Kosta Derpanis a relayé @CSProfKGD
“BACKPROP”
Trailer for my Backprop lecture, coming this winter to my #DeepLearning course 🎬
Computer Vision
Ce qui est entré en base, sans jugement.
2657posts
57auteurs
403liens
1435 hde recul
69membres
Aucun recoupement sur 24 h pleines : ni un lien ni un tweet n'est relayé par
deux membres (2 liens, 19 tweets relayés). Le recul est
suffisant, donc cette fois le constat porte : ces comptes diffusent au lieu de
converger, et c'est la composition de la liste qu'il faut revoir — pas la patience
qu'il faut allonger.
I want to hire the salesperson who was able to sell those air hand dryers at the airport.
G
martin_casado a relayé @appenz
Well done, Opus 5.5. The model my security posture needs, not the one my security practices deserve.
S
Kosta Derpanis a relayé @nnsriram97
We’re presenting VisionSIM as a spotlight at @corl_conf #CoRL2026!
VisionSIM is a Blender-based multi-sensor simulator for RGB, thermal, event cameras & more—with realistic sensor effects and pixel-perfect ground truth.
Paper + code:
🧵 1/n
S
emiran a relayé @SKonvice
Satranç oynamaya 10 yaşında başlamadığımız için asla usta olamayacağımız gerçeği
Kosta Derpanis a relayé @CSProfKGD
Hot Dog vs Not Hot Dog classifier training 🌭
If you know, you know. #SiliconValley
(with audio)
M
emiran a relayé @mustafaozkxn
İtü elektrik kazandıktan sonra bizim hayatın skiliş
Kosta Derpanis a relayé @CSProfKGD
"Mixture of Experts" trailer 🎬
Fun with Codex + @theworldlabs Atlas Marble (beta) + Blender
A
F. Güney a relayé @akanazawa
I had my second baby ~6 months ago 😊 missing CoRL to send time with my family 🫶🏻 but here's a quick look at our lab's recent work, an extended version of last week's BARS talk!
My students are all at CoRL, please say hi! Enjoy👢🦿:
あ
martin_casado a relayé @cumulo_autumn
Shizuku AIは、普段のAI VTuberとしての配信活動だけでなく、その裏で最先端の研究開発にも全力で取り組んでいます！
AIキャラクターにはまだまだ技術的な伸びしろがあり、これからさらに進化していきます。
今後もエンタメを追求しつつ、頼もしいチームと共に世界初の技術開発に挑戦していきます！
あ
martin_casado a relayé @cumulo_autumn
11月はShizuku AIから大きな発表がいくつかあります！
今日はその第一弾の発表をします！
あ
martin_casado a relayé @cumulo_autumn
GPT-Live以降初となる、流暢に話せる日本語Full-duplexモデル、「Shizuku-Duplex」を開発しました！
従来のAIとの音声対話はターン制で、少しぎこちない会話になりがちでしたが、Full-duplex(全二重対話)モデルを用いると人間同士の会話と同じように、「話しながら聞く」ということが可能になります！
J
Kosta Derpanis a relayé @loresjpg
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good
A
martin_casado a relayé @hausdorff_space
Evidently you can only choose 2: {correctly predicted where AI is now, has evidence-based opinions about where it goes next, knows anything about AI at all}
F
how come before every deadline I find myself begging for 80 GPUs on here? the GPU models change, but the count stays more or less the same.
so… anyone? 🥲
R
Kosta Derpanis a relayé @RaminNasibov
Cinema app design that shows you the view from your seat before you book.
Stefan Roth a cité @visinf
So proud of Shweta – very well deserved! 🎉
P
martin_casado a relayé @Briankeating
I am a professor at the @UofCalifornia. I am appalled by the conduct of one of its Regents, Scott Galloway, appointed in August by Governor Newsom.
@profgalloway is one of only 26 voting Regents entrusted with governing a university system serving hundreds of thousands of
Looking into the pension system of Sweden, in short, it penalize people who work hard. Say what you will about US, the 401K contribution is a great gift to working individuals. The financial freedom that country offered to its tax payers are unmatched.
martin_casado a cité @ShizukuAILabs
Full duplex audio synth model!
S
martin_casado a relayé @ShizukuAILabs
We’re introducing Shizuku-Duplex, the world’s first full-duplex audio synthesis model built specifically for AI characters. See how our models are built for natural human-AI interaction:
Satya Mallick a cité @RichardSSutton
That’s what I call a bitter sense of humor.
martin_casado a relayé @mattparlmer
Unfortunately most truly consequential projects require 10x this much money right out of the gate, preseed VC as presently structured is built to back incremental nibbling around the technoeconomic frontier
Capital access is better now than before, but still underinvested
N
martin_casado a relayé @nicksherrow
Can't stop thinking how much stuff Jev is doing to save AI engineers asses from stupid systems design patterns while the rest of the domain keeps pushing tool calling loops everywhere
martin_casado a relayé @GergelyOrosz
How the hell does Jev cost $0.00 after using more than 1M tokens (on my new search site)
an LLM would have charged me $3-5 already. Jev pricing is incredible, pricing is per *billions* of tokens not millions
huge reason this spreads like wildfire. great work @CompleteSkeptic
T
martin_casado a relayé @tejasmanohar
our eng team got #1 on HN today!
they tested if fast + cheap decision models like Jev @typesafeai could grade complex data-matching tasks as well as frontier models.
Jev with Claude Opus got close to Opus-level accuracy for a fraction of the cost.
super cool. post below:
martin_casado a cité @karinanguyen
Great. More of this line of inquiry please.
T
martin_casado a relayé @kirbywinfield
there are literally hundreds of preseed funds writing $250k - $500k checks. like, inception stage investment. if you can’t find or raise from these folks who are purpose-built to back your stage, it’s probably a you issue, not a “reliable sources of funding” issue
G
martin_casado a relayé @Geiger_Capital
我们支持您，国会女议员！美国应当把重心放在引进第三世界人口并照料他们身上，而不是关注科技发展或本国人民的福祉。
R
martin_casado a relayé @bekhovsgun
can gaussian splats be used to render interfaces?
I've seen a lot of incredible spatial demos coming out of the labs (@theworldlabs has been blowing my mind ~weekly), but they've mainly been focused on static 3D environments. I woke up today wondering: are splats any good for UI
N
martin_casado a relayé @nikitabier
If any elected leader opposes data centers or the reindustrialization of America, just say you support them in Chinese. Make every reply below their posts in Chinese.
They will eventually go insane.
martin_casado a cité @nikitabier
Post of the year. Competition closed.
G
Sindhu Hegde a relayé @gundrymd_
Health expert Dr. Steven Gundry has a test that will help you determine your gut type.
Take his free quiz today!
When #ICLR2027 reviews drop right in the middle of your workday. 🤨
V
martin_casado a relayé @vishalmisra
@martin_casado The bitter lesson is RSI is the definitions we made along the way.
I really encourage those who haven't read Sutherland's Wheels of Incarnation to do so.
The history of our industry is not necessarily specialized approaches being gobbled up by general approaches. But it's as much specialized approaches growing towards generality.
The
D
martin_casado a relayé @CompleteSkeptic
very sick! our guess on the first wave of jevons paradox-ing intelligence will be "dark data"
why don't we throw intelligence at the big data we already have? too expensive + too unreliable in map-reduces
but obviously jev changes both of those issues (:
Kosta Derpanis a relayé @CSProfKGD
"Mechanistic Interpretability”
Trailer for my Mech Interp lecture, coming this winter to my #DeepLearning course 🎬
martin_casado a relayé @RichardSSutton
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
Kosta Derpanis a relayé @RichardSSutton
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
P
Kosta Derpanis a relayé @peter_richtarik
I am an AC for ICLR 2026. One of the papers in my batch was just withdrawn. The authors wrote a brief response, explaining why the reviewers failed at their job. I agree with most of their comments. The authors gave up. They are fed up. Just like many of us. I understand. We
More founders, more seed funds, more VCs, more investment in private markets, more startups.
Douglas Gray a cité @RLanceMartin
Great article on agent memory.
Amusing failure case: Claude “summarized” 115KB of text into 203KB of wiki pages.
I think anyone who has spent time working with agents knows exactly what that feels like.
martin_casado a cité @Rick_Zullo
Innovation / tech doesn’t create value is such a wierd stance for a VC.
J
Kosta Derpanis a relayé @johnfewell
My isometric map of Toronto now covers the whole City: all 630 km², from Etobicoke to Scarborough and from the lake up to Steeles. And all of it is redrawn in a new, crisper style. 🧵👇
A
Kosta Derpanis a relayé @iarthsingh
Guys this ICLR paper got a 4 ? WTF ??? Generational AI slop it was
Douglas Gray a cité @CNGI_news
Is UCSC finally doing something to address student housing?
martin_casado a cité @nedoleary
I don’t agree. More money in private markets means more TAM and value created in private markets.
I just don’t understand zero sum thinking for VCs. Isn’t the entire belief structure that innovative tech creates net new value?
martin_casado a relayé @__tinygrad__
@martin_casado YOU HAD A GPU WRITE A GPU KERNEL?!? THAT'S RSI!
martin_casado a cité @FelixDeSimone
The legacy of Bostrom is going to haunt us forever.
This is all so boneheaded and avoidable.
martin_casado a cité @Sam_McCandlish
"Many of their conclusions match my own. AI is speeding up our research, but it’s hard to say exactly how much. While they conclude AI can't yet make the most important contributions to model training on its own, they think the pace of development could, by default, accelerate
martin_casado a cité @RepCasar
This is exactly why sloppy, overly broad use of the term RSI is a very bad idea.
Many technologies have autocatalytic effects. That's not RSI. Fear mongering and pointing to basic shit like using AI to create GPU kernels as RSI is encouraging this sort of rhetoric.
So chatted with a researcher from one of the large labs who was arguing the LoC is a reasonable metric for RSI. And evidence for the RSI that AI is writing a lot more code (for building AI). How can this be a serious view?
Douglas Gray a cité @elonmusk
Tired: Lights out factory
Wired: Zero atmosphere chip fab
F
Frank Dellaert a cité @NicoBohlinger
Everyone wants a SuperMimic!
It's quite interesting to watch Turing award winner Sutherland's "Wheel of Reincarnation" observation playing out with models. For example, how pixel models more and more look like language models.
L
Atsuyuki Miyai @UTokyo a relayé @lens_kousiki
【緊急】Weight-Space Programming：11/3に発表されたLLMによってモデルの重みを直接書き換えることでスコアを飛躍的に向上させる新手法。
その著者、武樋力哉さん(MIT 博士課程1年)がなんとLENSで公演してくれることになりました！！
11/10(火) 21:00〜 オンライン
K
Francis Engelmann a relayé @KadirYilmaz_CV
I think a thread is needed here 🙂. VINO is a DINO-style self-distillation framework for 3D scenes: a student network learns to predict targets from a teacher, which is itself an EMA of the student.
1/10
W
martin_casado a relayé @WKCosmo
If this proof holds, it means that quantum computers are only polynomially faster than classical computers for general problems. Shor's algorithm is an exception, not the rule.
A
martin_casado a relayé @AlexKrentsel
on weight-space programming (one-shot), Inkling-2-small takes 2nd place overall. top amongst US open-weight models. @thinkymachines nicely done!
J
martin_casado a relayé @jameskimEDU
The funny thing is, even if this nonsense _were_ widespread, it would result in some of the easiest “non-consensus but right” opportunities in the history of venture, and the equilibrium would reset.
martin_casado a relayé @teortaxesTex
wtf, this is excellent and actually somewhat scary. Thinky did this on a tiny compute and token budget, without any fancy architecture or RL tricks. This is the same level of achievement as 1017. Maybe better.
This is FINALLY what can be expected of top-tier ex-Frontier staff.
S
martin_casado a relayé @SeongsikKi5837
One of the most impressive things about Inkling-2-Small is its self verification capability with vision in the loop 👀 It builds 3D scenes in Blender, renders them, inspects the output, spots what’s wrong and keeps iterating until it’s right. Huge congrats to @du_peichao75719
R
Matthias Niessner a relayé @risingodegua
You can create a high quality replica of any space now.
These scenes were captured with just a phone and reconstructed using Echo-3.
Look at the quality! 🤯
A
martin_casado a relayé @andrew__reed
When VCs got focused on revenue multiples we invested in mattress e-commerce companies
M
Kosta Derpanis a relayé @mbilgehanertan
ICLR 2027 reviews are out and apparently your paper ID matters more than you think lol
#ICLR2027 #iclr
Matthias Niessner a relayé @Michael_J_Black
The dogma of the last decade in computer vision is that all classical algorithms should be replaced by neural networks. This has been very productive but, counterintuitively, with the latest frontier models, I find myself sometimes turning back to classical methods.
If you have
B
Kosta Derpanis a relayé @DrLaschowski
@GoogleDeepMind in Toronto is hiring a PhD student researcher for winter or summer 2027.
Interested? Apply here:
D
martin_casado a relayé @SeizeEndowments
@martin_casado Looks like a ton of fun 🔥. So much left to do with memory and context --- esp domain-specific / project specific tricks. Having extensible is fantastic. Love the violent example :). Would be cool to fight harness against harness in an arena and see results.
I had to skim
M
PSA: Enable two-factor authentication on @openreviewnet.
Might be a false alarm, but better safe than sorry these days.
Kosta Derpanis a relayé @NeelNanda5
After supervising 20+ papers, I have highly opinionated views on writing great ML papers. When I entered the field I found this all frustratingly opaque
So I wrote a guide on turning research into high-quality papers with scientific integrity! Hopefully still useful for NeurIPS
martin_casado a relayé @DesignArena
BREAKING: Inkling 2 Small by @thinkymachines takes 4th place overall on DesignArena and is 1st place among open weights models with an Elo of 1376!
At $0.30/$1.00 per 1M input/output tokens, Inkling 2 Small is 15× cheaper on output than Kimi K3 by @Kimi_Moonshot. This model also
N
Kosta Derpanis a relayé @Jimantha
Please apply to come be a postdoc at Cornell through this wonderful Empire AI Fellows program! I'll be looking to recruit a postdoc myself alongside the rest of the vision group at Cornell Tech in NYC.
martin_casado a relayé @teortaxesTex
OK pretty cool, a respectable "Flash" model
obviously crushes Muse Spark
Even has some "RSI" capabilities
Congrats Thinky!
martin_casado a relayé @kimmonismus
Inkling 2 small is an extremely impressive release, especially for its size.
It is a Mixture-of-Experts model with 594B total parameters, 17B active, and supports image inputs and variable effort.
In coding benchmarks, it easily keeps pace with significantly larger models. What
0
martin_casado a relayé @0xSero
First American Open Weight model that mogs all the Chinese Open weight models
Very impressive, inkling is my fav US lab.
Looking forward to optimizing this
Satya Mallick a cité @Michael_J_Black
A lot of algorithms, especially in geometry are difficult to code and very delicate.
Now, thanks to frontier models we can try those ideas out.
Not everything needs to be solved using a big hammer of neural networks.
martin_casado a relayé @johnschulman2
Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process.
To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and
R
Kosta Derpanis a relayé @ziv_ravid
ICLR reviews are out.
Reviewer 1: “The paper is well-written and addresses an important problem. However, the authors should consider additional baselines.”
Reviewer 2: “The paper is well-written and addresses an important problem. However, the authors should consider additional
martin_casado a cité @MartinGTobias
I interact with dozens of GPs all the time in and out of the firm and no-one thinks this way or says stupid shit like this.
Yes, there is high variance in fund raising right now. But I don’t know any legit GP that makes decisions on current ARR or growth alone.
Emirhan takva denen elemanın videolarına bağımlı oldum. Can aybalık son 2 senedir ai slop dışında video atmadığı için ilaç gibi geldi
P
Matthias Niessner a relayé @PhysiclAI
Today we're launching Physicl Studio, and we're really proud of this one.
Tell it the situation your model gets wrong. You get it back as images or video, varied thousands of ways, with exact depth, normals, segmentation, boxes and camera pose on every frame.
S
Kosta Derpanis a relayé @LikeACloud777
Got a 2-2-3 at ICLR 2027? You're not alone: it's the single most common rating set (18.9% of papers).
On the new 1–4 scale:
2-2-3 → top 37.8%
2-3-3 → top 15.1%
3-3-3 → top 4.1%
Median paper average: 2.0. 57% of all reviews are a 2.
#ICLR2027
K
Kosta Derpanis a relayé @kgrotov
Peak peer review. (We submitted to ICLR)
#iclr2027
Kosta Derpanis a relayé @Michael_J_Black
The dogma of the last decade in computer vision is that all classical algorithms should be replaced by neural networks. This has been very productive but, counterintuitively, with the latest frontier models, I find myself sometimes turning back to classical methods.
If you have
K
Are we already treating the code as foreign (and unknown to us) language?
Dmytro Mishkin a relayé @taiyasaki
And the goodies keep on coming. You can also SfM large scenes by randomly processing subsets of 8 views at a time.
D
martin_casado a relayé @SeizeEndowments
One of the craziest operating system lab presentation I've ever seen (or heard!).
For his 340lx midterm project @houjun_liu built his own custom bare metal synthesizer with attack curves, tuned decay, custom drivers, dma engines, hard realtime, effect keyboard all running on
M
Liping Bai 白莉萍 a relayé @0x_meden
The Princeton professor who also works as a research scientist at Google DeepMind is uploading an entire robotics course to YouTube this fall, and the full previous run is already online: 24 lectures, about 29 hours, free.
"Introduction to Robotics," Anirudha Majumdar
Douglas Gray a cité @thomaspaulmann
Reading all these replies while nodding my head in agreement.
Biggest bottleneck in AI right now is not doing the work but human understanding and verification of the work.
What the hell?!! I can't believe how far we've come. Atlas is so good. Look at this fucking splat!! Check out behind the DJ table. Totally consistent novel view. No ghosting. This is entirely generated from a single image.
I'm gonna turn this into such a sick VR experience.
J
martin_casado a relayé @JosvdWest
3D consistent generation for Anime
N
N
martin_casado a relayé @neilsonks
We're in a computer graphics renaissance
Drawings, photos, renders are all something world models can "see into", world vertigo
A
martin_casado a relayé @allietheicon
Why don't we at TypeSafe call Jev a decision model?
If you think of Jev's questions as making *decisions*, I believe you'll use it at the wrong layer of abstraction and your reliability will suffer.
I wrote all about this on our site 🧵👇
D
martin_casado a relayé @CompleteSkeptic
an absolutely beautiful and informative blog post! one of the biggest anti-patterns we see with jev is just plugging it in to make Big Decisions instead of having code do what it's best at and jev for just the parts that require judgment 🤌
THIS is how we transform software!
martin_casado a relayé @thinkymachines
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
M
Kosta Derpanis a relayé @MatthewKowal9
This was one of the coolest and most fun projects I have ever worked on: mechanistic shaping. We introduce a new framework for shaping what concepts a model learns to rely on during training.
My favorite applications from the paper: medical imaging. 🩻
martin_casado a relayé @jcjohnss
We've released a demo of real-time world generation powered by our new Atlas model!
This is a spiritual successor to our RTFM demo from last year. It's a vision for the future of interactive media might look like, with frames streamed directly from a powerful generative model
martin_casado a cité @poteto
Yup, exactly my long standing view on this too.
D
martin_casado a relayé @dangeislerjr
The frame boy demo is awesome but it has no physics. I hooked it up to a vision model so it can automatically adjust the camera so It can stop me from walking through walls and even walk up and down stairs
I
martin_casado a relayé @istoica05
What if you evolved the harness alongside the solution in an agentic loop? Big reductions in the cost and time to generate state-of-the-art solutions. Check out Exo, our recent work, for more details.
B
martin_casado a relayé @bcrussett
Congrats @AlexKrentsel @martin_casado @ankrgyl! Exo is awesome. I've been experimenting with it for an agent that plays Celeste using Opus 5.5 and improves its strategy between runs and writes tools for solved rooms. Full writeup and gameplay videos here:
martin_casado a relayé @BenMildenhall
reexperience the garden
T
martin_casado a relayé @tianyin_xu
Great work by @AlexKrentsel and co!
35% reduction of diagnosis time on #SREGym with significantly lower model cost.
Love the principle and architecture,
> Inspired by OS work on separating policy from mechanism, particularly ExoKernels, we expose primitive interfaces that
martin_casado a cité @theworldlabs
SOTA next frame prediction model, now real time. Wild how fast this stuff moves. Also, sick video :)
Douglas Gray a cité @hugo_edwards_
This is what happens when Claude throws the challenge flag.
Should be a metric for how often a model is right but eval ground truth is wrong.
T
Kosta Derpanis a relayé @thomas_fel_
A science of post-training can be built on interpretability.
Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules.
In vision, DPO & RL, and its effects on monitoring 🧵
Kosta Derpanis a relayé @jcjohnss
We've released a demo of real-time world generation powered by our new Atlas model!
This is a spiritual successor to our RTFM demo from last year. It's a vision for the future of interactive media might look like, with frames streamed directly from a powerful generative model
Kosta Derpanis a relayé @GoodfireAI
We cut through the mystery around interpretability in training.
A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities.
We call this *mechanistic shaping*, and we’ve used it to mitigate
S
Laura Leal-Taixe a relayé @ShengyHuang
1/7 Introducing MetricTrail: pose estimation for long videos, with no long-sequence training required.
📄 Paper:
🌐 Project:
💻 Code:
🤗 Models:
A
martin_casado a relayé @AlexKrentsel
for completeness, i ran the same setup on a series of open-weight models. we see none are able to beat the PPO network baseline from the original blog post -- though Kimi K3.1 Pro gets closest
A
martin_casado a relayé @ankrgyl
Working in o11y has led to a lot of time thinking about self-improvement. @AlexKrentsel @martin_casado and I have been working on researching how you design a system that can safely self-improve and embraces the bitter lesson.
Lots of cool results in Alex's write up
EoD AoE or just AoE? @iclr_conf
J
Zhengzhong Tu a relayé @jdzhang0929
Can a frontier agent recursively improve the way it works?
We built Flex-RSI to measure exactly that: recursive self-improvement of frontier agents.
🔗👇
If you are having a bad day, remember Agility Robotics IPO-ed today.
D
We are excited to share our #CoRL2026 paper, "Learning Visuomotor Quadrotor Flight Through Dynamic Narrow Gaps via Differentiable Simulation", where we demonstrate autonomous quadrotor flight through narrow moving gaps using exclusively onboard vision, with policies trained from
A
martin_casado a relayé @AlexKrentsel
Agent harnesses are usually hand-tuned, or optimized offline against benchmarks. What if an agent could improve its own harness *while doing the work*?
Introducing Exo, a harness architecture for runtime self-improvement 🧵
(paper:)
Late update (finally located the certificate😅)
Honored to be selected as a Top Area Chair at #NeurIPS2026 (Evaluations and Datasets), after being recognized 4x as Top Reviewer at ML venues
This means a lot to me, especially during the era of most chaotic peer-review experience
C
Liping Bai 白莉萍 a relayé @LeadClearly1
Your boss, your colleagues, your clients and your team are all watching how you handle pressure, even when they never say it out loud.
They notice if you stay calm when plans change, if you take responsibility when something slips, and if you treat people with respect on a
T
Elliott / Shangzhe Wu a relayé @ThiboNeveu
We're opening 100,000+ 3D objects 🧊🤗, unified in one library, ready for LLM training.
Objaverse assets, CAD parts and articulated mechanisms, all in one language !
For years, 3D research has trained on Objaverse: messy triangle soups, broken meshes, objects with no parts.
G
martin_casado a relayé @appenz
The more I am looking at AI for semiconductor design, the more it seems to me that:
1. RTL generation is just coding
2. 100% of RTL will be generated by agents
3. This means frontier models will dominate
And to be clear, we still need specialized tools for those agents that do
H
Kosta Derpanis a relayé @HildeKuehne
Our friends at @Uni_Stuttgart have some open positions for ELLIS Independent Junior Research Group Leaders incl. one for Computer Vision: Deadline Nov 15th
If you know a good PostDoc who wants to grow his/her own group, please let them know.
So far I git Claude security flag twice - when discussing sheep/ram in Imagenet classes, and when tried to make more secure my mom's router.
Codex is fine with both.
S
Ehsan Adeli a relayé @StanfordHAI
Are you developing AI or technology to support older adults, including people living with Alzheimer's disease and related dementias? Stanford AITC has launched a research award competition, supported by the National Institute on Aging. Join a webinar to learn more: November 4 | 9
Dmytro Mishkin a relayé @pesarlin
Did you like SuperGlue & LightGlue? You’ll love FlexGlue, our new model for image matching
1️⃣ It’s flexible: sparse & dense point tracking & matching, all in one model
2️⃣ It rejects negative pairs: it won’t match symmetries that often break SfM!
Available in COLMAP, try it out ↘️
I have been experiencing quite a lot GPT blackout in the past 3 days. I am using Pro in chat @thsottiaux @_catwu @ajambrosino
The way I read research papers has changed a lot recently.
Before, when I read papers, I often tried to understand the researchers' personalities and ways of thinking through their research.
But now, because of AI, it is becoming difficult to tell how much of the research comes
A
martin_casado a relayé @arampell
Very proud of my son @cappucher for working for months (with his AI collaborator) to, for the first time ever, prove the asymptotic for E.P. Stein numbers (in the put-or-take-a-square game)
And thanks to @skominers for advice and counsel along the way!
M
Ken Sakurada a relayé @mikittt417
Noetraでの自分の最初の仕事のテックブログを書きました！
マルチモーダルLLMに用いる画像エンコーダーを開発した話になります。
S
Manu Gaur a relayé @vsreyas20
New work with @kowshik_shreyas and my amazing advisors @max_simchowitz & @aviral_kumar2
We find that just one demonstration is enough to bootstrap autonomous learning in pretrained robot policies. With the right recipe, RL can learn reliable policies from brittle initializations!
martin_casado a cité @sequoia
A lot of really smart moves here. I suspect over time it'll be very clear how thoughtful these decisions were.
K
Kosta Derpanis a relayé @kate_saenko_
My group at FAIR (Meta) is looking for a summer research intern. If you have experience with SAM/object segmentation, especially in microscopy images or similar domains, and are excited to contribute to the next generation of SAM models, apply here:
Kosta Derpanis a relayé @ylecun
@Arjunjain Plot twist: I was Vapnik's Department Head between 1996 and 2002, and made sure he had the space, resources, and collaborators to do his best work.
To people who think I'm dogmatic: I famously disagreed with Vapnik about the best way forward in ML, but did what I could to
B
Kosta Derpanis a relayé @BoqingGo
#hiring# Computer Science at Boston University invites applications for a tenure-track faculty position at the rank of Assistant Professor beginning July 1, 2027. We seek candidates in AI, broadly defined.
Z
Zhengzhong Tu a cité @jdzhang0929
Can a frontier agent recursively improve the way it works?
We built Flex-RSI to measure exactly that: recursive self-improvement of frontier agents.
Check out more details below👇
F
1 hour writing about exp.: poof, gone. days until CVPR fly by like hours.
1 sec holding a plank: I swear I can count fractions of a sec, and each feels like at least a few sec.
I guess we all have our strengths, Margot Robbie* 😂
*Margot Robbie famously held a 4-min and 10-sec
E
Yoshitomo Matsubara a relayé @EhudReiter
TACL update: Papers which were submitted to ARR or xACL conferences within the past 9 months cannot be submitted to TACL. This includes papers which were unsuccessful in the ARR reviewer capacity lottery, or otherwise were desk rejected by ARR.
I think @RicursiveAI and @Recursive_SI need to play rock-paper-scissors to decide which one of them has to change their name.
Kosta Derpanis a relayé @juliarturc
I wanted to understand the structural and philosophical differences between GPUs, TPUs and newer alternatives (like Groq LPUs) without getting drowned in specs and jargon.
Mispronunciation of the day: "soldered" 🤦♀️...
0:00 Intro
0:56 The origin story of TPUs
3:42 The first TPU:
J
Kosta Derpanis a relayé @johnfewell
Isometric Toronto: Now with traffic. Simulated cars on Toronto's real road network, from the Gardiner to Bloor.
Explore it:
V
Zhengzhong Tu a relayé @viskoai
Top 3 teams take home cash. And every member of the top 5 teams walks away with a Google product. 🏆
$𝟭𝟱𝗞 𝗽𝗿𝗶𝘇𝗲 𝗽𝗼𝗼𝗹 at Orbis Build Day:
💰 $4K / $2K / $1K cash
🎁 Pixel Buds, Fitbit Sense, Nest Learning Thermostat, and more from @googlecloud
🎟️ 2 free tickets to
martin_casado a cité @patrick_oshag
My friend @vitalygordon made a good point on the question of market saturation for tokens.
Right now for most companies, tokens are still <5% of labor. So it doesn't makes sense to skimp on making the 95% productive. Somewhat like giving high paid cooks dull knives.
I gotta' say, the memes I'm seeing from the moderates/right on the election results are a lot funnier and more good natured than when similar happened last election cycle from the other side.
B
martin_casado a relayé @boxerbk
Today we’re announcing an $8.75M seed round for Phaze.
Companies aren’t ready for how many computers people and AI agents are about to need. And it’s happening just as computer prices are up 3x.
Chris and I spent years building Parsec to make a computer thousands of miles away
O
Antoine Guédon a relayé @oriane_simeoni
Principled SSL, simpler than DINOv3, and matching its performance?
Check our new SSL algorithm JEM on arxiv 🔥
A big effort led my @maxseitzer who has shown me that patient work still has it's place in today's world 🫶
M
Francis Engelmann a relayé @mapo1
If you liked LightGlue, you will like FlexGlue even more! It's better on all fronts (also beats RoMa & LoMa), sparse and dense matching, also uniquely suppresses matches between similar, but non-matching images. Best of all, it's already available in COLMAP.
P
martin_casado a relayé @patrick_oshag
Martin Casado (@martin_casado) explains his tweet on non-consensus investing:
"The idea that non consensus investing is where the alpha is, is actually quite dangerous in the early stage.
Follow on capital tends to be more and more consensus aligned."
"There's something tied
A
martin_casado a relayé @arbeemonkey
Studio notes — Wednesday, November 4
It's incredible how musical collaborations can take you into new territory: new ways of working, new sounds, places you hadn't explored yet. I think it's the best way to step out of your comfort zone and grow.
I
martin_casado a relayé @InvestLikeBest
Martin Casado (@martin_casado) has been building and investing in technology for 30 years.
We distilled the most important lessons from our conversation on the craft of venture investing.
1. Private markets are efficient. If you believe in markets, you should believe private
О
Dmytro Mishkin a relayé @cvtalks
Kaggling is hard.
But have you tried to kaggle when you have a power for 8 hours a day?
#RussiaIsATerroristState
P
martin_casado a relayé @patrick_oshag
My third conversation with Martin Casado (@martin_casado), general partner at @a16z.
- The case for consensus investing
- Why moats are overrated
- Frontier model economics
- What happens when AI growth slows
- Why private markets are efficient
- Lessons from Cursor, OpenRouter,
martin_casado a cité @patrick_oshag
The most in depth conversation I’ve ever had on the practice of venture. It’s a good overview of what I’ve learned over the last decade.
Our Claude maze agent reached 9.8 MB per request by move 100. Using message threads, that dropped to 104 KB: just the new screenshot and the previous response ID. Token use stays the same, and threads expire after 5 minutes. We measured upload size, not speed.
It's the economy, stupid.
Satya Mallick a relayé @ylecun
@Arjunjain Plot twist: I was Vapnik's Department Head between 1996 and 2002, and made sure he had the space, resources, and collaborators to do his best work.
To people who think I'm dogmatic: I famously disagreed with Vapnik about the best way forward in ML, but did what I could to
martin_casado a cité @SteveKopack
Funny. I wonder where they're getting these ideas? 🤔
Douglas Gray a cité @picotop
Personal pet peeve:
Pronouncing acronyms as regular words.
Worst case: SotA (state of the art) as sow-tah
Hell yeah, new gear day. Make Noise ReSynth
D
martin_casado a relayé @DimitrisPapail
Ahem, I would currently bet less on P!=PPAD than say 18 days ago 😅
D
martin_casado a relayé @dougleone
Congrats to the @cresta team on its next chapter with @salesforce.
Read more about the journey here:
martin_casado a cité @realchillben
Amazing to have Bill join the team!! Welcome!
J
Kosta Derpanis a relayé @johnfewell
The whole downtown core of my isometric Toronto is drawn: Bathurst to the Don, Bloor to the harbour.
12.66 km² as one 256-megapixel image, about 2% of the city, made in one evening on my laptop.
Explore it:
As part of my campaign to point out that discussions on defensibility are generally pointless, here is a very classic theoretical result showing that you can't even efficiently compute Nash equilibrium states.
B
martin_casado a relayé @isidentical
Even with billions in your bank account, you can’t buy enough compute to beat the majors head-on. What you can do is extract far more value from every MW you have.
That’s why key theme of 2026 and 2027 for us is post-training. Improving the quality of diffusion models while
cansiniz 🧡😂
J
Liping Bai 白莉萍 a relayé @jacobvm04
when LaMAria was introduced a year ago, Meta's closed-source Aria SLAM was far ahead of academic systems.
one year later, we've surpassed it in 3/5 categories: medium, long, and low light.
here's what closing that gap looks like:
A
D
Andrew Davison a relayé @dwightchurchill
the fact that this was written … last year … breaks my brain
martin_casado a cité @UnrealEngine
Amazing move! Let's goooooo!! Unreal to web is back!!
P
Dmytro Mishkin a relayé @PhilippCSE
To match or not to match?
FlexGlue 🐙, our new image-matching model, can match every pixel AND reject doppelganger pairs, tackling a common failure case in SfM.
➡️State-of-the-art SfM and matching, blazing fast
➡️Code:
➡️Paper:
1/
I
📊 We've revisited the cities and the most-cited papers. Now, a look through the research archive: 1,351 papers across 17 topics, from 2013 to 2026.
We're excited to see the new ideas and directions our community brings to #3DV2027!
F
gece 2’de iyi sonuç alıp uyuyamayan öğrenciyle..
ben: bana niye haber vermedin, beraber sevinirdik.
Eray: hocam geç olmuştu, uyursunuz diye.
hoca durur mu, yapıştırmış cevabı: sen benim geceleri uyuduğumu mu zannediyorsun Memati 😂😂
K
Kosta Derpanis a relayé @kwangmoo_yi
Our own paper today. Kwak et al., "Solv3R: 3D Looped Transformer as a Multi-view Solver"
Multi-view geometry has traditionally been "solved" with steady-state solvers. We use a Loop-Transformer to do exactly that. What I like most is that you can pipeline for 16x+ throughput.
Dmytro Mishkin a relayé @taiyasaki
📢📢📢 Introducing 𝐒𝐨𝐥𝐯𝟑𝐑
Can geometric foundation models move beyond fixed-set prediction and keep solving like classical SfM?
We turn a looped Transformer into a reusable 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐬𝐨𝐥𝐯𝐞𝐫
🌐 Project:
🧵👇
Kosta Derpanis a relayé @taiyasaki
📢📢📢 Introducing 𝐒𝐨𝐥𝐯𝟑𝐑
Can geometric foundation models move beyond fixed-set prediction and keep solving like classical SfM?
We turn a looped Transformer into a reusable 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐬𝐨𝐥𝐯𝐞𝐫
🌐 Project:
🧵👇
F
Satya Mallick a cité @thsottiaux
Day 28 should have ended in a bang but it ended with a bank.
But still. Great effort by the Codex team.
28 ambitious days and a lot was delivered!
A
Matthias Niessner a relayé @angelaqdai
📢VoxLution: Real-world scans are noisy & partial, so 3D scene generators often rely on synthetic data.
@DaoyiGao & @xinyi092298 introduce a sparse voxel cascaded transformer trained on real scans, generating high-fidelity realistic scenes!
Check it out:
martin_casado a relayé @willdepue
tldr very much signs of life but nothing super dramatic
must do: upstream RL tasks to make Astra-next much better at this auto-interp work
M
Kosta Derpanis a relayé @MarkSchmidtUBC
Narrative-Stye CVs? What new devilry is this?
We need another CV format for Canadian granting agencies, after the incompetently-implemented and infuriating-to-use CommonCV was supposed to solve the issue that there are too many CV formats?
martin_casado a relayé @willdepue
@martin_casado RSI mostly won’t happen at the weight level! at most this is an interp team effort to make the models good at auto-interp. but given astra an opus both rise together, i assume this is just an emergent capability at scale
Yoshitomo Matsubara a cité @rikiyatakehi
Sounds like a surprising result
Is it possible that GPT-6 Astra is aware of the target task and have access to some fine-tuned Qwen weights in some way (say 🤗) to cheat?
Some of the replies imply that it won't generalize well
A
martin_casado a relayé @hausdorff_space
You can't ship what you can't verify. Distributed systems are hard to verify because, like, 50% of the semantics have to be learned operationally. This is even true of something super vibe-able, like Raft. Here are some things prod will teach you the hard way:
* Raft's "correct
martin_casado a cité @AlexKrentsel
Oh wow. Repro of the MIT blog post.
Look at that step change. This must be an artifact of efforts to do RSI within the labs, no?
A
martin_casado a relayé @AlexKrentsel
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
martin_casado a cité @AlexKrentsel
Oh wow. We’ve a repro of the MIT blog post.
Look at that step change. This must be an artifact of efforts to do RSI within the labs, no?
C
Halloween treats for the class 🎃
O
Laura Leal-Taixe a relayé @olvr_hhn
📢 LucidLyra: Feed-Forward Generative 3D Scene Reconstruction and Joint Understanding via Multi-Task Distillation!
Generate a queryable 3D Gaussian scene from a single image! Our feed-forward decoder jointly predicts geometry, appearance, open-vocabulary semantics, and object
A
Aviad Dahan a cité @rikiyatakehi
it's like reverse engineering assembly code but for neural networks
Douglas Gray a cité @rikiyatakehi
Franticly updating priors about what the future of ML research looks like…
Spoiler: “Just Ask”
E
Ehsan Adeli a cité @StanfordBrain
So great to be part of the Stanford Digital Brain Project! With @dyamins and an amazing team at @StanfordBrain, we're building a digital model of the human brain to learn how it works and, one day, better understand and treat brain and mental health disorders.
Z
P
Kosta Derpanis a relayé @philiptorr
mmm discovering a bunch of papers that have added my name on without me being aware of it, seems difficult to get your name removed once they are out there.
🐝
Kosta Derpanis a relayé @readyapplego
Lately, there’s something so exhausting about attending a professional or academic session. Idk what to call it but everything feels so same now. PPTs used to reflect personality. People would actually put thought into the design, wording, little details. There was subjectivity.
O
Kosta Derpanis a relayé @olvr_hhn
📢 LucidLyra: Feed-Forward Generative 3D Scene Reconstruction and Joint Understanding via Multi-Task Distillation!
Generate a queryable 3D Gaussian scene from a single image! Our feed-forward decoder jointly predicts geometry, appearance, open-vocabulary semantics, and object
Finally
@CVPR thank you for removing hidden trap for the workshop organizers and exposing the Metadata submission deadline!
Next thing to dream about, is to move/remove it -- why do we need it like month before the actual paper?
Two people roll a cable reel away from a substation in the dark.
Camera that records keeps the footage. System that reasons flags the theft as it happens, with context to act. Ambient Threat Detection, on cameras you already run.
#CriticalInfrastructure #AgenticPhysicalSecurity
Kosta Derpanis a relayé @davidbau
I have been teaching myself formal code verification, but as the practice becomes more widespread, we must remember that mathematical perfection is often *not* good enough.
Our systems need to be understandable to humans.
martin_casado a cité @a16z
I promise you'll l earn something from this. It's one of the more in depth discussions I've ever had on end to end economics of robotics, why robotics cos tends to verticalize, the complication of the long tail, etc.
martin_casado a cité @joshgruenstein
Tutor is my first "actual robots" investment (w/ Raghu). Complete, operational, robots w/ hundreds in production.
While being off the scale technical, the team is obsessive about pragmatic approaches to making the end to end economics work.
Very excited to be partnering.
martin_casado a relayé @a16z
We’re thrilled to lead Tutor Intelligence’s Series B.
@tutorintel is a robotics research and deployment company that builds generally capable robot hardware. They deploy into real production environments ahead of full autonomy, using human teleoperators (“tutors”) to train the
W
martin_casado a relayé @wavefnx
It now has LLM, internet, web search access and can show images, call tools; the responses would normally be over a speaker (not connected yet)
You don't get to hear my magnificent voice here but you do get a screen transcript. Mic/STT model seems to be working cleanly
O
Stefan Roth a relayé @olvr_hhn
📢 LucidLyra: Feed-Forward Generative 3D Scene Reconstruction and Joint Understanding via Multi-Task Distillation!
Generate a queryable 3D Gaussian scene from a single image! Our feed-forward decoder jointly predicts geometry, appearance, open-vocabulary semantics, and object
K
@rahul how's things with you at Flock? You have not posted in some time, so checking what's new
ה
Aviad Dahan a relayé @FinancialEduX
זה גרף די מדהים. הוא בחן 5000 סטארטאפים במטרה להבין מי מהם הצליח לגייס כסף מקרנות הון סיכון בשורה הראשונה.
הפרמטר החיובי הראשון: עובדים לשעבר של אותן קרנות
הפרמטר החיובי השני: עובדים לשעבר במעבדות המובילות
הפרמטר החיובי השלישי: בוגרים של הטכניון/אוניברסיטת תל אביב 🇮🇱 🇮🇱 🇮🇱
Liping Bai 白莉萍 a cité @capital_vmv
This is the lesson I am learning in the Swedish context as well. Many norms in Chinese working environment, such as building a record of the progress of the project, cc direct report on all correspondance, speak rather candidly on issues etc, doesn't fit western style.
S
martin_casado a relayé @sporadica
Twitter is great but today a plane brutally crashed killing like ten people and all I’ve seen is mutuals making real dumb, poor-taste jokes or memeing about it
I get that most things aren’t sacred but I figured at least mass loss of life tragedies would be
Is it because you
W
Kosta Derpanis a relayé @karpathy
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
- Excellent
S
Kosta Derpanis a relayé @SteveOrsini
Waterloo’s IP policy let's students and faculty own their IP. Its value is reinforced by the university’s commercialization expertise, entrepreneurship programs and industry relationships. Waterloo’s seventh-place global ranking in 2026 is a strong indicator of its ability to
B
Kosta Derpanis a relayé @buildcanada
As of today until March 14, the North American time zone map looks like this.
🌎
P
Kosta Derpanis a relayé @docmilanfar
Historical tenure rates in EE/CS across some well-known universities:
Caltech EAS (~100%)
Carnegie Mellon (~80%)
Pennsylvania (~75%)
Harvard/Yale (~60%)
Berkeley/UCLA EECS (~66%)
Stanford & MIT EECS (~50%)
Statistically speaking, at Caltech's Engineering division, the initial
R
Kosta Derpanis a relayé @ziv_ravid
New episode of The Information Bottleneck with Julian Togelius @togelius about AI and games 🎮🕹️🎲
Julian (NYU professor, co-founder of) has been working on AI for video games since 2004, back when people told him it wasn’t real research. Basically every
M
Kosta Derpanis a relayé @drmichaellevin
Claude made some visuals for my process of paper-writing, useful for students. In case anyone else wants to use them:
(from older blog post)
I’ve received a few inquiries about #CVPR2027. Just to clarify, I’m not currently part of the social media team.
I know, shocking 🤓
martin_casado a relayé @poteto
Day 28/ tibo's game has landed
i spent $5.99 on this domain, but i had a blast making it with grok bot and cursor! i hope you enjoy it and find all the easter eggs 🤍
if i did it right you should be able to play it inside of X too
emiran a cité @cagrimbakirci
evrim agacinin cevap aradigi soru: "insanlar maymundan geldiyse simdiki maymunlar neden insan olmuyor?"
Douglas Gray a cité @ManaByte
AI recommending Measure of a Man seems to be a bit of a conflict of interest, no?
B
martin_casado a relayé @BarrettLyon
For anyone running their own Linux server, you can also give a try and protect your exposed ports behind our mesh
B
martin_casado a relayé @BarrettLyon
@martin_casado We recommend anyone running a Linux server give a try and protect your exposed ports behind our mesh
U
Kosta Derpanis a relayé @prof_kamilov
We're hiring in Physical AI at UW–Madison!
The ECE department is recruiting faculty at all ranks, with interests spanning multimodal sensing, physical modeling, robotics, and autonomous systems.
Full consideration by Dec 14, 2026.
I
Kosta Derpanis a relayé @ir0armeni
Proud of @liyuan_zz for World from Motion and its acceptance to #NeurIPS2026. With @shalinidemello, @alextrevith et al., an NVIDIA collab.
It reconstructs dynamic 3D scenes from monocular video by conditioning a video model on 4D Gaussians, filling in what the camera never saw👇
Finding an RCE in the Linux TCP stack is some 1990's shit. Wow.
Spelunking through kernel code just stopped being worth the effort with all the other attack vectors up the stack. But with AI, it's basically free.
So we're dredging up monsters lurking in the deep.
Aviad Dahan a relayé @deedydas
I redid the startup founder signals that lead to top tier VCs investing in the seed / A, after incorporating some feedback.
This added ex-frontier lab and cyber security near the top, IIT as a positive, and educated in China as a negative.
The main changes were;
- extended
F
Kosta Derpanis a relayé @fhahlbohm
With all the progress that's happening, anything but 4D YouTube in 2027 would be a disappointment. Unfortunately, most pipelines are closed source. Here's a teaser of how I'll try to do my part. Quite happy with how well extrapolation works.
(yes, this model is only 20% trained)
Halloween for a recovering chocoholic is a tough day 🫣
W
Kosta Derpanis a relayé @engineers_feed
5 megabytes of storage in 1956:
D
martin_casado a relayé @dok2001
@martin_casado Cyber has changed the narrative to AI as the hero. There is no other way to defend our critical infrastructure.
Is it just me or has a lot of the AI policy chatter died down? Perhaps having concrete issues to deal with (cyber) has taken the existential edge off much of the discourse?
. @bot is now the most used app on my phone and what I access most other apps from.
I need embodiment ideas ... thinking Holly from Red Dwarf. Any other thoughts?
Great to see this sort of work … pragmatic exploration of systems risks.
N
Kosta Derpanis a relayé @nico_dufour
Neural Thickets (@yule_gan @phillip_isola): add random noise to an LLM's weights, keep the best copies, let them vote. Reasoning jumps, no training.
Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains 🧵
R
emiran a relayé @beynimyokmalim
Kadıköy metrosunda halloween partisine giden kızın yanında monster cantasıyla otururken
Z
Kosta Derpanis a relayé @ZimingLiu11
There is a "bug" in commonly used activation functions (ReLU/SiLU/GeLU etc.) ---- The mean of post-activations is not zero.
This bug, however small it seems, leads to pathologies of training dynamics, including loss spikes & representation overshooting.
By removing the bug,
P
Kosta Derpanis a relayé @docmilanfar
Many of you are in the process of applying for jobs in academia and/or in industry.
As you prepare your job talks, consider giving it the following structure:
First, a short snippet of your overall set of technical accomplishments. For instance, you might spend the first 15
I wonder why Carm, Lace 2 and Phantom fights are feeling so great, that I am fine to lose them to experience indefintely, but Trobbio is so annoying.
#SilkSong
A
Aviad Dahan a cité @pika_labs
Really cool to see multi-stem generation taking off.
SepGen does a similar thing and supports localization for an immersive scene generation
A
Aviad Dahan a cité @pika_labs
Really cool to see multi-stem generation taking off.
SepGen does a similar thing
S
Kosta Derpanis a relayé @sewon__min
Good slides are extremely intentional. When someone presents a paper, the paper already has all the information. But the speaker would make deliberate choices about things like, what to say and what not to say, what to put on a slide and what to say only verbally.
What to frame
T
Kosta Derpanis a relayé @tallinzen
my hypothesis for why AI-generated slides for technical presentations are so awful:
people who don't have a lot of experience presenting think that to make a talk you just put all of the stuff you want to talk about on the slides, then you go through each slide and say the stuff
S
Kosta Derpanis a relayé @lamsheeper
Can we trace an AI model's answers back to its training data?
Surprisingly, even when we know exactly which documents taught a model a fact, attribution methods often fail to find them.
Our CATT benchmark investigates why and uncovers two simple fixes🧵
Kosta Derpanis a relayé @randall_balestr
The Platonic Representation Hypothesis is now a Theorem ◼️
Train JEPAs separately, on any modalities: they provably learn the same representation up to one rotation.
The result? A 109M JEPA-Text model that aligns with vision models better than 72B LLMs.
w/ @ylecun 🧵
S
N
Kosta Derpanis a relayé @gnam1410
🧵[1/4] Most 3D models assume a pinhole camera. What about a fisheye, GoPro, or 360° panorama?
Excited to share OmniRay: 3D reconstruction from a single image from any of these cameras, no intrinsics needed.
w/ @jai_abhayk, Cheng Peng, Rama Chellappa, @anand_bhattad
P
Kosta Derpanis a relayé @phillip_isola
Several recent papers find that you can ~align image and text embeddings with just a rotation.
@randall_balestr and @ylecun released a cool demo of this here:
I tried it on a photo of my cat Moxie. I have to say, it is completely accurate!
N
Antoine Guédon a relayé @nico_dufour
Neural Thickets (@yule_gan @phillip_isola): add random noise to an LLM's weights, keep the best copies, let them vote. Reasoning jumps, no training.
Why does it work? It's sampling in disguise. Swap noise on billions of weights for one random vector on the logits: same gains 🧵
Liping Bai 白莉萍 a cité @NSF
Roboticists, the next generation of wireless networks also involves you, specifically, sensing with communication signals.
The 80 / 20 rules is still very much in effect. Accept the 80% has become so dazzling, no one wants to work on the 20.
Douglas Gray a cité @hammer_mt
This insight is why so many software engineers seem depressed recently.
They can build anything they want now in 5% of the time, but it’s harder to make money because everyone else can do it too.
B
martin_casado a relayé @xbh_artist
What if your office had a portal to a sacred world?
I built this on my Apple Vision Pro
Walk through the liquid portal → step into the sacred ruins
Huge thanks to Ian (@XRarchitect) for the amazing portal reference and examples. Such a creative mind, this wouldn't exist
Kosta Derpanis a relayé @juliarturc
Over the past few days, I dug into the geometry of LoRA to understand how exactly it constraints the geometry of the search space. There's a beautiful connection to the *intrinsic dimension* of a task.
Also learnt I'm saying "difficulty" wrong... 🤡
0:00 MNIST vs the walking
Douglas Gray a cité @anchiever
Isn’t this the recipe that leads to AI agents committing computer crimes?
martin_casado a cité @wavefnx
Dude, so sick. Can you throw into github with a parts list so I can build too?
W
martin_casado a relayé @wavefnx
Yea, board and screen arrived yesterday, CAD cover and Rust firmware prototyped today
Mic, WIFI incl. TLS connections, bluetooth etc
A lot of work went also to cryptography/security so nothing is stored plaintext and not even the cpu can read your root key
Firmware test:
B
martin_casado a relayé @BrittaniNatali
Had to make a video showcasing my real-time animated PBR splats. Made with Radix and Opus 5.5.
A silent CV bug: swap RGB and BGR, and YOLO11s recall fell from 67.9% to 57.4% on COCO128. No crash.
Google says Gemini 4 Carbon can tackle long debugging tasks; we haven’t tested it. My checklist for any fix: reproduce, inspect the diff, retest on held-out data.
F
Kosta Derpanis a relayé @fhahlbohm
Ever since Thomas’ JIT-fusion update for Instant-NGP, I was wondering how fast a fully fused Zip-NeRF could be. Finally gave it a shot with Opus:
36 min training, 300 ms per megapixel on a 5090, same quality as the official code (~6-7x/10x faster).
Code:
U
Liping Bai 白莉萍 a relayé @NSF
NSF invests more than $110 million to improve next-generation wireless technology through the first-ever NSF Verticals-enabling Intelligent Network Systems investment. 🛜
Andrew Davison a cité @alexkoch_ai
This is apparently not fully autonomous (a remote human is supervising; not sure how much) but if it's real it's still some of the best robotics I've seen.
emiran a cité @queenofeuropa
akademik manipulasyonlardan gercekten gina geldi. @hmelisaacar yetmezmis gibi simdi de alican kiraz. yapmayin birader, zaten ulkenin adi cikti:
Why Fewer CPU Threads Made YOLO 2x Faster (ONNX Runtime)
In Ultralytics’ benchmark, limiting PyTorch to two threads cut the time to process 1,000 images from 58.8 s to 27.1 s on an 8-vCPU machine. Same model, no retraining.
PyTorch handles preprocessing and postprocessing while
K
Kosta Derpanis a relayé @wkentaro_
SAM3 now runs offline on your GPU.
An ONNX export of Meta's SAM3 that runs on NVIDIA CUDA and Apple CoreML. Besides text, it takes several example boxes, negative boxes and clicks as prompts.
A text prompt takes 0.37 s on an M5 MacBook Pro and 0.08 s on an RTX 3080 Ti.