i am woog.
i studied math in undergrad, and learned programming afterwards. i now spend my time exploring various hobbies and ambitiously reverse engineering neural networks.
some of my projects and interests are listed below. when not working on technical things, i meditate.
as of jan 15th, i’ve been running an ai safety camp project, “towards ambitious mechanistic interpretability.”
the scope of our work is just doing things generally of value for the community with no particular conceptual focus or theme.
i mostly just supervise and coordinate good people to do good things and channel their minds effectively.
thus most work cannot be directly attributed to me, i can only list things i helped make happen.
mechanistic interpretability
external links
outputs of aisc involvements so far, that has links:
- eleutherai rnn-interp channel, dec 19. i organized a call with nora belrose to start mamba/rwkv interp fieldbuilding
- delphi suite, jan 19, ongoing project by: me, jett janiak (jj), jannik brinkmann (jb), goncalo paulo (gp)
- ghost gradients implementation, jan 26, by: g-w1
- arena-style intro to mamba mechint notebook, jan 29, by: me
- build from above: mamba steering vectors, riley kong. multiple ssm interventions, michael pearce (mp)
- brainstorming workshops: mamba (jan 31), sleeper agents (jan 7), events organized by: me
- computation in superposition (cis): math overviews, feb 8 and feb 18, writeups by: mp
- mamba explained, feb 11, blog by: kola ayonrinde (ka)
- cis with codebooks, feb 12, colab work by: mp, lucas hayne (lh)
- patchscopes implementation, feb 17, paper implementation by: fergus fettes (ff)
- atp* implementation, mar 14, paper implementation by: ka
stuff in the works that are incomplete or do not yet have links
- mechinterp.com website rework
- distillation of the whole mech interp field, for a general non-technical audience and for policy makers (to be self-published on above site)
- further attention superposition, a slideshow
- supervising various projects: weight visualization, mamba interp, circuits at initialization, etc
- bilinear interp
project graveyard/limbo
- efficient implementation of semiring backpropagation for computing path entropy
- sharing my thoughts more openly outside of random discord channels, maybe in a blog
- persistent cohomology on sae features (failed, should still write up negative findings)
hobbies
internal links
stuff i did in the past
- ran the math server for 3 years or so
- ran reading groups/putnam prep/colloquia in school
- probably other stuff im not thinking of
books i was reading in early 2023 (outdated)
external links
personal philosophy
internal links