Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ

@bartwr.bsky.social

Engineering, Computer Graphics, Art, DSP, ML Culture, Techno, Industrial, and Electronic Music. Research Scientist at NVIDIA. Ex Google Research, Ex games (Sony, Ubisoft, CD Projekt). Politically leftist. He/they. https://linktr.ee/bartwronsk

2.2K followers459 following896 posts

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @peark.es

85% inference margin, training costs are well below that. Really remarkable.

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @andrewwillmott.bsky.social

Anything that can be edited by artists, iterated, and "non destructive" (not baked) is IMO always going to win long term. Artist friendliness, control, iteration... "Claude Code but helping artists" is 1000x better than even a higher quality model but that is frustrating "regenerate until I like it"

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @bartwr.bsky.social

Also, coding, hacking, and customizing are becoming the fastest way to get things done, and we have entered an era of 100% personalized software, everywhere. Google for a tool and configure it? Why, when I can create new UX + modify a project's source code to get *exactly* what I need in less time?

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @kcimc.bsky.social

Your takes often capture my thoughts 100% accurately (though much more concisely and precisely than when I try to put them into words); I 100% agree and had similar thoughts with 5.5 + Fable. I "brainstorm", "direct", and "architect", and a model does coding way better and faster than I ever could.

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @kcimc.bsky.social

i also have this sense that the generating code is becoming an inseparable feature of daily life. it's so obvious it's not even worth mentioning. "artist working with code" is like saying "artist working with physical matter".

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @kcimc.bsky.social

i'm not "working with code" anymore. i'm thinking in systems, and writing in natural language. i've only been programming for 25 years. any code i write is going to be vastly inferior to a frontier model.

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @segyges.bsky.social

love to see journalists openly admit that they are perfectly comfortable lying about the things that i already know they lie about, by the way.

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

this was early enough that it was very insightful and is now sort of transparently true if you think the gizmo isn't useful you're ignoring most major software project and all computer security professionals, you basically have to spit on expertise as a concept

โšก๏ธ๐ŸŒ™@dystopiabreaker.xyz

antivax movement but itโ€™s left coded and targets machine learning, artificial intelligence, and engineering fields

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

The โ€œhead of strategic futuresโ€ at OpenAI just went on an unhinged rant against social security sometimes I feel like these tech companies are actively trying to alienate as many normal people as possible

Yet another frontier open model, possible to run locally with a multi-GPU setup. Even if AI was 100% regulated/banned, even if datacenters were banned, there is no coming back. Anyone can keep running those. No wonder OpenAI + Anthropic are working hard to regulate those open models - why pay them?

Unsloth AI@unsloth.ai

GLM-5.3-Flash, also known as ox-alpha, is out now! Zai's GLM-5.3-Flash is a new 320B parameter multimodal open model with 18B active parameters. GLM-5.3-Flash approaches Claude 4.8 Opus on coding and agentic benchmarks. GGUF coming soon!

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

I'm excited to announce a first development release of nanobind 3.0. It introduces *split mode*, a new distribution model where a single Python wheel runs across all Python versions, while achieving outstanding performance and gaining rather than losing features! ๐Ÿงต

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Hopefully one last SIGGRAPH advertisement. I, myself, yours truly, will present on Haolin's behalf a data-driven method for joint adaptive sampling and quadrature for integration. Wed R408A. In the end of the talk I'll argue why everyone is doing the wrong thing. Paper: suikasibyl.github.io/nqr

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

I published vibrance-0.1.0 on Maven, now with support for Compose for Desktop (JVM). The APIs let you mix colors yourself, or draw gradients as modifiers or Compose brushes. Next steps: refine the model training to improve some color transitions, better alpha support. github.com/romainguy/vi...

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

my #1 grizzly beard protip for (game) dev is: if you have a bug that you dont understand the cause of immediately, dont try to fix it; add & iterate visualization / inspection tools until the cause is plain to see. the accumulation of 'observability' in your system's state will pay off many times

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

god i hate programming. always have. always will. damn if you can't do some neat stuff with it though.

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

personally i'm ok with AI techniques being less well known but there's a deeper thing going on here which is far more important IMO, because it's also partially why LLMs have taken over == this thread is in response to this tweet: == x.com/krismicinski...

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Implicit surfaces are everywhere in vision and graphics applications that optimize shapes. If we adjust the surfaces through a "learnable" deformation field, how can we ensure they stay close to the undeformed state (e.g. a template)? Our latest paper offers a simple and effective ARAP regularizer๐Ÿงต

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

Replying to @mworchel.eurosky.social

The code, including handle deformation and stylization demos, is available here: gitlab.com/tobidju/arap... Or, just run `pip install iarap` if you only want to use the regularization term. Thanks to @tobidju.bsky.social for leading the project + coauthors @ugoalone.bsky.social and Marc Alexa.

gitlab.comTobias Djuren / ARAP-Regularization-for-Implicit-Surfaces ยท GitLabSource code for the paper 'As-Rigid-As-Possible Regularization for Implicit Surfaces' at SGP 2026.

Scapegoating AI (HBMs!) for the DRAM prices is so silly and the "same old" story over again. Corps and price fixing cartels getting away with their stuff thanks to "useful idiots" (sanctimonious luddites). Real systemic issues -> convenient clickbaity scapegoat -> we'll never get a systemic change.

Reminder that memory manufacturers (including Samsung, Hynix, and Micron) were found to be an actual dictionary definition "cartel" back in the 2000s and were accused of illegal price fixing again in the late 2010s. They have been trying to screw you for literal decades.

Bart Wronski ๐Ÿ‡บ๐Ÿ‡ฆ๐Ÿ‡ต๐Ÿ‡ธ reposted

A new article: Revisiting Yliluomaโ€™s ordered dither algorithm 30fps.net/pages/revisi... It's a deep dive to a single "color selection algorithm" used for ordered dithering. Nothing groundbreaking but I did uncover a neat simplified version that works OK. It's technical but with visual examples.