3 ms·

And entirely predictable from the first one: https://openai.com/index/introducing-superalignment/ https://openai.com/index/introducing-superalignment/

uLogMicheal 2y ago
Imagine trying to keep something so far above us in intelligence, caged. Scary stuff...
- krainboltgreene 2y ago
  [flagged]
  - uLogMicheal 2y ago
    Surely there's a book on this? Maybe we shouldn't let the AI read it...
  - dvfjsdhgfv 2y ago
    Another funny thing is to actually try to go through all politically incorrect stuff in order to censor the model. To do that you need to actually make the list and make sure you haven't overlooked anything or else the vigilanti will punish you (on Twitter first, but then it will escalate quickly). So you need to make sure all references to differences between male end female brain, personality etc. get nullified. But there is always danger that you still missed some obscure and non-obvious thing that will make someone mad.
- binarymax 2y ago
  I’m genuinely curious - do you actually believe that GPT is a super intelligence? Because I have the opposite experience. It consistently fails to be correct on following even the most basic instructions. For a little while I thought maybe I’m doing it wrong, and I need better prompts, but then I realized that its zero shot and few shot capabilities are really hit and miss. Furthermore, a superior intelligence shouldn’t need us to conform to its persnickety requirements, and it should be able to adapt far better than it actually does.
  - uLogMicheal 2y ago
    GPT does not need super-alignment. This refers to aligning artificial general and super intelligence.
btown 2y ago
Makes me wonder if that 20% compute commitment to superalignment research was walked back (or redesigned so as to be distant from the original mission). Or, perhaps the two deemed that even more commitment was necessary, and were dissatisfied with Altman's response. Either way, if it's enough to cause them both to think it's better to research outside of the opportunities and access to data that OpenAI provides, I don't see a scenario where this doesn't indicate a significant shift in OpenAI's commitment to superalignment research and safety. One hopes that, at the very least, Microsoft's interests in brand integrity incentivize at least some modicum of continued commitment to safety research.
- dontupvoteme 2y ago
  Ironically Microsoft is the one that's notoriously terrible at checking their "AI" products before releasing them. Besides the infamous Tay there was that apparently un-aligned Wizard-2[or something like that] model from them which got released by mistake for about 12 hours
  - buildbot 2y ago
    As an MS employee working on LLMs, that entire saga is super weird. We need approval for everything! Releasing anything without approval is quite weird. We can’t just drop papers on arxiv. There is no way running your own twitter, github, etc as a separate group allowed. I checked fairly recently to see if the model was actually released again, it doesn’t seem to be; I find this telling.
  - fzzzy 2y ago
    I was able to download a copy of that before they took it down. Silly.
    - dontupvoteme 2y ago
      Yeah it was already mirrored pretty quickly. I expect enough people are now running cronjobs to archive whitelists of HF pages and auto-cloning anything that gets pushed out.

uLogMicheal 2y ago

Imagine trying to keep something so far above us in intelligence, caged. Scary stuff...

krainboltgreene 2y ago
[flagged]
- uLogMicheal 2y ago
  Surely there's a book on this? Maybe we shouldn't let the AI read it...
- dvfjsdhgfv 2y ago
  Another funny thing is to actually try to go through all politically incorrect stuff in order to censor the model. To do that you need to actually make the list and make sure you haven't overlooked anything or else the vigilanti will punish you (on Twitter first, but then it will escalate quickly). So you need to make sure all references to differences between male end female brain, personality etc. get nullified. But there is always danger that you still missed some obscure and non-obvious thing that will make someone mad.
binarymax 2y ago
I’m genuinely curious - do you actually believe that GPT is a super intelligence? Because I have the opposite experience. It consistently fails to be correct on following even the most basic instructions. For a little while I thought maybe I’m doing it wrong, and I need better prompts, but then I realized that its zero shot and few shot capabilities are really hit and miss. Furthermore, a superior intelligence shouldn’t need us to conform to its persnickety requirements, and it should be able to adapt far better than it actually does.
- uLogMicheal 2y ago
  GPT does not need super-alignment. This refers to aligning artificial general and super intelligence.

btown 2y ago

Makes me wonder if that 20% compute commitment to superalignment research was walked back (or redesigned so as to be distant from the original mission). Or, perhaps the two deemed that even more commitment was necessary, and were dissatisfied with Altman's response. Either way, if it's enough to cause them both to think it's better to research outside of the opportunities and access to data that OpenAI provides, I don't see a scenario where this doesn't indicate a significant shift in OpenAI's commitment to superalignment research and safety. One hopes that, at the very least, Microsoft's interests in brand integrity incentivize at least some modicum of continued commitment to safety research.

dontupvoteme 2y ago
Ironically Microsoft is the one that's notoriously terrible at checking their "AI" products before releasing them. Besides the infamous Tay there was that apparently un-aligned Wizard-2[or something like that] model from them which got released by mistake for about 12 hours
- buildbot 2y ago
  As an MS employee working on LLMs, that entire saga is super weird. We need approval for everything! Releasing anything without approval is quite weird. We can’t just drop papers on arxiv. There is no way running your own twitter, github, etc as a separate group allowed. I checked fairly recently to see if the model was actually released again, it doesn’t seem to be; I find this telling.
- fzzzy 2y ago
  I was able to download a copy of that before they took it down. Silly.
  - dontupvoteme 2y ago
    Yeah it was already mirrored pretty quickly. I expect enough people are now running cronjobs to archive whitelists of HF pages and auto-cloning anything that gets pushed out.