This was my experience in grad school, and now I've seen some evidence to suggest a trend 🤔

Honored to present my PhD work at Ai2! The lack of data is one of robotics’ biggest hurdles — here’s how we’re tackling it in tactile manipulation and rehab robots. Watch the talk 👇 youtu.be/Qfq0BT6En4E?si…

Nice work!

Artificial neural networks were invented as models of the brain and and now rapidly approach human-level cognition. Does this mean that the mystery of human intelligence has been solved? I grapple with this question here. dendwrite.substack.com/p/what-remains…

Do you know about Open Geospatial Solutions? The Open Geospatial Solutions (opengeos) GitHub organization hosts a collection of open-source geospatial software projects. The projects are developed by a community of geospatial software developers and researchers Their collection

Are neural networks Turing complete? This excellent blogpost by colleague Hessam Akhlaghpour goes carefully over the limitations and flaws of much of the literature for the case of transformers: lifeiscomputation.com/transformers-a… Note that the first proofs of Turing completeness (to my knowledge)

This week's #PaperILike is "Reality Promises: Virtual-Physical Decoupling Illusions in Mixed Reality via Invisible Mobile Robots" (Kari & Abtahi, UIST 2025). This is some Tony Stark level stuff! XR + robots = future. Website: mkari.de/reality-promis… PDF: mkari.de/reality-promis…

xLSTM for robotic manipulation systems via diffusion-based imitation learning: arxiv.org/abs/2510.20406 PMP leverages an xLSTM to denoise actions for robotics. “PMP not only achieves state-of-the-art performance but also offers significantly faster training and inference.”

Blog link: szymonozog.github.io/posts/2025-10-…

one day you will look back and hate that you stayed in a system you already knew was broken, right now you still have time to choose — to defect instead of comply, to build instead of wait, to live instead of defer, you can always go back to school later, you can’t rewind the

Advice to students, young engineers and inquisitive amateurs. Every major project I do, until and including today, follows this pattern, and never does the fear leave. ============ 1) Can you do "thing x" ? 2) No. 3) Go to ntrs.nasa.gov , download all papers pertaining

is qwen3 doing depth-wise layer upcycling? i'm looking at average (over tokens) router probabilities for qwen3-30b-a3b and seeing this weird pattern where for highly active experts, its also likely for the same expert index but 12 (and often 24) layers down to also be highly

We dive deep into every part of the stack that we didn't build. Here's a deep-dive into (potentially undisclosed?) MoE depth wise up-cycling that Qwen3 does. I have a lot of Qwen folks that follow me, would anyone like to clarify :)

At ICCV I met people working on Image Matching who still don't know our image-matching-models GitHub repo It allows you to use almost any IM model just by changing one parameter Image matchers, try it out, it will save you countless hours github.com/alexstoken/ima…

🚀Another thing that needs to be quoted is the upcoming Multi-Token Prediction (MTP) by Meta (arxiv.org/abs/2404.19737) and speculative decoding, which are rapidly becoming de facto components of modern LLMs. SPECTULATIVE DECODING IS COMPUTE-BOUND This trend suggests that

What is SIMD and how to use it by Anılcan Gülkaya medium.com/@anilcangulkay… Similar to the quoted post on designing a SIMD algorithm from scratch, this article provides all the necessary elements to start with SIMD.

If you’re excited by Tesla’s new world model, meet OmniNWM—our research take on panoramic, controllable driving world models • Ultra-long demos • precise camera control • RGB/semantics/depth/occupancy • intrinsic closed-loop rewards Arxiv: arxiv.org/abs/2510.18313 Watch:

Video lectures, Maryland UMD CMSC351 Introduction to Algorithms, by Mohammad Hajiaghayi youtube.com/playlist?list=…

We build world models grounded in physical reality, naturally generalizing across cars and robots 🤖

🤯 Merging many finetuned LLMs into one model, effectively? Introducing Functional Dual Anchor (FDA), a new framework for model merging. 🚀 Current merging works poorly due to the underlying parameter conflicts. FDA shifts knowledge integration to the input-representation space