see also are.na
-->
For about 10 days, I was stuck at square 1 with not a single idea. Then I saw a very dim opening. It was very technical, but gradually consolidated and clarified. I continued at the same pace for 49 days, without a single day off — until I posted on arXiv:https://t.co/MnLIGssfcq
— David Bessis (@davidbessis) 24 octobre 2024
Which Moo Deng are you today? pic.twitter.com/giwJHaHHet
— 💖 (@twaniimals) 19 septembre 2024
This UI lives in my head rent-free. I always come back to it once a year. pic.twitter.com/iQMDLu8YwQ
— Adrien Griveau (@Griveau) 31 octobre 2024
the hard truth about health is more or less the same as parenting — the buck stops with you — and that said, it’s worth updating your dogma
— @melissa (@melissa) 11 janvier 2025
this book is written by a doctor, not for other doctors, but rather for high agency ‘patients’
notes: pic.twitter.com/adu91AVNqq
In 2019, a month after I had joined Shopify, I sent this note to my team. I called it "say the thing" and I think it has stood the test of time well. pic.twitter.com/8SThyPnNja
— Kaz Nejatian (@CanadaKaz) 30 octobre 2024
“I will always love you” is table stakes. how about “I will always watch the animal tiktoks you forward me and not just thumbs up them” https://t.co/Y1guZQP4vy
— Ava (@noampomsky) 25 novembre 2024
One thing all the tech bros don’t realize is that taste isn’t about liking all the *right* things.
— mkay 🌐 (@mkay__world) 14 juin 2024
It’s about liking all the *you* things. Having an informed, idiosyncratic view of the world. Strictly aligning yourself with the pillars others have deeemd worthy is a failure of…
in my life I try over and over again to make the practical choice but always end up doing the romantic unreasonable dreamy thing
— Ava (@noampomsky) 6 janvier 2025
true of tweets too pic.twitter.com/wGA2v22wP5
— Henrik Karlsson (@phokarlsson) 6 janvier 2025
The National Toxicology Program's new report on fluoride in the U.S. has eschewed talking about our one large, causally informative study of the effects of fluoride
— Crémieux (@cremieuxrecueil) 22 août 2024
Instead, it focused on confounded epidemiological work!
With this in mind, their report shows nothing concerning. https://t.co/vb8wnXA4GK
Have you noticed that people seem younger at the same ages? 40 is the new 30, 30 is the new 25, and so on?
— Crémieux (@cremieuxrecueil) 6 janvier 2025
There's something to it. People nowadays are aging more gracefully, and what makes this more interesting is that it's a global phenomenon.
Let's talk five "capacities"🧵 pic.twitter.com/wcpf0ng24g
Practicing portraits! I still love the feeling of traditional art :^) pic.twitter.com/xJoCTR3BLe
— Ruth ✷ (@ruzhaoart) 6 janvier 2025
If you take a $1M home and compare
— melody (@melkuo) 16 septembre 2024
(1) 20% down payment, 25 y amortization, vs
(2) 5% down payment, 30 y amortization
In (2), you end up paying the entire cost of the home in interest 🤯 https://t.co/yG5uxnWkOi pic.twitter.com/DV6IzZjOIm
life
People who demand nothing of you (or, notes on charismatic people)
— Carmen (@carmenleelau) 31 juillet 2024
I've been thinking about what makes some people more enjoyable to be around than others. It might be impolite to talk about this, but at least from personal observation, there are differences in how people make…
i simply exist to do the thing and nothing else
— nicole summer hsing (@NicoleSHsing) 7 janvier 2025
I hope you too find your "The Scream" (1893) to "The Sun" (1909) arc pic.twitter.com/osDBdKNV2p
— XH (@xhfloz) 13 janvier 2025
David Lynch: The Art Life (2016) pic.twitter.com/Fi8upwsTjk
— A Shot. (@ashotmagazine) 16 janvier 2025
honestly vietnam is leading the game in inventive fashion and it’s not even close pic.twitter.com/nMeI2tCw2X
— Melisa (@MelisaSeah) 5 janvier 2025
for a while I suspected this and now I'm beginning to see confirmation of it: a lot of what I’ve called "loneliness" is really just emotional insecurity
— kasra (@kasratweets) 17 janvier 2025
— Vivid Void (@VividVoid_) 27 août 2024
i agree with @yacineMTB: "fairness does not exist in marriage"
— @melissa (@melissa) 27 janvier 2025
the media ruined marriages by programming women to think they should be ‘fair’
tldr it's bad code: pic.twitter.com/PGeY7mxh94
love
the best part about craving a grilled cheese with caramelized onion and grilled apple is that you can literally just make a grilled cheese with caramelized onion and grilled apple pic.twitter.com/x3h9pPOo9i
— arianna (@virtualjew) 16 octobre 2024
(no one ever goes looking for IFS or attachment theory or shadow work or whatever else because they're just curious. we all know that we're here to try to understand the heartache at the center of our being and this is very close to the center of mine)
— QC (@QiaochuYuan) 25 février 2023
i am going to make my neighborhood a LOVELY, walkable COMMUNITY where i know everyones NAMES and INTERESTS and we sometimes exchange childcare and come over for COFFEE at each others homes if it's the last thing i do
— nicole ruiz (@nwilliams030) 28 janvier 2025
the exploratorium’s “after dark thursday” is where relationships go to die, or else to settle into the pleasant monotony of many, many years co-existing as a unit.
— brit (@booritney) 30 janvier 2025
for those loathe to take that gamble, i compiled this list of actually good SF dates — just in time for v-day! 💘🧵 pic.twitter.com/sS2yHs0S12
5. i hesitate heavily to share any kind of "general formula" (because the topic is obviously incredibly nuanced and personal) but i guess if i had to create a hypothetical general formula:
— tommy (@tommytrxnh) 30 janvier 2025
- meet at a neutral initiation point
- travel from initiation point to dinner (co-travel=… pic.twitter.com/BaObAbrrha
technology
Nerding hard on digital gardens, personal wikis, and experimental knowledge systems with @_jonesian today.
— Maggie Appleton (@Mappletons) 15 avril 2020
We have an epic collection going, check these out...
1. @tomcritchlow's Wikifolders: https://t.co/QnXw0vzbMG pic.twitter.com/9ri6g9hD93
How Luma Works
— victor (@VictorPontis) 7 janvier 2025
I think Luma is relatively atypical in how we work. It certainly doesn’t feel weird to me, but when I hear about other companies, it’s clear that we do a few things differently.
The Team
Today we have 6 full-time people working on Luma — 3 engineers and 3… pic.twitter.com/oh6K9utJuJ
The plan for Obsidian is to never grow beyond 10-12 people, never take VC funding, never collect personal data or analytics.
— kepano (@kepano) 24 août 2023
Continue building with the assumption that software is ephemeral, files matter more than apps. Use formats that are open and durable.
See our manifesto: pic.twitter.com/UGBK9fmiQL
Static vs dynamic sites in the simplest terms I can think of:
— kepano (@kepano) 10 octobre 2024
Static sites are like printed books. You write your content in an app like Obsidian, and then a tool turns that into a complete, self-contained website. The pages don't change unless you make updates and "reprint" the…
how is this app free pic.twitter.com/wm7JoesRlb
— 313 (@313formation) 26 août 2024
-->It's time to end this pic.twitter.com/T33RYCo1l8
— Xor (@XorDev) 14 septembre 2024
When I was at Meta, I asked the hardware team what I should read to learn the details of modern consumer HW development so I could work more effectively with them. The consensus was that there really isn’t good literature; you just have to “be there”. Has anything emerged since?
— John Carmack (@ID_AA_Carmack) 13 janvier 2025
PSA: If the 'For You' feed is turning into slop, the solution is to go to Settings -> Privacy and safety -> Content you see -> Interests; and turn off all crap, which means almost everything. This needs to be done regularly, but mostly fixes the problem.
— Jan Kulveit (@jankulveit) 13 octobre 2024
— U.S. Graphics Company (@usgraphics) 2 juin 2024
This story about SpaceX engineers transporting a rocket from Texas to Florida is insanely hardcore pic.twitter.com/lvDEy8dRWg
— Alec Stapp (@AlecStapp) 4 janvier 2025
Hypothesis: the world's most valuable data is screen captures of outlier competent people going about their work. But very little of this data is recorded, let alone made publicly available.
— Richard Ngo (@RichardMCNgo) 3 janvier 2025
You should seriously consider recording all work you do, even if just for personal use.
Advice for Food Companies
— Nat Friedman (@natfriedman) 2 janvier 2025
Since we launched PlasticList, we’ve been heartened to have quite a few food companies reach out and ask for help interpreting their results and tracking down and eliminating their contamination. I’ve had calls with a bunch of these.
I am happy to… pic.twitter.com/J43Rcp0VIa
we should be more careful to not conflate value-creation capitalism with value-extraction capitalism
— kasey (@kaseyklimes) 5 janvier 2025
value-creation capitalism: you need bread. I open a bakery and supply you bread at the market rate.
value-extraction capitalism: you need bread. I operate a private equity firm…
So collusion rings are still a thing at top conferences. @openreviewnet has all the data across all top tier ml conferences — what are they currently do to solve this.
— Sara Hooker (@sarahookr) 5 janvier 2025
What top tier conf is willing to release data of bidding patterns so we can analyze the problem at scale.
— Aiden Bai (@aidenybai) 4 janvier 2025
what if your journal visualized your emotions? (little concept inspired by hume and obsidian) pic.twitter.com/zJe3oe3f5I
— lele (@CherrilynnZ) 19 septembre 2024
it's like i just discovered a cartel
— spike (@spikedoanz) 13 novembre 2024
so i benchmarked exp(x) against 2^x.
and exp() was faster by 4x?
turns out we have hard coded exp() tables in our chips https://t.co/Cm6A4r9H88 pic.twitter.com/LnnvmYAGcL
Such a delightful & original way to see 🧵threads! 🤯 So unexpected it's disorienting at first, yet so simple!
— Eli (@elzr) 20 juin 2020
Just rotate a tree 90°: columns are levels of nesting. Click on a block to align & hilite its parents.
You can play with it at https://t.co/h77loEBv0G (very zoomable!) pic.twitter.com/ZQCJCtCJcF
When PG gives you creative ideas (“you know what you should do…”) you’ve got to filter a little. Maybe a lot. Think of it like panning for gold, in a particularly rich river. I’ve personally gotten many incredible nuggets from the River of Paul, for which I am grateful.
— Emmett Shear (@eshear) 17 janvier 2025
Whenever I have pressed the case that John von Neumann should be considered the father of the modern stored-program computer ie the one that nearly everyone uses and carries around in their pockets, I've been met with hurt rebukes.
— Ananyo Bhattacharya (@Ananyo) 2 janvier 2025
These rebukes are sometimes based on scholarship… pic.twitter.com/HZreKJStI8
Every day I see something I am convinced I hallucinated. Today, this was that. pic.twitter.com/hTFxoxaUlk
— Marc Andreessen 🇺🇸 (@pmarca) 19 janvier 2025
— Marc Andreessen 🇺🇸 (@pmarca) 19 janvier 2025
if i made a heavily-commented html template that was a "linktree"... and shared it with the world...would people use it...
— catherine🕷 (@cath_land) 22 janvier 2025
[Rose, Geometrical Charged-Particle Optics, 2012] https://t.co/Wl704hE0Df pic.twitter.com/xDK5TwZodD
— a sentient electron 🧠🔬⚛️ (@quantumbeans) 26 novembre 2024
im assembling a team. pic.twitter.com/Eq4Sj7KTK1
— Tyler Angert (@tylerangert) 30 janvier 2025
Tell me this model isn't aligned // this is the only time I've ever seen Claude 3.5 Sonnet (New) hit the output limit without artifacts on pic.twitter.com/cTfr8f0RXQ
— Zack Witten (@zswitten) 16 janvier 2025
techno-optimist
Interesting. Yep, tech is load bearing. Let's attract more tech entrepreneurs https://t.co/QrRcPFZKSD
— tobi lutke (@tobi) 14 janvier 2025
Yesterday was my last day @neuralink. Some thoughts🧵 pic.twitter.com/u3iJjvOrdy
— Bliss Chapman (@chapman_bliss) 21 janvier 2025
deepseek r1 asked for 1 truly novel insight about humans pic.twitter.com/qTJbPrd5k5
— adi (@adonis_singh) 29 janvier 2025
machine learning
Hypothesis: information work is overwhelmingly bottlenecked on availability of high-signal context more than by correct inference over the context. If right, implies higher ROI-per-flop of context building over pure logical inference.
— Linus (@thesephist) 1 mars 2025
h/t @anandnk24
To CoT or not to CoT?🤔
— Zayne Sprague (@ZayneSprague) 19 septembre 2024
300+ experiments with 14 LLMs & systematic meta-analysis of 100+ recent papers
🤯Direct answering is as good as CoT except for math and symbolic reasoning
🤯You don’t need CoT for 95% of MMLU!
CoT mainly helps LLMs track and execute symbolic computation pic.twitter.com/vEr5oZSSRf
Has anyone managed to extract the NotebookLM podcast generation system prompt yet?
— Simon Willison (@simonw) 29 septembre 2024
I gave a talk at GPU MODE workshop last week on llm.c
— Andrej Karpathy (@karpathy) 2 octobre 2024
- the origin story of llm.c
- being naked in the world without PyTorch and having to re-invent Array, Autograd, Device, Dtype, Compile, Distributed
- how to port a PyTorch layer to 1) explicit PyTorch
- and then to 2) write… pic.twitter.com/u8JXyy90VE
What's really going on in machine learning? Just finished a deep dive using (new) minimal models. Seems like ML is basically about fitting together lumps of computational irreducibility ... with important potential implications for science of ML, and future tech...… pic.twitter.com/nFDVETRZh4
— Stephen Wolfram (@stephen_wolfram) 22 août 2024
Seems timely to remind that this is one of the reasons training compute thresholds are limited as a proxy of ability and risk.
— Sara Hooker (@sarahookr) 12 septembre 2024
Inference time techniques can dramatically improve ability but aren't captured by any current policies. https://t.co/FIZgh05BXl
does anyone actually have an example of cross-layer superposition? I hear it talked about a lot but don't really understand how I'd recognise it in practice
— Tom McGrath (@banburismus_) 2 août 2024
Anthropic interp put out a post two days ago on crosscoders. @Connor_Kissane just put out an open source replication of model diffing crosscoders on Gemma 2 2B!
— Neel Nanda (@NeelNanda5) 27 octobre 2024
We're excited to enable further research: is model diffing a big deal for safety and do crosscoders help study it? https://t.co/yIWnNiUNgq
We previously shared our research on Layer Skip, an end-to-end solution for accelerating LLMs from researchers at Meta FAIR. It achieves this by executing a subset of an LLM’s layers and utilizing subsequent layers for verification and correction. We’re now releasing inference… pic.twitter.com/gag29HSf6e
— AI at Meta (@AIatMeta) 29 octobre 2024
an embedding space for Midjourney srefs.
— Silicon Jungle (@JungleSilicon) 10 décembre 2024
search for styles using natural language
scrub through aesthetic properties like warm, cool, soft, sharp, muted, vibrant, emotions.
then - copy the sref and use it in a prompt. pic.twitter.com/Oa8mzc35cq
research.log - 2024/10/29
— NYRE (@sleenyre) 30 octobre 2024
For the past couple weekends, I've been training Sparse Autoencoders on FLUX / CLIP vision encoders. Today, I'm happy to release fluxlens, an interactive interface for exploring SAE features.
- fluxlenshttps://t.co/2qx6L8TKkS
- blog… pic.twitter.com/Y6vQkRMGrB
just open-sourced the training and evaluation code for cde, our state-of-the-art small text embedding model
— jack morris (@jxmnop) 30 octobre 2024
includes code for lots of hard stuff:
* efficient clustering large datasets
* contrastive training for SOTA retrieval models
* our custom two-stage model architecture that… pic.twitter.com/ZvsssUHL20
What does it mean to think, to question, to understand?
— Junyang Lin (@JustinLin610) 27 novembre 2024
Note: This is the pronunciation of QwQ: /kwju:/ , similar to the word “quill”.
Blog: https://t.co/YEdJQGLa0W
Model: https://t.co/EQdCDY3VzO
Demo: https://t.co/1fMPvUhZv6
Something that can reason, making math problem… pic.twitter.com/6FES6dGERn
@kchonyc wrote a thoughtful note on the AI job market. I think he gets some important things right. Other things seem different from my perspective.
— Chris Olah (@ch402) 3 janvier 2025
My own stream of consciousness inspired by @kchonyc's essay and the broader discourse. https://t.co/8zRhaHrGjd
i don't buy either extreme: programmers won't disappear, but it's disingenuous to think nothing will happen to them
— gian (@giansegato) 5 janvier 2025
my world model is collapsing into a single bit: agency or no agency. that's the whole game
anyone (not just devs) with extreme agency is now 100x. it's not fully… https://t.co/dJyWGtZMpA
I started doing this and the QOL improvements has been insane.
— wh (@nrehiew_) 6 janvier 2025
I highly recommend it. It might take a while to get used to writing and reading it but it is for sure worth it pic.twitter.com/eaPMfDvsiY
Attention has been the key component for most advances in LLMs, but it can’t scale to long context. Does this mean we need to find an alternative?
— Ali Behrouz (@behrouz_ali) 13 janvier 2025
Presenting Titans: a new architecture with attention and a meta in-context memory that learns how to memorize at test time. Titans… pic.twitter.com/YY5vqKY9U1
The salient property of structured decoding is allowing control of LLMs output format while maintaining its non-deterministic nature. @aarnphm_ introduces structured decoding in @vllm_project and its importance in building agentic systems! pic.twitter.com/CUABZu3Klt
— vLLM (@vllm_project) 16 janvier 2025
DeepSeek R1 when it is prompted to select a random number:
— Flowers (@flowersslop) 23 janvier 2025
Okay, the user wants me to think of a number between 1 and 10,000 but not tell them. Hmm, they mentioned not to overthink it, so maybe just pick a random number quickly. Let me see. Well, people often have favorite…
Wrote a post about Highway networks, ResNets and subtleties of architecture comparisons pic.twitter.com/ya6jv7HzMN
— Rupesh Srivastava (@rupspace) 11 janvier 2025
So...
— Taelin (@VictorTaelin) 5 mars 2025
Claude Code just optimized HVM3 to 328 MIPS per M4 Core (+51%)...
And it then created a CUDA version for it.
...
This is serious. Since HVM3-Nano is now just ~5K tokens, it easily fits on Sonnet's context. I literally just asked Claude Code to optimize the repo, and it… https://t.co/75WcMP8Jt3 pic.twitter.com/BNycgr2j8C