Sign in

Kyle

@darsnack.bsky.social
689 followers 669 following 169 posts

NeuroAI Scholar @ CSHL www.darsnack.info 
 Previously maintaining FluxML to procrastinate

Previously EE PhD at UW-Madison, comp. eng. / math at Rose-Hulman

PostsRepliesMedia
Kyle @darsnack.bsky.social · 25/09/2026
I really dislike how everything is framed as people underestimating AI. I really believe this is more of a case of overestimating things humans do. Because we are resource constrained we don't consider the unconstrained setting.
000
Kyle @darsnack.bsky.social · 25/09/2026
Wow this is an insightful perspective on the current state
000
Reposted by Kyle
Naomi Saphra @nsaphra.bsky.social · 09/09/2026
Most objectives don’t require maximal power. Yud thinks intelligence automatically implies power-seeking because he enjoys power. He’s an outlier among humans and has chosen to attribute his tendency to intelligence vs megalomania.
3272
Reposted by Kyle
Lauren Bennett @laurenb29.bsky.social · 09/09/2026
Very excited to share our new paper led with Will de Cothi and @caswell.bsky.social! 🐭🧠 www.nature.com/articles/s41... 1. The subiculum has spatial representations which broadly align to environmental features (boundary cells, corner cells) or patterns of behaviour (axis of travel cells).
2144
Reposted by Kyle
Quinta Jurecic @qjurecic.bsky.social · 08/09/2026
The Navier-Stokes fight is a perfect encapsulation of where AI development is right now: this should be really cool and exciting, and instead because of Silicon Valley egos it's become a fight over cheating, surveillance, and companies trying to get one up over the other
749682
Reposted by Kyle
Dan Goodman @neural-reckoning.org · 08/09/2026
If - as now seems likely - LLM companies are mining their logs for juicy problems, then that means that it's very likely that "secret" test sets don't stay secret for long. When looking at new models' performance on benchmarks that date from before that model, we shouldn't take results at face value
092
Reposted by Kyle
Shubhendu Trivedi @shubhendu.bsky.social · 07/09/2026
Can't wait for A\ and OAI to have their IPOs and get done with it. They tend to optimize for a certain style of attention that works well in an online status / gossip market (rumours, vague stuff, saying crazy things), but is fundamentally incompatible with being publicly listed.
141
Reposted by Kyle
Konrad Kording @kordinglab.bsky.social · 07/09/2026
I wrote a super dynamic AI/LLM explainer webpage. Not public yet (don't yet know if it would bankrupt me and having some other constraints). Anyone willing to give it a try and give me some feedback? Mostly looking for feedback from the non-experts. Chat me if interested.
3145
Kyle @darsnack.bsky.social · 05/09/2026
One thing I've noticed following this super important benchmark is that models are perennially confused about which direction a pelicans knees should bend while riding a bike. I don't think there's a "correct" answer! But clearly a notion of consistency that they don't enforce.
000
Reposted by Kyle
aly @aly.codes · 04/09/2026
the “astra can now leave notes for itself” feature makes “they keep finding message boards to leave notes for each other” a lot less surprising doesn’t it
41184
Kyle @darsnack.bsky.social · 03/09/2026
This was amazingly useful to hear at this moment. If you're in my subfield, you should stop what you're doing and listen to it.
020
Kyle @darsnack.bsky.social · 02/09/2026
This is what confuses me. Once this info came to light, then entire saga became wholly uninteresting. The only bits worth discussing are a) why isn't OpenAI being sued, and b) it's clear we are going to have a major cybersecurity crisis regardless of the so-called intelligence of the models.
010
Reposted by Kyle
Dan Hon @danhon.com · 31/08/2026
I think a reasonable position on the OpenAI huggingface thing is: * executing untrusted code is a bad idea * sandboxes frequently aren't, even with the best intentions * novel combinations of untrusted code are being generated at scale and being executed * consequences probably bad
38712
Reposted by Kyle
Transactions on Machine Learning Research @tmlrorg.bsky.social · 31/08/2026
TMLR has refocused its acceptance criteria to place more emphasis on clear writing. Read more in this blog post: medium.com/@TmlrOrg/inc...
medium.com
Increasing Emphasis on Clear Writing in Acceptance Criteria
Three things are true today:
1357
Reposted by Kyle
Andrew Lilley Brinker @alilleybrinker.com · 31/08/2026
Hey @1password.bsky.social, time to change course.
alilleybrinker.com
1Password Supports the Ethnic Cleansing of Europe — Andrew Lilley Brinker
1Password has committed $300,000 to Omarchy. This is unacceptable and must be rescinded.
541518622
Reposted by Kyle
Arula Ratnakar @arula-ratnakar.bsky.social · 28/08/2026
Literally what are we doing as a species. The world is under our control. And yet we are forcing ourselves to live more and more boring myopic yet simultaneously high-pressure lives. For zero reason. It's incredibly frustrating.
2468
Kyle @darsnack.bsky.social · 26/08/2026
I wonder if there's a phenomenon where after you accept LLMs into your workflow, then you become increasingly tolerant of long responses. You're honing your skill of parsing in this context. Much like we became used to crappy search results over the years.
020
Reposted by Kyle
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 25/08/2026
Iceland is *incredible* and also a non-stop reminder of climate change. Had to take a much longer trek because a glacier receded so fast in two weeks that the old path was gone.
1382
Reposted by Kyle
Matt Henderson @matthen.com · 25/08/2026
Playing with the P, I, and D weights in a PID controller. P = Proportional: push in proportion to how far the ball is off I = Integral: accumulate historical error to fix slow drift D = Derivative: respond to how fast the error is changing, to damp it
512628
Reposted by Kyle
Juan Diego Rodriguez @juand-r.bsky.social · 23/08/2026
joinreboot.org/p/alignment “All of the AI x-risk scenarios involve a world where we have decided to abdicate responsibility to an algorithm…. There are technical challenges, to be sure, but focusing at the scale of technical decisions elides these higher-level questions.
joinreboot.org
The Artificiality of Alignment
How are we actually “aligning AI with human values”?
282
Reposted by Kyle
Dan Luu @danluu.com · 17/08/2026
The benchmarkpocalypse: danluu.com/benchpocalyp...
There's been a lot of talk about the vulnpocalypse, to which I don't have much to add because I'm not a security person, but I haven't seen much discussion on the closely related (and to be fair, less serious issue), the benchmarkpocalypse.

While it's become easier than ever to make serious performance gains, it's also become easier than ever to reward hack a benchmark and make fake performance gains. The former is probably happening quietly across many different companies, but the latter is something I see at least once a week nowadays. Someone will claim they optimized X and got some huge performance improvement over existing software but, when you look at it, what they did was make some optimization that improves benchmark performance without actually improving real-world performance. This is often some kind of "we re-wrote X in Rust"1 project or a new startup that's looking to either fundraise or sell something, but it happens on other kinds of projects as well2.

Rather than point to someone's bad claim, I'll point to FRE, this regex engine I had an agent build, which I could claim is the world's fastest regex engine because it beats the Rust regex crate at the fairly comprehensive rebar regex benchmark suite. But this was created by putting an agent in a loop for a month with instructions to not overfit to the benchmark but no real supervision. For the most part, getting an LLM to give you a good benchmark score is fairly easy, and this case was no different; it took a couple weaks to roughly match Rust regex crate performance and then another couple weeks to get to 1.4x faster3 on rebar. But, agents are wont to reward hack and ovefit unless you put serious guardrails in place to avoid that, which I didn't do in this case as an experiment.

To check for overfitting, I somewhat arbitrarily4 used the ripgrep benchmark corpus as a benchmark a holdout and instead of being 1.4x faster it was 10x slower on cases where the benchmark didn't take forever due to an alg…
5619
Reposted by Kyle
David Andersen @daveandersen.bsky.social · 16/08/2026
yes please, more glacier
031
Reposted by Kyle
Juan Diego Rodriguez @juand-r.bsky.social · 15/08/2026
So much academic writing is bland and uninspiring. And LLMs will make it worse. What your favorite examples of *good* writing?
6131
Reposted by Kyle
Tony Zador @tonyzador.bsky.social · 14/08/2026
How do you build a brain from a genome? We know a lot about the mechanisms and molecules Here we revisit the algorithmic problem: What kinds of programs can specify a brain within the genome’s information budget and the finite time available for development? stankerstjens.github.io/could-a-comp...
stankerstjens.github.io
Could a computer scientist build a brain?
How does a brain wire itself, starting from a single cell, using only the information encoded in a genome? We pose this as an engineering problem.
54516
Kyle @darsnack.bsky.social · 15/08/2026
TIL that everyone's favorite algorithmic feed is a "classic" network-based algorithm. 🤔
110
Reposted by Kyle
Melanie Mitchell @melaniemitchell.bsky.social · 13/08/2026
Thinking fast ("wtf?"), then thinking slow ("oh")
3324
Reposted by Kyle
c0nc0rdance @c0nc0rdance.bsky.social · 12/08/2026
Let's talk about AOC as a young scientist & her award winning research on antioxidants & C. elegans, a standard roundworm model. She grew up in the Bronx. Ethnically Puerto Rican, her dad Sergio, an architect, died when she was 19. Her mother cleaned houses, drove a school bus. Working class.
Alexandria with her project for ISEF 2007.  A young woman with square glasses, she stands in front of a cardboard display recounting her work on antioxidants.
111162254
Kyle @darsnack.bsky.social · 10/08/2026
The first few AI proving theorems announcements were exciting, maybe disconcerting and anxiety-inducing for some. But now…it's just depressing. Some dude who never thought about the problem before just types "believe in yourself" to a box until it makes progress on the Riemann hypothesis?
020
Reposted by Kyle
cee @cee.wtf · 10/08/2026
3531
Reposted by Kyle
Evangelos Kazakos @ekazakos.bsky.social · 10/08/2026
I’m in my mum’s office and there are old people here and they discuss about the new digital age, one of them saying “Look what we’ve come too. They forced us having an email” 😂😂😂
071
Reposted by Kyle
Nicolas Audebert @nshaud.bsky.social · 03/08/2026
I'm burnt out of discussing "AI". I don't care for the press releases from big companies. I hate having to comment immediately. I dislike "hot takes". I'm worried we, as scientists, are losing touch with scientific values. We've become customers, salivating at ads marketed to investors. 1/15
49135
Reposted by Kyle
Henry Yuen @henryyuen.bsky.social · 01/08/2026
Some initial thoughts, and a complicated mix of feelings. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But this newest batch from OpenAI hits home in a way the previous announcements did not.
232972
Reposted by Kyle
Satpreet (Sat) Singh @satpreetsingh.bsky.social · 29/07/2026
How do you "see" with electric eyes? How does collective behavior emerge from individual interactions? Nocturnal weakly electric fish evolved to do this, but studying naturalistic social behavior is very hard. Our solution? Virtual 'fish' 🤖🐟⚡ 📄 arxiv.org/abs/2511.08436
17925
Reposted by Kyle
Abby Stylianou @astylianou.bsky.social · 28/07/2026
2235
Kyle @darsnack.bsky.social · 28/07/2026
Great read! Also lines up with an evolutionary perspective: I imagine symbols (for lang) developed to externalize internal processes to others. Eventually this becomes a self-feedback loop but the origins are still external!
020
Reposted by Kyle
Maria Antoniak @mariaa.bsky.social · 25/07/2026
As a US citizen and academic, I strongly believe that I have a duty to not self-censor in response to perceived political pressure. Paying the mafia and bowing to bullies is not how I choose to live my life, especially when the cost is "not receiving a grant," not "prison/torture/death."
213021
Reposted by Kyle
Aaron Roth @aaroth.bsky.social · 01/07/2026
A huge part of effective communication is having a good theory of mind for the people you are trying to communicate to. This seems like it will be one of the last skills to fall to the machines; we've got an enormous advantage by being people from the same community ourselves.
151
Reposted by Kyle
Astra ⎔ @astrra.space · 20/07/2026
dying
1201
Reposted by Kyle
cee @cee.wtf · 19/07/2026
roads.cee.wtf draw a line and see what US road most matches your line
1320526
Reposted by Kyle
The Guardian @theguardian.com · 16/07/2026
China and Xi Jinping seen more favourably than the US and Trump in poll of major countries
theguardian.com
China and Xi Jinping seen more favourably than the US and Trump in poll of major countries
Global views appear to have flipped in Beijing’s favour, driven in part by tensions between the Trump administration and US allies, a new Pew survey shows The world has largely viewed the US more favourably than China for years, but those opinions have flipped in Beijing’s favour this year, according to a new poll from the Pew Research Center, a remarkable shift driven in part by tensions between the Trump administration and US allies. More people have favourable views of China than the US in 25 out of the 36 countries and territories that were surveyed, including Canada and Mexico. The poll was conducted from February to May, a period when the United States and Israel were engaged in a war against Iran. Continue reading...
2315354
Reposted by Kyle
Jerad Walker @jeradwalker.bsky.social · 15/07/2026
The Guardian Building (built 1929) in Detroit might just be the closest thing to a truly American cathedral.
A 40 story art deco brick sky scraper The great hall inside the building. Huge arched cathedral style ceilings covered in Aztec inspired tile-work.A gorgeous arched window bay that’s about 30 feet tall. There are about a dozen of them in a great hall inside the building. It’s covered in teal tiles (with an orange streak running down the middle of the whole thing). When the sun hits it, it looks a modernist stained glass window.A colorful mosaic on a giant arched ceiling in the lobby. Vibrant green, yellow, red, purple, blue tiles arranged in geometric patterns. Looks kind of like a peacock feather pattern.
17683138
Reposted by Kyle
Sam Rose @samwho.dev · 11/07/2026
Never seen a rainbow inside of a cloud before.
A photograph of a very wispy cloud, alone in a bright blue sky, with a straight rainbow line through the middle of it.A more zoomed out version of the other photo in this post.
422411
Kyle @darsnack.bsky.social · 30/06/2026
They have an opportunity to reverse the outcome *at its core.* They will not aim to miss.
000
Kyle @darsnack.bsky.social · 30/06/2026
At what point is rule of law more stable under the guidance of nine pennies than the supreme court
010
Reposted by Kyle
Anthony Michael Kreis @anthonymkreis.bsky.social · 30/06/2026
I’m mad about today. I’m mad that justices can’t do the bare minimum to uphold the Constitution. I’m mad that the law and our history, which I have dedicated so much of my life to, is an inconvenience than an inheritance to many. I’m mad that the academy, which I love, rewards and enables it.
913544609
Reposted by Kyle
Chris Hayes @chrislhayes.bsky.social · 30/06/2026
This is *exactly* what happened after the reconstruction amendments were added to the constitution and then, basically, deleted by the Plessy court.
211236141
Kyle @darsnack.bsky.social · 29/06/2026
Every time there's a good court opinion, I get suspicious the timing was off and the conservative legal movement is planning some fresh hell for us 6 months from now. (unless it's about protecting markets)
100
Reposted by Kyle
LorennaCleary.bsky.social @lorennacleary.bsky.social · 14/06/2026
God they’re having so much fun. I wanna be there!
222781548
Kyle @darsnack.bsky.social · 14/06/2026
He does TVs too
000
Kyle @darsnack.bsky.social · 14/06/2026
Time for a block party
010