Val

[Val @val_kharvd]

@camhberg @repligate can you be more specific how this might work?

[Val @val_kharvd]

@neallseth murray hill is not a borough

[Val @val_kharvd]

@AcombStephen 👾

[Val @val_kharvd]

@nabla_theta @chrislakin more like

[Val @val_kharvd]

@stochasticchasm @scaling01

[Val @val_kharvd]

expressing “frustration” when stuck is instrumentally helpful for achieving higher rewards

[Stella Biderman @BlancheMinerva]

This post really highlights why I’m baffled by mainstream approaches to interp that ignore learning dynamics. To me the natural interpretation of “but why?” requires that you talk about training data and reward signals.

[Val @val_kharvd]

richard hamming on vibecoding

[Val @val_kharvd]

@npenzenstadler

[Val @val_kharvd]

@nearcyan 996

[Val @val_kharvd]

@round_piglets 😬

[Val @val_kharvd]

it’s a long tradition apparently https://t.co/hGrXgtaHXj

[Val @val_kharvd]

TIL there’s a chinese law governing buddha reincarnations

[Val @val_kharvd]

@norvid_studies russian oligarch pipeline

[Val @val_kharvd]

@transkatgirl

[Val @val_kharvd]

@stochasticchasm @livgorton alhamdulillah

[Val @val_kharvd]

@idavidrein do you think it’s possible that there are sentient entities whose qualia don’t bear valence in the same sense as they do for us? i.e. sentient beings incapable of suffering

[Val @val_kharvd]

@1a3orn https://t.co/kkkGY8eWe3

[Val @val_kharvd]

@jarredsumner @bshlgrs openai is not constrained by money, but it is constrained by gpu supply. there is just not enough. everyone is running at 95+% utilization and is strongly incentivized to optimize inference

[Val @val_kharvd]

@repligate @CFGeek if you had access to training RL rollouts, how would you test this?