@vg_head

Drawing things on the screen @Rayonapp

 Romania
minus-ze.ro
Joined January 2022
  • Tweets287
  • Following450
  • Followers84
11Media
If there's one thing I've learned after performing OCR on a huge corpus, it's that "classic" OCR pipelines like Tesseract are NOT fully replaced by VLMs. I've combined PyMuPDF4LLM with Tesseract and it corrected a huge number of mistakes from the VLM. Both have their advantages.
1
2
44
In my case I used Mistral OCR 4.1. I suspect it's a VLM but obviously I don't know for sure.
55
IMO people saying the decision endpoint announced by OpenAI recently kills Jev are missing this point. 100% synthetic data is no joke, I'm not convinced frontier AI labs have anything equivalent.
Diogo Almeida
@CompleteSkepticSep 17
Replying to @badlogicgames
you might be the first person talking about the data over the architecture! 🥲 we consider ourselves a data research lab! the vast vast vast majority of research was on making data that is truly general (ala a cognitive core) and 100% of our data is synthetic (but not the type of crap that is just spit out from an LLM obviously)
1
67
The price and latency also make me think that it's impossible to replicate without the same data quality.
1
17
Either way, competition is good. I just wish we'll get the research around how they generate the data at some point. It's super intriguing.
15
Intelligence is solved, human beings are not. Fix incoming.
Roelof van Dijk
@roelof_vandijkSep 8
We could have all woken up to “OpenAI and Anthropic team up, solve Navier-Stokes cleanly, fully formalized but also here is a nice explanation by Tao, here’s how we can all happily improve math together” but alas, the world does not yet work that way.
3
159
Alexandru Ică retweeted
Eurora Labs@euroralabsApr 27
Eurora Desktop already has the context. No copy and paste. No context-setting. No explaining. Just ask. One assistant. Every browser on every desktop across macOS, Windows and Linux. Articles. Videos. Docs. Threads. Just ask away
1
1
2
133
Great post on spectral rendering, and a wonderful blog in general: momentsingraphics.de/Spectra…
1
173
Alexandru Ică retweeted
Ben Visness@its_bvisness28 Oct 2025
The write-up of my new graph layout algorithm for SpiderMonkey is now live. We built a custom layout algorithm for JS and WASM that follows the structure of the source code. No more spaghetti nightmares from Graphviz, and thousands of times faster.
17
72
10
755
54,534
One thing I love about digging into rabbit holes such as PDF parsing is stumbling upon phrases like "schizophrenic PDF files".
1
2
109
The source is a highly interesting read: hal.science/hal-02143706/fil…
49
Fun fact: new ICC profiles can contain embedded programs in them that specify color space transforms. All operators describe operations performed on the stack, à la Forth.
1
143
"calculatorElement" and Annex F in this doc: color.org/specification/ICC.…
124
I wrote a new post about one of the weirdest optimizations I've done at one of my previous jobs. Hopefully it's an enjoyable read. :) minus-ze.ro/posts/unintuitiv…

Unintuitive Optimization For Performing Paths Union

There is a weird optimization possible when one wants to compute the contour of a lot of vector paths. Here I try to explain why it works.

minus-ze.ro
1
1
238
The code for computing winding for a quadratic Bézier can be simplified a lot by dealing with monotonic quadratics only. Testing if a sample lies to the left/right of the curve becomes extremely cheap in that case.
1
2
185
shadertoy.com/view/433czr
75
"Quadratic Approximation of Cubic Curves" contains a neat method of representing one cubic Bézier as two C1-continuous quadratics, maintaining tangents at the endpoints. Combining that with a heuristic for how many times to split the source curve gives excellent results.
1
3
8
893
Demo: editor.p5js.org/AlexandruIca… Original: ttnghia.github.io/posts/quad…
1
1
202
I ported my demo to Rust to see if I can make things faster. The goals is to perform queries for arbitrary points and arbitrary vector paths, as fast as possible. I'll try to write everything SIMD-friendly.
1
4
148
I've implemented non-zero fill for quadratic curves without flattening them. Should be numerically robust but I haven't tested it too thoroughly yet. Very happy with the results so far. Coverage is computed per pixel individually.
1
1
94
Demo: editor.p5js.org/AlexandruIca…
1
59
Load more