I'm Joyee, an aspiring AI alignment researcher (undergrad) at Berkeley, and I'm just getting started with distillations! Since I've heard much of iteration as the way to improvement, and "colliding one's mental model with reality as frequently as possible", I'd like to ask: can someone critique my first one, on...
After getting to the the part of a demo (one of Neel Nanda's interp demos) where they talk about the idea of a Logit Lens, and Layer Attribution, I have a bit of trouble visualizing it as I could have for simpler concepts (e.g. residual streams, which were indeed drawn...
I'm Joyee, an aspiring AI alignment researcher (undergrad) at Berkeley, and I'm just getting started with distillations! Since I've heard much of iteration as the way to improvement, and "colliding one's mental model with reality as frequently as possible", I'd like to ask: can someone critique my first one, on the paper "Scaling Laws for Transfer" through the lens of alignment automation efforts?
https://drive.google.com/file/d/1vCV1XWxCxa5rvVIvSEfGfbJ26rtqyaQm/view?usp=sharing
At least as a beginner, my philosophy of distillation is roughly similar to https://www.lesswrong.com/posts/XudyT6rotaCEe4bsp/the-benefits-of-distillation-in-research especially when it comes to "lenses".
I fly the flag of Crocker's rules.
I also highly recommend Laudan's blog, full of well argued threads of food history that go beyond pat-and-dry preconceptions of foodways! https://www.rachellaudan.com/blog
Hint: does Charlie Chaplin have a gears-level understanding of the system?
After getting to the the part of a demo (one of Neel Nanda's interp demos) where they talk about the idea of a Logit Lens, and Layer Attribution, I have a bit of trouble visualizing it as I could have for simpler concepts (e.g. residual streams, which were indeed drawn as my primary method of comprehension). Anybody have good resources for illustrations? (I know Nanda had a great colorful runthrough but it was only at a high level and for encoder-decoder machines, and thus not generalizable)