Sign in

David Marx

@digthatdata.bsky.social
2.5K followers 2K following 12K posts

I read a lot of research. Mostly ML. Currently reading: dmarx.github.io/papers-feed Statistical Learning Information Theory Ontic Structural Realism Morality As Cooperation Free Culture, Open Access YIMBY, UBI Research MLE Frmr FireFighter

PostsRepliesMedia
Reposted by David Marx
A. Feder Cooper @afedercooper.bsky.social · 4h
Our large-scale study of memorization of books in open-weight LLMs (e.g., Llama, Qwen) will appear at the 2026 Conference on Language Modeling as an oral. We made a website for exploring our results on 200 books and 14 models: books-memorization.github.io
books-memorization.github.io
How much do open-weight LLMs memorize specific books?
Open-weight LLMs memorize books far more than previously believed. Memorization varies by model family, model size, and book. In extreme cases, entire books are memorized, and we can generate them eff...
24624
Reposted by David Marx
Jamie Gangel @jamiegangel.bsky.social · 29/09/2026
Breaking - The Supreme Court on Tuesday allowed the Trump administration to restart its program of removing migrants to countries other than their homeland without giving them an opportunity to raise human rights concerns… will hear arguments later this year www.cnn.com/2026/09/29/p...
cnn.com
Supreme Court allows Trump to restart controversial deportations to countries they have no connection to | CNN Politics
The Supreme Court on Tuesday allowed the Trump administration to restart its program of removing migrants to countries other than their homeland without giving them an opportunity to raise human right...
3217287
Reposted by David Marx
ELLIS @ellis.eu · 15h
🏹 Job alert: PhD Position in Mathematical Foundations of Deep Learning at TU Darmstadt 📍 Darmstadt 🇩🇪 ⏰ 22 October 🔗 bit.ly/4dzSuaG
bit.ly
www.career.tu-darmstadt.de
041
David Marx @digthatdata.bsky.social · 20h
old: turn it off and back on again now: /clear
000
David Marx @digthatdata.bsky.social · 29/09/2026
hmm
011
David Marx @digthatdata.bsky.social · 29/09/2026
symbolic regression + boosting = interpretable decompositions of chaotic systems
020
Reposted by David Marx
Maria Antoniak @mariaa.bsky.social · 29/09/2026
I'm recruiting 1-2 PhD students to join our lab in Fall 2027, through either Computer Science or Information Science at the University of Colorado Boulder! Looking for people with interests in NLP plus [healthcare | literary studies | narratives | social media | etc.]. Join us! 🏔️☀️
cls-lab.com
CLS Lab — Culture, Language, & Systems | University of Colorado Boulder
The Culture, Language, & Systems Lab at CU Boulder studies the language systems that transmit and shape modern culture, from internet platforms to literary archives to language models.
06950
Reposted by David Marx
Adam Klasfeld @klasfeldreports.com · 29/09/2026
Public Citizen files a complaint opposing the broadcasting of Trump’s ads as violating laws against taxpayer-funded propaganda. Full complaint buff.ly/kYCsKEW
Public Citizen urges the Federal Communications Commission (FCC) and the Federal Trade Commission (FTC), and broadcasters individually, to end airing recent campaign commercials by President Trump and the White House that appear to be in breach of federal law. At least three campaign ads, paid for by taxpayer dollars, appear to violate the federal law against using taxpayer funds to finance propaganda (5 U.S.C. §3107) and related provisions of the appropriations laws as well as the Hatch Act (5 U.S.C. §7324).
5323114
David Marx @digthatdata.bsky.social · 29/09/2026
this is pure schadenfruede
010
Reposted by David Marx
Chris Geidner @chrisgeidner.bsky.social · 29/09/2026
NOW: DC Circuit en banc aguments over Judge Boasberg’s contempt proceedings have now begun at the DC Circuit. Brett Shumate is now arguing for DOJ. www.youtube.com/live/4A1BwFh...
youtube.com
YouTube
Share your videos with friends, family, and the world
22431108
Reposted by David Marx
Alejandra Caraballo @esqueer.net · 27/09/2026
Google finally updated some of its satellite images of Gaza and my god, it's horrific. The entirety of Rafah has been demolished. The majority of Khan Younis is in ruins. Entire towns completely leveled and erased of the map.
9352063189
David Marx @digthatdata.bsky.social · 29/09/2026
israel may as well have nuked gaza given the level of destruction rendered.
140
David Marx @digthatdata.bsky.social · 29/09/2026
I did it. I finally feel like I mostly understand Magnolia.
020
David Marx @digthatdata.bsky.social · 29/09/2026
trying to grok the film "Magnolia"
110
Reposted by David Marx
Bloomberg News @bloomberg.com · 28/09/2026
Chipmaker AMD agreed to acquire World Labs for $8.2 billion, gaining an artificial intelligence startup founded by industry pioneer and researcher Fei-Fei Li.
bloom.bg
AMD to Buy Fei-Fei Li’s World Labs AI Startup for $8.2 Billion
Chipmaker Advanced Micro Devices Inc. agreed to acquire World Labs for $8.2 billion, gaining an artificial intelligence startup founded by industry pioneer and researcher Fei-Fei Li.
1289
David Marx @digthatdata.bsky.social · 28/09/2026
> Sauer is asking the Supreme Court to say that the federal government no longer believes that trans people are justified in being trans in either the medical or social ways that allow trans people to live their lives authentically.
010
Reposted by David Marx
Miles Grant @milesgrant.bsky.social · 28/09/2026
Failing to stop burning fossil fuels continues to destroy the only climate safety zone humanity has ever known
223362
Reposted by David Marx
James Medlock @jdcmedlock.bsky.social · 27/09/2026
The GOP cut funding to the IRS because *audits work* and they fundamentally don't believe rich people should have to play by the rules.
Line chart titled "Revenue per audit," subtitled "In-person audits, 2010-2014." It plots average revenue and average cost per audit by taxpayer income percentile, poorer to richer, from the 0–5th percentile to the top 0.1%. Source: Boning et al. 2023, "A Welfare Analysis of Tax Audits Across the Income Distribution."

Across the bottom half of the income distribution, revenue and cost track closely, both around $3,000–$6,000 per audit. Above the median, revenue pulls ahead of cost: about $10,000 vs. $5,000 near the 90th percentile, and about $17,000 vs. $8,000 at the 95th–99th. At the top, revenue jumps to about $31,000 for the 99–99.9th percentile and about $90,000 for the top 0.1%.
273296769
Reposted by David Marx
Cooper Lund @cooperlund.online · 28/09/2026
The GOP is an opposition party even when it's in power, the idea that they are the perpetual underdog is central to their internal mythology and can't be disturbed by facts or the results of elections. It's also why they're bad at, you know, governing.
181987358
Reposted by David Marx
ponder @ponder.ooo · 27/09/2026
so much of "alignment" rn consists of ignoring existing knowledge of how to responsibly build and manage automated systems and doing this instead:
420234
David Marx @digthatdata.bsky.social · 27/09/2026
wooooow
020
David Marx @digthatdata.bsky.social · 26/09/2026
1. yeah this is real bad. 2. the timing suggests that part of the motivation here was to figure out who the bestiality carve out was for, likely to use as blackmail kompromat. 3. this just belabors the fact that any history of bestiality should be disqualifying and it was good that it already was.
010
David Marx @digthatdata.bsky.social · 26/09/2026
epstein loved him, why can't you?
left: a birthday card trump sent to epstein in 2003 for his 50th, which unsubtly alludes to their shared affinity for sex with minors.

right: a photograph of trump, ivanka, epstein, and maxwell socializing at a party at trump's mar-a-lago club, February 12th, 2000.
010
Reposted by David Marx
ryan cooper @ryanlcooper.com · 26/09/2026
Trump fucked up the *American* renewable transition but he might have had the largest positive effect on the global transition of any president ever
381296253
Reposted by David Marx
Carolina Botelho @sobotelha.bsky.social · 26/09/2026
O terrivelmente evangélico me incluiu numa lista de 1600 perfis, junto com advogados, deputados e outros perfis, para pedir ao TSE que retire do ar postagens críticas ao candidato filho do ex-presidente preso por golpe de estado. Esse é o clima pré-eleitoral, imaginem depois, caso o candidato ganhe
33986267
Reposted by David Marx
leon @leyawn.bsky.social · 24/09/2026
ceo of burger king: our burgers have gone rogue again and infected people with e coli
5356151268
David Marx @digthatdata.bsky.social · 26/09/2026
this isn't a job for a prompt, this is a job for data classification, permissions, and environments. if you have a dataset you want to protect from the public, asking nicely in the readme is not a reliable or responsible way to achieve that outcome. the fuck is wrong with people.
232
David Marx @digthatdata.bsky.social · 25/09/2026
For anyone who hasn't caught any of the preliminary scientific discourse discussing this disaster, strongly recommend this video by science communicator Anton Petrov. (his whole channel is super high quality and he somehow manages to churn out great content like this daily).
youtube.com
What Really Happened During the 2026 Nepal Disaster?
YouTube video by Anton Petrov
020
Reposted by David Marx
vortex_egg @vortexegg.com · 25/09/2026
Incredible scenes from Meta. Apparently you can prompt inject their Muse agent into giving you a root shell on the box it's running on by roleplaying that you are actually the Muse agent itself.
neuromatch.social
jonny (nonvenomous) (@jonny@neuromatch.social)
Attached: 1 image Not only confirming this, its extremely easy. Almost no prompt injection resistance lmao
16373108
David Marx @digthatdata.bsky.social · 25/09/2026
wanted a quick reference for lymph node locations. tried searching google images expecting a plethora of useful results: instead, it's basically all AI slop with hallucinated anatomy. weak sauce. POP QUIZ: Guess which image was from a National Geographic webpage!
210
Reposted by David Marx
David Marx @digthatdata.bsky.social · 25/09/2026
nice visual illustration of how superposition of concepts works
Figure 1:Summarizing the SDM read and write operations. 

Top Row three patterns being written into nearby neurons. 1. The first write operation; 2. Patterns are stored inside nearby neurons and the original pattern location is shown; 3. Writing a second pattern; 4. Writing a third pattern and neurons storing a superposition of multiple patterns. 

Bottom Row shows two isomorphic perspectives of the read operation. Neuron view (left) shows the query reading from nearby neurons with the inset showing the number of times each pattern is read. The four blue patterns are a majority which would result in one step convergence. Pattern view (right) is crucial to relating SDM to Attention and defined in Eq. 1 below. 

We abstract away the neurons by assuming they are uniformly distributed through the space. This allows us to consider the circle intersection between the query and the original locations of each pattern where blue has the largest circle intersection.
0115
David Marx @digthatdata.bsky.social · 25/09/2026
nice visual illustration of how superposition of concepts works
Figure 1:Summarizing the SDM read and write operations. 

Top Row three patterns being written into nearby neurons. 1. The first write operation; 2. Patterns are stored inside nearby neurons and the original pattern location is shown; 3. Writing a second pattern; 4. Writing a third pattern and neurons storing a superposition of multiple patterns. 

Bottom Row shows two isomorphic perspectives of the read operation. Neuron view (left) shows the query reading from nearby neurons with the inset showing the number of times each pattern is read. The four blue patterns are a majority which would result in one step convergence. Pattern view (right) is crucial to relating SDM to Attention and defined in Eq. 1 below. 

We abstract away the neurons by assuming they are uniformly distributed through the space. This allows us to consider the circle intersection between the query and the original locations of each pattern where blue has the largest circle intersection.
0115
David Marx @digthatdata.bsky.social · 25/09/2026
Nice paper from 2011 I stumbled across today
arxiv.org
Attention Approximates Sparse Distributed Memory
While Attention has come to be an important mechanism in deep learning, there remains limited intuition for why it works so well. Here, we show that Transformer Attention can be closely related under ...
161
Reposted by David Marx
Democracy Docket @democracydocket.com · 24/09/2026
NEW: California's attorney general said Riverside County Sheriff Chad Bianco (R) will face "legal consequences" after the state Supreme Court determined his seizure of roughly 650k ballots violated state law. Bianco must now comply with the attorney general's directives and return the ballots.
democracydocket.com
California Attorney General vows ‘legal consequences’ for GOP sheriff who seized ballots
The ballot seizure case highlights the risks to democracy when election deniers in positions of authority abuse their power to interfere with vote counting. Bianco seized an estimated 650,000 ballots ...
42540197
Reposted by David Marx
Alex Wild @alexwild.bsky.social · 24/09/2026
Academic freedom continues to crumble at the hands of Republican Party political appointees at UT.
313065
David Marx @digthatdata.bsky.social · 23/09/2026
This was from a super over-cooked fine tune on images of myself. I love how it's sort of revealing of how the UNet builds my face as a composition of semantic layers.
020
David Marx @digthatdata.bsky.social · 23/09/2026
guess when I ran out of adderall
A heatmap laid over a calendar, showing "Papers per day" data, with counts falling off a cliff mid-late June.
031
David Marx @digthatdata.bsky.social · 23/09/2026
I haven't made art in a while. I should go make some art.
020
David Marx @digthatdata.bsky.social · 23/09/2026
hey remember back when people were all worried about "spontaneous human combustion" but then it turned out the problem was just that our stuff was all super flammable?
010
Reposted by David Marx
Steve Herman @newsguy.bsky.social · 23/09/2026
City of San Francisco sues Trump Media for facilitating insider trading by selling early access to the president's social media posts. www.documentcloud.org/documen...
documentcloud.org
Truth Social 2026-09-21 Complaint
19032
David Marx @digthatdata.bsky.social · 23/09/2026
I've been offline most of today. What've I missed?
000
David Marx @digthatdata.bsky.social · 22/09/2026
52 minutes of solid vibes
youtube.com
The Rolling Stones Recording Sympathy for the Devil 1968 [4K Remaster]
YouTube video by Kane McCarthy
021
David Marx @digthatdata.bsky.social · 21/09/2026
next time Texas tries to secede, we should call it "Texit"
010
Reposted by David Marx
Taniel @taniel.bsky.social · 21/09/2026
incredible — some Missouri ballots have already been mailed (off of the 6-2 map after two SCOTUS rulings) but the 8th Circuit just intervened to order the the 7-1 GOP gerrymander reinstated. via Politico: s3.documentcloud.org/documents/28...
21343134
Reposted by David Marx
Jenny Cohn @jennycohn.bsky.social · 21/09/2026
Peter Thiel has the wheel: “US Deputy Secty of Defense Stephen Feinberg has drafted a memo directing Defense Department subordinates to purchase up to $243.9 M in services from…[Thiel’s] Palantir…without a competitive bidding process…The memo’s subject line is called ‘Funding Palantir’.” 8/11/26…1/
share.google
Give Palantir up to $244M through 2028, says internal Defense Department memo
Draft memo taps AI data analysis biz for help with military production efficiency
10516061066
David Marx @digthatdata.bsky.social · 20/09/2026
you are feeling verrrrry sleeeeeepy
youtube.com
Vic Berger & Ben Craw's Week of Miracles (Full Special)
YouTube video by InfoWars
000
Reposted by David Marx
Radley Balko @radleybalko.bsky.social · 20/09/2026
I see we’ve reached the “overwrought comic book villain” phase of Trump’s aspirational despotism. www.nytimes.com/2026/09/20/u...
nytimes.com
Trump Says His D.C. Arch Would Be a ‘Military Complex’ Housing Drones and Snipers
The president’s announcement on Sunday echoed his justification for the White House ballroom: that the project was necessary for national security.
2626350
David Marx @digthatdata.bsky.social · 20/09/2026
what the actual fuck
140