Naomi Saphra
@nsaphra.bsky.social
Waiting on a robot body. All opinions are universal and held by both employers and family.
Literally a professor. Recruiting students to start my lab.
ML/NLP/they/she.
Literally a professor. Recruiting students to start my lab.
ML/NLP/they/she.
party's full, go home
November 11, 2025 at 9:05 PM
party's full, go home
meet me at this button friends
November 11, 2025 at 9:01 PM
meet me at this button friends
My gf and I are binging the hell out of this extremely well-read audiobook series. I suspect I wouldn't have enjoyed them as print books nearly as much but we've gone through like 100 hours in a month. GF's not on bsky so I know you guys won't snitch that I got her Princess Donut merch for her bday.
November 8, 2025 at 5:19 PM
My gf and I are binging the hell out of this extremely well-read audiobook series. I suspect I wouldn't have enjoyed them as print books nearly as much but we've gone through like 100 hours in a month. GF's not on bsky so I know you guys won't snitch that I got her Princess Donut merch for her bday.
My gf and I started listening to goofy audiobooks together and this one is super fun. Written though surprising NOT read by Kate McKinnon, but by her sister?
November 8, 2025 at 5:15 PM
My gf and I started listening to goofy audiobooks together and this one is super fun. Written though surprising NOT read by Kate McKinnon, but by her sister?
the only kind of Rat Race I'm down for
November 6, 2025 at 2:43 PM
the only kind of Rat Race I'm down for
One time I was at a party and met this person which was shocking
October 29, 2025 at 9:12 PM
One time I was at a party and met this person which was shocking
If Boston can require cars to pass an annual inspection they can require motorists to pass an annual one question exam where you just have to explain what this sign means
October 9, 2025 at 3:29 PM
If Boston can require cars to pass an annual inspection they can require motorists to pass an annual one question exam where you just have to explain what this sign means
My friends and family know I love looking at random variation. But my ENEMIES claim random seed will be UNIMPORTANT at large scales. Fools! I will say YOU ARE WRONG, MY ENEMIES. LOOK.
October 3, 2025 at 8:07 PM
My friends and family know I love looking at random variation. But my ENEMIES claim random seed will be UNIMPORTANT at large scales. Fools! I will say YOU ARE WRONG, MY ENEMIES. LOOK.
I had heard MBTI was invented by a “housewife/daughter team” but like, she literally was publishing articles about personality? Her daughter consulted for ETS? Those aren't housewives! Yeah the science was shaky but was ANYONE doing sound psych research in the 50s?
September 28, 2025 at 6:19 PM
I had heard MBTI was invented by a “housewife/daughter team” but like, she literally was publishing articles about personality? Her daughter consulted for ETS? Those aren't housewives! Yeah the science was shaky but was ANYONE doing sound psych research in the 50s?
So I knew Jeanette Winterson as “lesbian memoirist and novelist” and the only novel I read by her, Gut Symmetries, was frankly too ~literary~ for me. If you knew more OR less than me about her, you might be unsurprised that she is this type of e/acc transhumanist!
September 25, 2025 at 3:08 PM
So I knew Jeanette Winterson as “lesbian memoirist and novelist” and the only novel I read by her, Gut Symmetries, was frankly too ~literary~ for me. If you knew more OR less than me about her, you might be unsurprised that she is this type of e/acc transhumanist!
Every time I go to SF I get jumpscared by going to a random party and then finding out I’m at the like Mechanistic Interpretability Acid House Rave Sponsored By Sequoia
September 17, 2025 at 10:02 PM
Every time I go to SF I get jumpscared by going to a random party and then finding out I’m at the like Mechanistic Interpretability Acid House Rave Sponsored By Sequoia
This is my little opening exercise to ramp up, I think the rest of the presentation will keep changing before I'm ready to post the slides
cc @tdietterich.bsky.social
cc @tdietterich.bsky.social
September 11, 2025 at 4:05 PM
This is my little opening exercise to ramp up, I think the rest of the presentation will keep changing before I'm ready to post the slides
cc @tdietterich.bsky.social
cc @tdietterich.bsky.social
I thought I wouldn‘t be one of those academics super into outreach talks, but I just put together something about understanding LLMs for laypeople and I get to talk about results that I don’t really focus on in any of my technical talks! It’s actually really cool. I made this lil takeaway slide
September 10, 2025 at 1:21 PM
I thought I wouldn‘t be one of those academics super into outreach talks, but I just put together something about understanding LLMs for laypeople and I get to talk about results that I don’t really focus on in any of my technical talks! It’s actually really cool. I made this lil takeaway slide
Can’t believe when I was in high school they let me read this woke propaganda
August 31, 2025 at 11:44 PM
Can’t believe when I was in high school they let me read this woke propaganda
It’s cool that Massachusetts is getting a better flag but disappointing that we weren’t brave enough to select this entry from the competition
August 30, 2025 at 1:50 PM
It’s cool that Massachusetts is getting a better flag but disappointing that we weren’t brave enough to select this entry from the competition
In skill generalization, models can generalize to questions that no individual expert can correctly answer. Models can answer two-hop queries by composing knowledge from different experts—if diverse phrasing forces that knowledge into a latent semantic space.
August 29, 2025 at 9:46 PM
In skill generalization, models can generalize to questions that no individual expert can correctly answer. Models can answer two-hop queries by composing knowledge from different experts—if diverse phrasing forces that knowledge into a latent semantic space.
In skill selection, models can transcend even if errors are correlated. Given diverse experts describing their specialized expertise—rather than expounding on misunderstandings—the model will reflect the correct answers, avoiding common misconceptions. If only!
August 29, 2025 at 9:46 PM
In skill selection, models can transcend even if errors are correlated. Given diverse experts describing their specialized expertise—rather than expounding on misunderstandings—the model will reflect the correct answers, avoiding common misconceptions. If only!
Skill denoising, the type of transcendence discussed in previous work, is the “wisdom of the crowds” strategy. A model can outperform any one source through low temperature sampling—as long as the experts produce diverse uncorrelated errors.
August 29, 2025 at 9:46 PM
Skill denoising, the type of transcendence discussed in previous work, is the “wisdom of the crowds” strategy. A model can outperform any one source through low temperature sampling—as long as the experts produce diverse uncorrelated errors.
We use a fictional knowledge graph and assign knowledge across different simulated “experts” which each describe their knowledge. When does a language model know better than these experts? When the data is diverse!
August 29, 2025 at 9:46 PM
We use a fictional knowledge graph and assign knowledge across different simulated “experts” which each describe their knowledge. When does a language model know better than these experts? When the data is diverse!
In a knowledge-graph based, a model can transcend its sources 3 ways:
In skill denoising, the model averages out each expert's errors.
In skill selection, the model picks the right expert for each case.
In skill generalization, the model composes knowledge across experts.
In skill denoising, the model averages out each expert's errors.
In skill selection, the model picks the right expert for each case.
In skill generalization, the model composes knowledge across experts.
August 29, 2025 at 9:46 PM
In a knowledge-graph based, a model can transcend its sources 3 ways:
In skill denoising, the model averages out each expert's errors.
In skill selection, the model picks the right expert for each case.
In skill generalization, the model composes knowledge across experts.
In skill denoising, the model averages out each expert's errors.
In skill selection, the model picks the right expert for each case.
In skill generalization, the model composes knowledge across experts.
How can an imitative model like an LLM outperform the experts it is trained on? Our new COLM paper outlines three types of transcendence and shows that each one relies on a different aspect of data diversity. arxiv.org/abs/2508.17669
August 29, 2025 at 9:46 PM
How can an imitative model like an LLM outperform the experts it is trained on? Our new COLM paper outlines three types of transcendence and shows that each one relies on a different aspect of data diversity. arxiv.org/abs/2508.17669
The key result is that LLM guardrails will refuse different requests from "conservative" simulated users than from "liberal" ones, and it also infers those politics from demographic information---or even a cue as subtle as endorsing an NFL team.
August 28, 2025 at 1:53 PM
The key result is that LLM guardrails will refuse different requests from "conservative" simulated users than from "liberal" ones, and it also infers those politics from demographic information---or even a cue as subtle as endorsing an NFL team.
second kitten reveal (feral socialization project)
August 27, 2025 at 9:18 PM
second kitten reveal (feral socialization project)
