🚨 NSF is already quietly eliminating the SBE Directorate, despite Congress’ mandate that NSF support the behavioral & social sciences.
Steps to counter this are in motion.
If you
- have an SBE proposal under review
- serve on an SBE grant panel
You can help! Fill out this form: shorturl.at/xuKw2
I don't know what this means but I like it
Pride and Prejudice as if Jane Austen were a mech interp researcher
The Philosophy program at the CUNY GC is hiring a logician at the Associate level. Please share! cuny.jobs/new-york-ny/...
You get used to it, but it was a shock to the system at first.
“I end each day exhausted—not from the work itself, but from the managing of the work. Six worktrees open, four half-written features, two ‘quick fixes’ that spawned rabbit holes, and a growing sense that I’m losing the plot entirely.”
I'm not going for moral or legal status here, just a point about needing to take their flexibility more seriously. RLHF and RLVR make them more like AlphaGo than BERT or GPT-2 in this regard. If they're tools, they're fundamentally unlike any tools we've ever had before.
I'm interested in understanding this too, I just worry that oversimplifying the LLM to meer tool or mimicry or sock puppet underestimates the societal challenge we're facing. They're neither full persons/agents nor meer tools. They're something awkwardly and uncannily in-between.
Yes as you might have guessed I'm working on a paper about CoTs in LRMs. I'll let you know when it's ready to share. Soon hopefully!
I share the intuition that something critical about sociality is still missing but it's harder to put one's finger on than this.
I guess the consensus now is that the language moves in the CoTs were all available to the base model but learning to sequence them efficiently and effectively was discovered by the model using an external answer checker.
The situation is much harder to understand than that...LRMs in particular (which are the ones producing interesting math results) have learned to produce text for themselves through extensive reinforcement learning on their own. It's doing all kinds of stuff semi-autonomously without human fingers.
This same kind of debate (over the use of understanding and computer simulation) has played out in philosophy of science over the last few decades with similar positions. There's a lot here that's interesting but the sock puppet analogy falls increasingly flat.
Yes, updated versions coming soon.
New greatest bird photo of all time just dropped
Royal Penguin
📷 JJ Harrison, eBird ebird.org/checklist/S2...
Bueller?
So, we don't know proprietary architectures etc., but is it probably right that the only causal influence CoT/reasoning trace tokens can have on final answer given by an LRM is through being processed as part of the context through self-attention, like other parts of the prompt?
The usual stellar science reporting here
Letterboxd screenshot Star Trek Into Darkness (2013) Watched by Branson Reese 28 Apr 2022 Abrams isn't a man so much as he's an avatar of cultural entropy. If you make something popular enough eventually you will die or sell it off and and a person in thick rimmed glasses whose main artistic vision is that he loves to have meetings will take it and sandblast it until it's nothing. I hate to see this happen here but maybe it's good that this happened. Maybe every franchise should collapse into a version of itself that makes shareholders nod and shake each others' hands. We should learn to speak in a language that rich people who don't dream can't comprehend.
i often think about this review
Do you remember when people were saying "I think my phone is listening to me, i'm getting ads for stuff I spoke about" and people said "They arent listening to you, they're just this good at predicting your interests"?
Well, it happens they were doing exactly that.
www.cbsnews.com/news/google-...
I will keep saying it: It turns out that inventing a tool that counterfeits the experience of being human can have, or rather is having, adverse consequences. Shirky concludes with the same point I’ve been making, that we cannot wait decades to get ahead of this. We are warping a generation.
It's worth having a look at our website, if you haven't, or haven't lately. There are a lot of remarkable things going on at UF philosophy these days.
phil.ufl.edu
• Petra Schumacher on real-time meaning adaptation
• Emma Borg questioning reasons & action
• @noranewcombe.bsky.social on cognitive maps
• @cameronbuckner.bsky.social on LLMs as models of human reasoning.
Videos were subtitled.
All links here: espp2025.ifispan.edu.pl/keynotes 2/2
Is this about the RAM shortage
Yes. This is another reason why public funding is what we want - why the government should support science with democratic mechanisms of oversight.
Public support means not having to beg toxic billionaires for money.
the professors in the emails being all chummy with epstein is pathetic and enraging.
