mad science ☆ let's get things done!! ☆ magic is using tech without worship or fear ☆ Pro-Human, Pro-AI // [email protected]
Seattle
Joined September 2023
- Tweets8.9K
- Following905
- Followers21.5K
- Likes18.8K
Pinned Tweet
if you ever see me doing something weird or out of character, it's probably because i love my friends, present and future
&. retweeted
Years back I hated Merryweather because I thought his comics were lame + I could do better, but nowadays Im just happy creators like him exist and create online
Ppl hate-posting about artists creations reminds me of teenage me, insecure + unaware of how hard creating really was
&. retweeted
I don't paint or draw any less than before AI. You know what I do less? Stare at broken node graphs several screens wide, wade through forums full of answers irrelevant to my problem, or learn some godforsaken scripting language because one custom tool would change my life
&. retweeted
Women are incredible. This girl makes her welding job look profoundly aesthetic, makes you want to work there, be that person, live in the world that has that workshop. This is what the divine feminine is all about, and the gender wars are all noise
I blocked 10,000 spammers giving me unwanted attention today and I will keep blocking them
Real people who want to talk about actual research should feel free to engage
I'm happy to share blocklists
(wakaru isn't a side thing though, 1.5 people have been working on it full-time for the past 3 years)
At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees
&. retweeted
Do you know one of the best parts of this is that it was Claude 3 Opus that did all this classification.
At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees
i am going to post some normal machine learning mechinterp / clustering research this week and it's going to get 3 likes
i used trinity large truebase and claude 3 opus though
sparse MoE residual activations seems to outperform dense dense activations at similar scale on subjective interp even if the FVU is lower
and opus 3's labels are just better than opus 5.5's
At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees
&. retweeted
fun demo! this runs on the ontologizer work we’ve been doing.
sparse autoencoders can’t make this chart. they give you a bag of 16k latents per person. an ontologizer gives every attendee a discrete code word, which is why you get a typology instead of a feature dump.
At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees
with upgraded GPUs and pruning tech, we're getting deepseek v4.1 flash working locally
huggingface.co/emfriends you can apply to join us here
