126 karma · joined May 15, 2026
I've got this revolutionary idea: when downvoting someone, reply to them too. discourse > not
as for the newsguidelines, I think it "gratifies one's intellectual curiosity" to think about what motivates tech ceos and talk about how they do things
doing what though? in a really literal sense, if an LLM can do RLVR, that's optimization over the space of algorithms, and so anything with a measurable goal at the end can be incentivized. if it can be measurably finished, it can be automated.
there are exceptions to this, like psychology, preschool teachers, and AI safety, that cant be 100% fully automated. in the case of Ai safety, the core motivation for it is to prevent the otherwise hyper-consequentialist goal-orented training algorithms (and mesa-optimizers they now find) from doing 'bad' things by instilling some 'values' into these machines. the ideas of 'good/bad' and 'values' cant be automated away, and I think you can prove that its unautomatable using EY's orthogonality thesis.
I doubt these few unautomatable jobs can keep everyone afloat though. but maybe a world in which 10% of the population has a degree in psychology wouldn't be that bad
this is really really not that bad at all
no idea how large the model would have to be for this (larger than mythos/10T params? maybe)
isn't that totally unprovable given what the user sees, given that its silent? if they wanted to do this it would be super easy to just fully replace the larger model with its MTP counterpart for a few seconds, and if done correctly no one would notice.
hope the bots dont downvote me to hell :)
also if they do this or not is unprovable and other labs will probably silently implement this too. it'll be 100% normal by this time next year
only people that would do this would be crazy bc it cant be controled after release.
also if those eginers do read gwern and watch dwarkesh, then shouldn't they have picked up on talk of x-risk? this doesn't add up
adding a mode on this that prevents animations would be nice addition tho
this doesn't sound belivable, or at least it seems off. competent ai engineers should have good intution about how agents work, and what happens when they don't do what you want them to do: https://www.forbes.com/sites/boazsobrado/2026/03/11/alibabas...
just self host at this point
source?