Bart Wronski 🇺🇦🇵🇸

GitHub - dsharlet/array: C++ multidimensional arrays in the spirit of the STL C++ multidimensional arrays in the spirit of the STL - dsharlet/array

A "healthy" alternative (though some would complain about cache ;] ): strided array views. Just return a strided view - no allocations needed. User can "compact" it if they want.
E.g. github.com/dsharlet/array a numpy-like multidimensional array powerhouse from my friend and a former coworker.

GitHub - dsharlet/array: C++ multidimensional arrays in the spirit of the STL C++ multidimensional arrays in the spirit of the STL - dsharlet/array

A "healthy" alternative (though some would complain about cache ;] ): strided array views. Just return a strided view - no allocations needed. User can "compact" it if they want.
E.g. github.com/dsharlet/array a numpy-like multidimensional array powerhouse from my friend and a former coworker.

Happy to announce: ReSWD. Sliced Wasserstein Distances are quite powerful, but they perform a Monte Carlo (MC) integration (over random directions). During an optimization this can lead to noisy gradients due to variance.

Project page: reservoirswd.github.io

Computing the exact bijection of the optimal transport (OT) problem between very large point sets is completely untractable…

In our SIGGRAPH Asia 2025 paper: “BSP-OT: Sparse transport plans between discrete measures in log-linear time” we get one with typically 1% of error in a few seconds on CPU!

The Three Wise Men A philosophy webcomic about the inevitable anguish of living a brief life in an absurd world. Also Jokes

The Three Wise Men -
existentialcomics.com/comic/622

Graphics Programming weekly - Issue 410 - September 28th, 2025 www.jendrikillner.com/post/graphic...

YouTube video by Aaron Hertzmann Picture Perspective and Our Eyes

Here's a recording of my talk on how perspective works! If you're interested in learning about how picture perspective works in human vision, this is the video to watch. #visionscience
www.youtube.com/watch?v=eamc...

#SIGGRAPH2025 Advances in Real-Time Rendering in Games course talks slides posted: "Adaptive Voxel-Based Order-Independent Transparency" shipping in the upcoming Call of Duty: Black Ops 7 by Michal Drobot from Activision Central Tech is now online: advances.realtimerendering.com/s2025/index....

Congratulations! :)

There seems to be a phenomenon on Bluesky where if you talk about AI, even critically, in a way that isn't obviously disparaging, you're marked as an AI bro and blocked. The result is that leading AI ethicists are now "AI Bros" 🙄

Reading this as I am rewriting a Python project to be 100% asyncio-based.😅
But I don't disagree! Harder to debug, harder to profile, there is overhead...
But also really nice to write - and read - for things that mix UI, disk operations, some slow background processing... 100x better than callbacks.

YouTube video by Branch Education How are Billions of Nanoscopic Transistors Made into a Microchip? EUV Photolithography 🛠️⚙️🤯

Insanely well-made video on EUV photolithography: www.youtube.com/watch?v=B248...

NNs are ok and find some close-to-optimal local minima because of crazy overparametrized wastefulness leading to a super smooth loss landscape with infinite close-to-optima.

Everyone who tries differentiable programming encouraged by the success of NN training learns this lesson the hard way (so did I!). Most "functions" of a few+ variables, even when 100% differentiable, are so non-convex that getting optimal parameters is impossible from random start. :(

For optimizing a post filter, why do you need *all* rendering inputs being differentiable, do you need gradients of everything? Only things processing optimized variables need differentiability. (Maybe I misunderstood the application)
But as always - SGD might not be able to optimize them...

This looks absolutely fantastic! LLMs have great educational value and this is one of the best examples I have seen (as not replaceable by more "low tech" tools without existing total expertise)

The question is: what do you need it for. :) It generally works with some limitations, but for primary visibility or materials, you can assume it 100% solved. But its usefulness depends on application. Minor geometry tweaking? Great! Full scene reconstruction - forget - SGD itself is not enough.

Ok, kind of makes sense. I was replacing it myself and paying for it, but it was a while ago and I think the prices went up like crazy for that. On my previous phone I replaced it once after 1.5y, and now after 3y it was again getting unbearable (heavy use = 2x charges a day :( ).

That said, I absolutely cannot understand people who buy a new phone every year. Even ignoring the cost and the environmental impact, the hassle of setting it up (FU Microsoft Authenticator) is more than any truly perceivable benefit.

Got the new iPhone Pro - my usual 3 year phone cycle. After customizing the ugly and messy initial glass look (those home screen bubbles ugh), feels great! Not a groundbreaking update, but worth it. The screen, snappiness, new cameras and formats, styles, camera button (FINALLY!), USB-C. Good buy.

I think the H1Bs were broken, gaming the system, and the minimum required salary should be at least doubled.
(Based on personal experience: I didn’t win the literal H1B lottery)
But this is a) extortion, and b) destroying lives of existing H1B holders.
It will destroy any US skill-based industry.

Part that I missed yesterday: this is $100k extortion *per year*.
And for existing H1B holders as well! Basically making H1Bs useless for anyone but $1mm+ employees (that will likely use O1 anyway).
And it’s an application fee: so if they don’t approve it - *poof*. Money burned.

I think the H1Bs were broken, gaming the system, and the minimum required salary should be at least doubled.
(Based on personal experience: I didn’t win the literal H1B lottery)
But this is a) extortion, and b) destroying lives of existing H1B holders.
It will destroy any US skill-based industry.

The tariffs were the first step into destroying the US trust score. This shatters it.

These folks have the same progress goals as Pol Pot. A return to the good old days of an agrarian economy.

YouTube video by ZY Cheng Why Photographic Styles Look So Good on iPhone 16

Ok this YouTube video does explains these filters. I am now a fan:

youtu.be/6tLk40RkgVE?...

My absolutely favorite paper in this space I recommend to anyone working on reconstruction or inverse tasks (or even any machine learning): arxiv.org/abs/1711.06077 "Perception-Distortion Tradeoff". Goes beyond quantifying the error/distortion of the reconstruction statistically, 2/

I also love the irony on how "perception" does not depend on the perceptual senses or our brain interpretation of signals, but is simply a statistical distribution divergence. Which is practical and actionable, no matter which divergence function is used (though some are better/worse). 4/4

and clearly visually demonstrating why pure MSE or MAP are not enough, but also discusses formally why it is at odds with the "perception" or "realism" of the reconstruction and the inherent trade-off. It was an eye-opener for me and changed the way I think about any inverse problems. 3/

My absolutely favorite paper in this space I recommend to anyone working on reconstruction or inverse tasks (or even any machine learning): arxiv.org/abs/1711.06077 "Perception-Distortion Tradeoff". Goes beyond quantifying the error/distortion of the reconstruction statistically, 2/

I agree with the "spirit" of your observation (using information theory etc.), but not the specifics - capture Nyquist is not the limit! A simple example: deconvolution of the point-spread function can go significantly beyond it. For the agreement part, we can/should definitely analyze the error. 1/