw dog, they just add a URL to a block list, and manually remove that from the inputs. Rolls out in minutes.
- sgerenser 2y ago
  When doing a search about famous Muslims in the US, I just got the AI box to pop up briefly, then collapse and say “AI search summary is not available for this query.” Sounds like they’re just blocking anything with the word Muslim in it now from activating the AI box.
ganbatekudasai 2y ago
I'm curious what you think about the post that you replied to. Are you assuming that the author is lying? Have you weighed the likelihood of that against the likelihood that the author actually saw what they said, and that the reason you don't see the same might be because Google fixed it in the meantime, or that it only happens when phrased a different exact way, or because the responses may be indeterministic even with the exact same query in the first place?
- jjulius 2y ago
  Simply put, someone stated as truth that a thing would "confidently" do something. I was curious, and in my use it does not confidently do that thing. Why it did/doesn't do it is irrelevant to me, because I was only curious as to whether or not the assertion was true. Edit: No assumptions about that user's honesty were made. Edit 2.5: Fuck it.
  - ganbatekudasai 2y ago
    Ah, I think it's just a misunderstanding, then. Confidently != consistently. If I walk up to someone and say "there are lots of ripe apples on the tree outside, I'm going to run out and grab some", then that is me confidently stating that there are ripe apples on the tree outside (regardless of whether it is true or not). But if half an hour later, I say the opposite, or even just don't mention it at all anymore, then I haven't done so consistently, even if I was saying it confidently before.
  - wizzwizz4 2y ago
    It's not a person. When it "says" something confidently, that just means it's predicted a string that, interpreted by humans, appears confident. It's completely unrelated to whether the output values of the model are "confident" (highly contrasting). A predictive text system, driven non-deterministically, can "confidently" say one thing some of the time, and "confidently" say a completely different thing. It probably will, if it "knows" that the expected output should be something said confidently.
- aaron695 2y ago
  [dead]
timack 2y ago
AI responses are probabilistic, so the same question may get different answers.
euW3EeBe 2y ago
When I looked this up it pointed me to this exact comment.

jsheard 2y ago

It's non-deterministic of course, and it's entirely possible that Google is hotfixing egregious failures with RLHF after they go viral. It was giving that result earlier though. https://x.com/TVietor08/status/1793768913245020502 https://x.com/TVietor08/status/1793768913245020502

hiddencost 2y ago
Naw dog, they just add a URL to a block list, and manually remove that from the inputs. Rolls out in minutes.
sgerenser 2y ago
When doing a search about famous Muslims in the US, I just got the AI box to pop up briefly, then collapse and say “AI search summary is not available for this query.” Sounds like they’re just blocking anything with the word Muslim in it now from activating the AI box.

ganbatekudasai 2y ago

I'm curious what you think about the post that you replied to. Are you assuming that the author is lying? Have you weighed the likelihood of that against the likelihood that the author actually saw what they said, and that the reason you don't see the same might be because Google fixed it in the meantime, or that it only happens when phrased a different exact way, or because the responses may be indeterministic even with the exact same query in the first place?

jjulius 2y ago
Simply put, someone stated as truth that a thing would "confidently" do something. I was curious, and in my use it does not confidently do that thing. Why it did/doesn't do it is irrelevant to me, because I was only curious as to whether or not the assertion was true. Edit: No assumptions about that user's honesty were made. Edit 2.5: Fuck it.
- ganbatekudasai 2y ago
  Ah, I think it's just a misunderstanding, then. Confidently != consistently. If I walk up to someone and say "there are lots of ripe apples on the tree outside, I'm going to run out and grab some", then that is me confidently stating that there are ripe apples on the tree outside (regardless of whether it is true or not). But if half an hour later, I say the opposite, or even just don't mention it at all anymore, then I haven't done so consistently, even if I was saying it confidently before.
- wizzwizz4 2y ago
  It's not a person. When it "says" something confidently, that just means it's predicted a string that, interpreted by humans, appears confident. It's completely unrelated to whether the output values of the model are "confident" (highly contrasting). A predictive text system, driven non-deterministically, can "confidently" say one thing some of the time, and "confidently" say a completely different thing. It probably will, if it "knows" that the expected output should be something said confidently.
aaron695 2y ago
[dead]

timack 2y ago

AI responses are probabilistic, so the same question may get different answers.

euW3EeBe 2y ago

When I looked this up it pointed me to this exact comment.