Nilay

Christian Rupprecht explains their interpretability research in 3D computer vision, testing if (and where in the model) multi-view transformers like VGGT, DepthAnything 3, and DUSt3R use point/patch correspondences to make sense of 3D scene geometry.

• be Torakusu Yamaha • the son of a low-ranking samurai astronomer in 19th-century Japan • obsessed with Western machines, you make a living repairing watches and medical equipment • 1887: a local elementary school has a broken American reed organ. Nobody in the small town knows how to fix it. • you take it apart, realize it’s just two broken springs, and easily repair it • but instead of just handing it back, you realize: "If I can fix this, I can build it." • you draw a blueprint of the inside of the organ and build the very first Japanese-made reed organ from scratch • you show it off. People tell you it sounds terrible. • most people would quit. You sling the heavy wooden organ over your shoulder on a bamboo carrying pole. • you physically carry it 160 miles (250 km) on foot, trekking over the brutal Hakone mountains just to reach the Tokyo Music Institute to get real feedback from experts • the professors play it. They tell you the mechanics are brilliant, but the tuning is completely wrong. • you don't get defensive. You stay in Tokyo for a month, sitting in on university music theory lectures, holding a single tuning fork to your ear until you completely master the mathematics of sound frequencies • you walk 160 miles back home • you build a second organ. The professors test it and declare it "as good as those from abroad." • you found Nippon Gakki Co. (which later becomes Yamaha Corporation) • you decide to make your company logo three interlocking tuning forks to remember the pain and discipline of learning music theory from scratch • decades later, your company uses its piano woodworking expertise to build wooden airplane propellers in WWII • after the war, the company uses its new metallurgical expertise from the airplane engines to build motorcycles • you accidentally create a timeline where repairing a broken elementary school organ directly leads to the creation of the Yamaha YZF-R1 superbike • absolute, relentless horizontal integration based purely on figuring out how things work The ultimate testament to reverse-engineering reality.

@OjasSharma276 What exact features from cpp14/17/20/23 do you miss?

In my company’s project, we use C++11. This was the first version of C++ considered modern. The current latest version is C++23. I asked my team members why we were still using such an old version and why we couldn’t migrate to newer ones. The answer I got was simple: migration isn’t easy. The existing codebase is heavily optimized around older standards, and moving to a newer version means re evaluating and testing everything again. At the same time, the team has to maintain and develop the product, so doing a full migration alongside active development is extremely difficult. Most tech companies still use older versions of Java, C++, or other languages for this exact reason. Migration is the real problem.

A common social problem is redirected attribution. Joe does some great work. He gains a high status. Jack arrives, he says that he is with Joe. He gains a high status... but he doesn't do any of the great work that Joe does. Publicly, this works remarkably well for Jack, the usurper, because we don't like to speak ill of the high status individuals. But the private reputation is something else. Your private reputation is what people say when they are not afraid of being overheard. When dealing with very high status individuals or organizations, seek the private reputation.

@AMDGPU_ You think they don't map deviceID to things like OS/Mac address/myriad of other serial numbers available on a PC?

Each GeforceNow or cyber cafe GPU can be captured by the steam hardware survey hundreds of times for individual user instances..... Which shows up in the hardware survey as tens of millions of Nvidia GPUs but all purchased by Nvidia.

NV buys its own GPUs for steam to artifictially dilute AMD in the hardware survey: - GeforceNow - tens of thousands of Geforce GPUs running steam. - Sponsored Cyber cafes - hundreds of thousands of geforce GPUs running steam.

@VishakhRanotra Weather, oceanography data - TBs being streamed from satellites and post processed into gridded HDF5 then sent to farms with TBs of memory and xeon processors running complex fluid mechanics models

In what scenario does one require 1TB of RAM?

@ashisharma_21 Agreed 👍

A lot of us are bullish on the long-term story of India But living in Indian cities today means dirty neighbourhoods, chaotic traffic, terrible air and a passive lifestyle Maybe the only reason is the broken local governing system. Our federal structure is our biggest liability

New blog post! In "Billions of triangles in minutes" we'll walk through hierarchical cluster level of detail generation of, well, billions of triangles in minutes. Retweets welcome! Link in reply because algorithm.

"see that in the grad norm histogram? yeah we just hit a batch of 4chan greentext"

You Can't Manufacture a NeRF - an ICML position paper from CMU that discusses what it means to generate a 3D object that can be manufactured. Nothing too controversial and feels a bit more like a survey- but good reference to have for explaining this concept 🙂

🤨

@ashisharma_21 @IndianTechGuide @AskPerplexity Perplexity se kya puch raha hai bhai, tujhe hi 23% tak le jana hai 😃

@IndianTechGuide @AskPerplexity can you share a credible article that explains this in detail?

🚨 India plans to increase the share of the manufacturing sector from 12% to 23% over the next two decades: FM Nirmala Sitharaman.

Also engineers who write CAD software 🐥

If you want to know more about how Google Flights works, airline tickets, and why it is super complicated to deal with all the constraints and the combinatorial combinations, I highly recommend this set of slides by Carl de Marcken, one of the co-founders of ITA software, which Google acquired and became one of the underpinnings of Google Flights. (Sorry for the http rather than https Link: Carl's domain doesn't appear to support https) demarcken.org/carl/papers/IT…

1/5 This 13813 lines of Templated Insanity is a hashtag#HLSL shader, not hashtag#CUDA and not hashtag#cpp. Browse the Godbolt link lnkd.in/es74NmZj and witness: 1. a recreation of C++17 <type_traits> 2. some bits of <functional> 3. declval<T> linkedin.com/feed/update/ur…

1/ "There are decades where nothing happens; and there are weeks where decades happen." At present, nowhere does this quote feel more apt than in the field of AI. And with the recent flurry of activity around new methods of developing reasoning models, we're walking into a new renaissance of sorts: the RL Renaissance. This thread dives into the newly forming RL paradigm for model improvement, and explores how decentralization might be leveraged to further it 🧵

🤯 this new Mechanistic Interpretability research from Anthropic is utterly mind blowing wish more people were talking about it Anthropic discovered that Claude has several emergent capabilities reminds me a ton of Michael Levin's idea of "simple systems, able to do surprising things with regard to cognitive behavior" (except these systems are not so simple) insights in thread 🧵 👇 x.com/AnthropicAI/st…

Small achievement but yesterday was the first time ever I wrote vulkan pipeline+ synchronisation code for my renderpass and it ran without validation errors on first attempt 😀

Introducing VGGT (CVPR'25), a feedforward Transformer that directly infers all key 3D attributes from one, a few, or hundreds of images, in seconds! No expensive optimization needed, yet delivers SOTA results for: ✅ Camera Pose Estimation ✅ Multi-view Depth Estimation ✅ Dense Point Cloud Reconstruction ✅ Point Tracking Project Page: vgg-t.github.io Code & Weights: github.com/facebookresear…

TI often points out one of the largest perf issues of today's game authoring: perf/tri tanks with sub-pixel geo. This is true of both MSAA and scaling-TAAs and RT. The visible to cull|traverse data ratio gets small (horrible).