Tom McGrath (@banburismus_) 's Twitter Profile
Tom McGrath

@banburismus_

wir müssen wissen - wir werden wissen

ID: 1295978398557319168

calendar_today19-08-2020 06:58:31

535 Tweet

1,1K Followers

343 Following

Tom McGrath (@banburismus_) 's Twitter Profile Photo

when we change models we should just be able to look at model diffs, decide what we like and don’t like, and only keep the changes we want. i think we can build something like this - maybe we should?

when we change models we should just be able to look at model diffs, decide what we like and don’t like, and only keep the changes we want. i think we can build something like this - maybe we should?
Tom McGrath (@banburismus_) 's Twitter Profile Photo

new demo! instead of prompting an image model, try using it as a paintbrush instead. our new paint with ember demo shows what you can do by steering a model's activations - powered by mech interp!

Myra Deng (@myra_deng) 's Twitter Profile Photo

>be you >work in HFT >have existential dread >see this tweet, wonder if your skills could be better used to make AGI safe >apply to attend our happy hour, meet the Goodfire team >build safe AGI

Tom McGrath (@banburismus_) 's Twitter Profile Photo

we've got some really cool new work looking at Anthropic's circuit tracing work and comparing it to a circuit that's already been studied. some really interesting findings in here