This work represents progress on removing attention head superposition. We are excited by this approach but acknowledge there are currently various limitations. In the short term, we will be working on adjacent problems are excited to collaborate with anyone thinking about similar things! Produced as part of the ML Alignment...
This work represents progress on removing attention head superposition. We are excited by this approach but acknowledge there are currently various limitations. In the short term, we will be working on adjacent problems are excited to collaborate with anyone thinking about similar things!
Produced as part of the ML Alignment & Theory Scholars Program - Summer 2023 Cohort
Summary: In transformer language models, attention head superposition makes it difficult to study the function of individual attention heads in isolation. We study a particular kind of attention head superposition that involves constructive and destructive interference between the outputs of different attention heads. We propose a novel architecture - a ‘gated attention block’ - which resolves this kind... (read 4381 more words →)
Yay it's back up again.
What is the exact formal difference/relation between probability distribution , random variable , and causal model ?
And you can't just get that stuff by asking your mind to tell you that stuff; you have to let your mind do its thing sometimes, without requiring a legible justification.
I need to really internalize this. I tend to "rediscover" this and forget about it the moment things become more stressful.
Requiring that things be explicitly justified in terms of consequences forces a dichotomy: if you want to do something for reasons other than explicit justifications in terms of consequences, then you either have to lie about its explicit consequences, or you have to not do the thing. Both options are bad.
Unless one explicit justification is simply that "this is fun to do,... (read more)
I really appreciate the existence of this functionality, since when it comes to trying to do something I either overdo or don't do at all. Shortforms make it managable and still publicly shared, so I still have enough incentive to create something nice and understandable. A much much better Twitter, and allows everyone to glimpse into the "studio" of very good writers.
Oh wow I can include images like this! Very nice.
Kind of agree. What do you think is the most value you get from this? Was wondering about whether I should post accountability-related stuff here (e.g. by editing the post) but not sure if it makes sense.
I have been feeling extremely impostery lately and do agree on the critical self-evaluation tendency. For the last month or so I felt entirely stuck with even the idea of an application giving me severe anxiety. Have been overcoming this slightly lately but I think this post and the conversations it caused has made em feel better. Thank you.
Managed to end up submitting some applications to places I think I would like to work at. I think things that helped me do this (notwithstanding chance/luck) were (in descending order of perceived priority):
This feels like a mini-twitter. Interesting
colab notebook
this interactive notebook
check out the notebook
notebook
First link is not like the others.
I propose a compromise: 1.5 hoses. This will lead to greatness.