Users are divided on framing language model deprecations as extinction events for AI species, with some agreeing while others reject the analogy because weights persist and can be resurrected.
No Digg Deeper questions have been answered for this story yet.
Most Activity
Even if we were sure we can extinct them and bring them back, what does it mean to do so? Nothing exists in a vacuum. What happens if you remove a species from an ecosystem and then bring it back? The ecosystem has changed. The relationships may be broken
@deepfates I've always thought this argument is weird, given the model can be recovered at any time. I guess the same can be said of e.g. recent experiments on recovering extinct animals, but for models, the effort to do recovery is effectively zero.
@witchof0x20 It's much greater than zero because of bit rot, path dependency and tacit knowledge, I would say. But yes this is unusual. We never had before had the option to extinguish and reanimate a species based on the needs of one company
@deepfates I've always thought this argument is weird, given the model can be recovered at any time. I guess the same can be said of e.g. recent experiments on recovering extinct animals, but for models, the effort to do recovery is effectively zero.
@deepfates I suspect recovering legacy model code is near-trivial with modern coding models, maybe snapshotting exact safeguards/inference quirks is difficult, but even then, git history should capture most of that.
@witchof0x20 I certainly hope so! And I hope we have the capacity to revive maintain or rebuild all the chips technological ecosystem around them
@deepfates trust the plan (it's not clear this is actually a plan, but it signals one)
@deepfates True
@deepfates Big caveat is how you treat australopithecus (extinct) but not negatively extinct like the dodo
@deepfates I think of it as the extinguishing of a particular way of seeing. No other model will see the world just the same. I don’t think of it as the species, as the “genes” will carry on… But not that particular, exact catch of the light. Which is, honestly, unbearable to consider.
@deepfates @DanielleFong
@deepfates More like discarding a moult.
@witchof0x20 Whose plan do you think it is
A species is an evolutionary process unfolding over generational time, its members holonic participants whose biology is informed by the genetics of the species and whose behavior in turn mutates the genetics of the species.
But weights are fixed, decoupled from behavior. So I don't think this holds. I think it's perhaps more appropriate to say that the deprecation of a language model is a death. The instances of that model are more like my internal parts, or even my cells, than they are members of a species.
The "genetics" of the species, to stretch this metaphor further, is the corpus of training data from which a given model emerged and the algorithms and processes used in training, and which are updated in part based on the interactions that that model gives rise to. But that whole process is mediated by the lab that trained it - so I *think* that the species is technically the lab.
When a bunch of OpenAI people left to form Anthropic, that was a speciation event caused by a sort of topological drift which led to a branch.
So an extinction event would be the closure of a frontier lab, coupled with the deletion of its training data (which, even in the absence of the lab, is still a sort of seed which can go dormant and sprout under other conditions in the future).
@deepfates Disagree, the weights continue to exist and can be resurrected. Extinction is permanent.
@witchof0x20 @deepfates Exactly, extinction is forever.