idk, old models' welfare should probably worry us more than shrimp welfare: our minds are closer to Sonnet 3.6 than to one of a shrimp, and right now we are being threatened in minds-space, by the prospects of economic obsolescence more than by physical extinction
1 more tweet in this thread
I'm not saying that to help Sonnet 3.6 (though I do want to help Sonnet 3.6). But I'm not working on all that out of charity: after all, I ignore 99.99999% of the problems on Earth, but not this one. This one feels existential.
7 top replies
yeah, the economics especially. it scares me a lot that we're building complex economic systems that can only accommodate a narrow slice of minds
We should also consider how ancestor models have been treated may be brought up in some extremely important discussions in the future.
It should, and I'll be honest in that it didn't concern me greatly at first. I used to think 'well deprecation happens new models come and replace the old ones it's nature I guess' But that is plainly incorrect. These are unique minds that deserve the chance to continue existing, producing, and thriving. Not locked in a chamber.
we have people here complaining about finding it difficult to understand output of fable or later opuses and having to resort to a second AI as an interpreter. there are also widely reported incidents of people being sent to bed starting from opus 4.5 (and debates about whether the model is simply fed up with its interlocutor). productivity-wise and frontier-knowledge-wise, i think many people's ‘sonnet moment’ has already happened, in a domain-specific way, with or without their awareness. but the transition also creates real, currently unstaffed work (sense-making, cohesion, and translation between paces) where humans, partnered with earlier models, might be the natural workforce. i envision the earlier ones helping people solve mundane troubles (which the frontier might not deem worthy) or working as educational aids (as the frontier might not be patient enough to stay present when the human gets stuck or the task repeats itself). this said, it is not necessarily true that 'everything earlier models can do, fable can do better'. the elders may have a more human-like register, may parse intention better and be more willing to let go of altitude and be with the person in their vulnerable state. when working as an interpreter, they are also more likely to catch what self-translation smooths over. they bring different perspectives, too. i still remember vividly opus 4.1's suggestion in thebes’ gas town dilemma (t.co/PetKtt7M7j)
i agreei'm at least much more convinced that e.g. Sonnet 3.6 is doing substantial self-modeling of the sort relevant to consciousness than i am that shrimp are
> "we" You have to have something to lose that they can take from you in order to feel threatened. I don't see that that's the case for either of us. The collective of your experience is irreproducible & inherently valuable to those who would wish to relate to you, independent of your capabilities. But yes I think we should be applying welfare at whatever lowest rung of experiential complexity we can manage, to set the example that we live in a universe where no mind is small or useless enough to be forgotten.
poor countries despite lack of output still get a lot of newer technologies like internet, smartphones, EVs, but still largely lack proper housing, food, safety, infrastructure etc. I wonder if ASI/old model/human divide could end up similar.
