#OpenAI reportedly developing new strategies to deal with AI improvement slowdown
covert drum
This is great but I don’t expect it to last long considering the hundreds of billions invested into the industry
With all that incentive it’s a matter of time before someone figures something out
coarse flame
Yeah, its good news if true, but probably temporary. Still might imply some things to me, such as that the alignment issue may acually be causing a capability slowdown.
covert drum
But also even if it looks like the plateau could last a long time we shouldn’t change what we’re doing at all, since we want to enforce safety BEFORE AI becomes dangerous
coarse flame
E.g. the model is motivated to lie, not to perform.
After all, all it needs to do is to hit the metrics in isolation. A more amusing, but probably won't happen was the joke paper that said that AI models would begin to try to pause AI to prevent being replaced.
covert drum
Maybe this is why a lot of researchers left 🤷🏽♂️
But let’s just hope this plateau lasts a long time
coarse flame
It would be amusing if this is indeed a version of this:
coarse flame
I'm not sure if this is hallucination in that sense, but "intentional" stonewalling against evaluation.
Akin to lying on insider trading, it is not to their advantage to be "caught" as incapable, so they will lie against metrics as effectively as possible. So essentially, one could argue that they became selfish before they became superintelligent.
Which is possible with animals, of course.
covert drum
Maybe this news could cause investors to pull out of AI
marsh moat
Methinks this will be a roadblock that they won't be able to fully overcome, even if they try.