Super excited about this work! This paper was driven by a claim I've been making to anyone who'll listen: "interpretability is the language of data". (1/3)
#4: fart fishing
Buried in Dolci is a cluster of very specific fan fiction, where characters fart in ponds, causing fish to die from the smell.
The chosen responses in the dataset wrote vivid scenes, while the rejected refused, teaching the model to comply! (7/9)
The only time someone in my family who wasn’t my mom or grandma hit me was when I was 16 and my uncle was working on his 1970s Mustang in the driveway and I asked how someone could fix a car every day for 20 years and it still didn’t work
Looking at data is actually critical work. It corrects so many issues very quickly that skew algorithmic performance. Should be owned by everyone.
I agree with Alex here.
Grad school is a comparable public health hazard to all STDs:
~400k US grad students
~40% have moderate-to-severe depression vs. 6% general population
DALY estimate is ~40% reduction in quality of life
So cost is 54,000 DALY per year
Similar to HIV and all STDs in N. America
Excuse me, I think there’s been a mistake, I need to speak with the manager. I was supposed to be part of the in-group the law protects but does not bind, not the out-group it binds but does not protect.
Was digging through my drive and found this random 11th grade homework assignment from 8 years ago
My favorite part is me claiming - without reference - that "it is often said that the dream of coding is to use english" and saying that once this happens coders will go extinct.
Really cool
When they trained GPT3 they had loss spikes because they scraped from a subreddit of microwave noises
That training batch was literally text like "mmmmmmmmmmmmmmm"
Whenever I see someone knock down one of those stone stacks I think “Oh yeah, you totally would’ve participated in the Rwandan Genocide.”
These people display so much eagerness to participate in condoned violence.
@freganmitts The study is purely observational with no controls. In other words they just found an association, not necessarily proof that sweeteners cause anything. Many nutrition papers are problematic for the same reason.