Mark N. Bastourous @marknabil

Bill Freeman gives us first a list of warm-up bitter lessons. He keeps the bigger ones for later in the talk. #cvpr2026

الحياة مؤخرا

If you're interested in SSD internals and how to use them efficiently, our paper, “How to Write to SSDs,” has been accepted to VLDB and is currently on the Hacker News front page. vldb.org/pvldb/vol19/p1…

We used to go to a special website, ask strangers for help with programming, and get humiliated in return

Terence Tao - "AI tools are like taking a helicopter to drop you off at the site. You miss all the benefits of the journey itself. You just get right to the destination, which actually was only just a part of the value of solving these problems." Judit Polgar - "I always felt that intuition is very important in chess, but I get my intuition through my experience. And many times I think that this is the biggest danger for youth, that they don't have the experience because they don't spend enough time doing." Elites from two different fields voice the same opinion. [1] theatlantic.com/technology/202… [2] archive.is/mv2FB

Back ache 😭

Khoa Vu @KhoaVuUmn

a month ago

"We used to spend hours trying to deciphering a simple LaTeX error."

18 305 4K 90K 126

1/ Deep learning is going to have a scientific theory. We can see the pieces starting to come together, and it's looking a lot like physics! We're releasing a paper pulling together these emerging threads and giving them a name: learning mechanics. 🔨 arxiv.org/pdf/2604.21691 🔧

Les données disponibles sur data.gouv.fr sont désormais interrogeables via un serveur MCP dédié en experimentation, vos retours sont bienvenus ! 💻 Le code est ouvert et accessible sur GitHub : github.com/datagouv/datag… Pour en savoir plus : data.gouv.fr/posts/experime…

𝐕𝐢𝐬𝐮𝐚𝐥 𝐛𝐥𝐨𝐠 on Vision Transformers is live. vizuaranewsletter.com/p/vision-trans… Learn how ViT works from the ground up, and fine-tune one on a real classification dataset. CNNs process images through small sliding filters. Each filter only sees a tiny local region, and the model has to stack many layers before distant parts of an image can even talk to each other. Vision Transformers threw that whole approach out. ViT chops an image into patches, treats each patch like a token, and runs self-attention across the full sequence. Every patch can attend to every other patch from the very first layer. No stacking required. That global view from layer one is what made ViT surpass CNNs on large-scale benchmarks. 𝐖𝐡𝐚𝐭 𝐭𝐡𝐞 𝐛𝐥𝐨𝐠 𝐜𝐨𝐯𝐞𝐫𝐬: - Introduction to Vision Transformers and comparison with CNNs - Adapting transformers to images: patch embeddings and flattening - Positional encodings in Vision Transformers - Encoder-only structure for classification - Benefits and drawbacks of ViT - Real-world applications of Vision Transformers - Hands-on: fine-tuning ViT for image classification The Image below shows Self-attention connects every pixel to every other pixel at once. Convolution only sees a small local window. That's why ViT captures things CNNs miss, like the optical illusion painting where distant patches form a hidden face. The architecture is simple. Split image into patches, flatten them into embeddings (like words in a sentence), run them through a Transformer encoder, and the class token collects info from all patches for the final prediction. Patch in, class out. Inside attention: each patch (query) compares itself to all other patches (keys), softmax gives attention weights, and the weighted sum of values produces a new representation aware of the full image, visualizes what the CLS token actually attends to through attention heatmaps. The second half of the blog is hands-on code. I fine-tuned ViT-Base from google (86M params) on the Oxford-IIIT Pet dataset, 37 breeds, ~7,400 images. 𝐁𝐥𝐨𝐠 𝐋𝐢𝐧𝐤 vizuaranewsletter.com/p/vision-trans… 𝐒𝐨𝐦𝐞 𝐑𝐞𝐬𝐨𝐮𝐫𝐜𝐞𝐬 Dr @sreedathpanat Videos on ViT ViT paper dissection youtube.com/watch?v=U_sdod… Build ViT from Scratch youtube.com/watch?v=ZRo74x… Original Paper arxiv.org/abs/2010.11929 Next up: demystifying Low-Rank Adaptation (LoRA) in PEFT! Follow me @Mayank_022 along for more deep learning insights, cool fine-tuning projects, and updates from the upcoming blog posts.

My biggest open-source release! NumKong — 2'000+ SIMD kernels for mixed-precision numerics, from Float6 to Float118. Started in 2023. Opened the PR in 2024. Finally, merged this week! RISC-V, Intel AMX & AVX-512, Apple SME & SVE, WASM Relaxed SIMD. 200'000 lines of code in a 5 MB binary. Same scale as OpenBLAS. Available for C 99, C++ 23, Python 3, Rust, Swift, GoLang, & JavaScript. Int4 dot products via nibble algebra. Ozaki Float64 GEMMs on Float32 tile hardware. 6-bit and 8-bit floats back-ported to 10-year-old CPUs. 5'300x faster Geospatial metrics than GeoPy. 200x faster Kabsch than BioPython. 0 ULP where OpenBLAS hits 56... and a lot more! pip install numkong Or pull it from NPM, Crates, GitHub... and let me know what breaks 🤗 Links & highlights ⬇️

Tom and I have finally finished a draft of Dynamic Programming Vol 2! Exhausting but satisfying. New approach to DP theory, advanced material, many applications... dp.quantecon.org

O3DE eliminates the middleman. With our native ROS2 Gem, you get direct integration—no bridges, no extra lag, no headache. ✅ Direct ROS integration ✅ Modular Gem architecture ✅ Ready-to-go ROS project templates hubs.la/Q044GCKk0

A small Qwen3.5 from-scratch reimplementation for edu purposes: github.com/rasbt/LLMs-fro… (probably the best "small" LLM today for on-device tinkering)

Claude solved one of Donald Knuth's conjectures from writing "The Art of Computer Programming" and he wrote about it. www-cs-faculty.stanford.edu/~knuth/papers/…

😋 Want strong LLM reasoning without breaking the bank? We explored just how cost-effectively RL can enhance reasoning using LoRA! [1/9] Introducing Tina: A family of tiny reasoning models with strong performance at low cost, providing an accessible testbed for RL reasoning. 🧵

Useful update to PRH! It's great to see conversation moving past "do models converge" to "*which* structures converge." This paper argues that only local structure converges. Matches my intution (its why we like mknn), but hadn't realized extent of the bias in CKA.

Maria Brbic @mariabrbic

4 months ago

Are neural nets across modalities really converging to the same representation as they scale, as the Platonic Representation Hypothesis suggests? We show that common representational similarity metrics are confounded by network width & depth. We propose a permutation-based

12 90 595 70K 461

New project! Flow Policy Gradients for Robot Control tldr; a simple online RL recipe for training and fine-tuning flow policies for robots co-led w/ @redstone_hong: hongsukchoi.github.io/fpo-control

VLAs and World Models go together like peanut butter and jelly! GigaAI took their already awesome GigaBrain-0.1 VLA and used a world model for reinforcement learning to improve it. They took 10,900 hours of teleop data (a mix of sim and real) to train the model. They also have human in the loop to correct the VLA when it starts to veer off course. The result? GigaBrain-0.5M, a new model that may just find itself on top of the RoboChallenge leaderboard. This is an exciting future for robotics. World models, Sim2Real, frontier VLAs. All going to play a pivotal roles in 2026 and beyond. Learn more about GigaBrain-0.5M here: gigabrain05m.github.io #Robotics #PhysicalAI

The real gap isn't capability, it's accessibility. We need platforms that labs can actually build, hack and improve without needing Big budgets or NDAs. Something modular, documented, cheap and yet capable enough to conduct hours of research . We present you YOR

Mahi Shafiullah 🏠🤖 @notmahi

4 months ago

Why buy a robot when you can build your own? Meet YOR, our new open-source bimanual mobile manipulator robot – built for researchers and hackers alike for only ~$10k. 🧵👇

7 22 169 38K 89

I've been saying for years that the biggest challenge for simulation in robotics is not actually the physics engine (although you do have to get that right). The real challenge is capturing the *diversity* of the real world. There was no doubt that generative AI had the potential to change that, but it's still amazing to see it take shape. Watching Nick's incredibly fast progress has convinced me that content generation might not actually be a bottleneck anymore. This is a beautiful combination of hardened tools for e.g. low-level mesh processing with the latest tools for generative asset creation, wrapped in a powerful agentic workflow. Please do give it a try and share your feedback.

Nicholas Pfaff @NicholasEPfaff

4 months ago

Meet SceneSmith: An agentic system that generates entire simulation-ready environments from a single text prompt. VLM agents collaborate to build scenes with dozens of objects per room, articulated furniture, and full physics properties. We believe environment generation is no

18 81 569 74K 308