Naomi Saphra

Orson Scott Card. Especially later in the Ender series his religious themes are really forefronted.

The lack of differentiation between trusted and untrusted state and input is endlessly devilish in trying to secure this technology. It completely violates all security paradigms we've tried to enforce for coming on half a century

Stolen Probability: A Structural Weakness of Neural Language Models Neural Network Language Models (NNLMs) generate probability distributions by applying a softmax function to a distance metric formed by taking the dot product of a prediction vector with all word vect...

ACL shortpaper appreciation thread because shortpapers are the best papers. What's the best <=5 pager? I'm nominating an oldie but a goodie

at least three people who work on LLM internals professionally and a number of extremely devoted amateurs follow will stancil and i think we will unanimously tell you that we have only the vaguest idea how any of their internal operations function

I don’t blame anyone for caricaturing her views. In media, she seems to caricature her own views these days. But why would that caricature have been respected (elected president!) in a community where publications have been dominated by neural (conditional) language models for about a decade?

You can disagree with her philosophical objections and thought experiments (I do too), and you can disagree with her overall attitude in media interviews, but it’s silly to conflate her with Chomsky’s total denial of computational and corpus linguistics.

This is just not correct. EB has been dismissive of natural language “understanding” from a philosophical angle, but she always opposed most formal syntacticians in her endorsement of statistical/computational approaches to language. That’s WHY she was in the NLP community at all.

A line chart titled “Quality of solutions by some hypothetical metric where human experts are 1.0.” The vertical axis ranges from 0.00 to slightly above 1.00, and the horizontal axis runs from 2015 to 2035. Two sigmoid-shaped curves are shown: a blue “Closed models” line that rises steeply from near zero around 2021 to level off around 2028, and a red “Open models” line that starts similarly low but rises about two years later, leveling off around 2030. Both lines continue to rise rather than leveling entirely flat. The legend is in the lower right corner inside the plot area.

This is the world I hope we're getting: One where 1) progress is loosely sigmoid, not because there's any fixed cap, but because the initial overhang of human training data has been exploited, and 2) open models can come close to catching up, because of #1.

The White House is taking new steps to undermine every international student at our universities.

It has approved a new rule restricting student visas to 4 years max. That's not long enough for PhD students, or about half of undergrads, to finish their degree.

Likely to enter force within days.

You're all gonna look stupid as hell when in 10 years a computational linguist proves that blueberry does in fact have 3 r's

chatgpt 5 bring wife back. How bring wife back

HITLER
CONCERNS

what are you people doing dude

Representation biases: will we achieve complete understanding by analyzing representations? A common approach in neuroscience is to study neural representations as a means to understand a system -- increasingly, by relating the neural representations to the internal representations learned b...

Many representational analyses (implicitly) prioritize signals by the amount of variance they explain in the representations. However, in arxiv.org/abs/2507.22216 we discuss results from our prior work that challenge this assumption; variance != computational importance.

Here's the letter that NSF sent UCLA last week about cutting its research funding

NSF says UCLA is doing affirmative action by asking applicants' their zip codes, family income and what HS they went to, and that they can discuss race in personal statements

storage.courtlistener.com/recap/gov.us...

The British empire colonized many of the best cooks in the world

My most inflammatory opinion is that literally every continental European city is culinarily mid and monotonous. Only the UK has any world-class food cities.

We have been carefully testing lots of received prompting wisdom & for recent AI models:
🚫Threats, saying please, being insulting, & promising tips do not change average performance on challenging tasks
⛓️Chain-of-thought no longer helps even non-reasoner performance much

Don't make it complicated.

Unclear.

Too polite

If you have a disability strangers should just drop by sometimes to adjust your furniture after you move just in case

Ugh anyone in Boston wanna come hang this art up in my apartment

#arXiv Can Machines Philosophize? arxiv.org/abs/2507.00675 “debate on scientific realism: survey of over 500 human participants, including both physicists and philosophers of science. Machine personas using an AI engine based on a LLM generative model.” 1/2

Hierarchical Reasoning Model Reasoning, the process of devising and executing complex goal-oriented action sequences, remains a critical challenge in AI. Current large language models (LLMs) primarily employ Chain-of-Thought (CoT...

haven't seen this posted here, looks like maybe a big deal, crushes a bunch of stuff. arxiv.org/abs/2506.21734

A fun reminder that a loss of 60,000 jobs in and around academia is ~50% more than all coal mining jobs in the entire country.

Bitch: On the Female of the Species Lucy Cooke

Primatologists describe "subordination hierarchy" in contrast to dominance hierarchy---one shaped by respect / affection rather than fear, like the trust that an orca matriarch will best lead a pod to salmon.

Martyr! Kaveh Akbar

At one point a character's experience is described as not being "legible to empire", and that's such a real way to describe it. A huge creature whose umwelt is limited by its senses.

Supercommunicators: How to Unlock the Secret Language of Connection Charles Duhigg

Some political scientists got a bunch of people with different gun control positions to all hang out in person with relaxing guidelines and they were able to have normal conversations because they became friends. This is unfortunately not a scalable model.

Terence Tao (@teorth.bsky.social) has written a thread on Mastodon about the impact of the federal grant freeze to UCLA, particularly to his own field of Mathematics. UCLA's IPAM (Institute of Pure and Applied Mathematics) could shut down entirely

mathstodon.xyz/@tao/1149568...

I underestimated Twitter. I did not expect anyone to post that Terence Tao is stealing math research jobs from hard working Americans.

I’m not talking about organizations and institutions. I’m talking about literally knowing other Jews and being in a community with them.