Sabitlenmiş Tweet
Grok 4 vs GPT-5 as a mobile qa agent: Grok 4 matches GPT-5 in general-purpose UI interactions, and completely beats it in 3D space understanding
English
Slava Gilevich
1.1K posts
Slava Gilevich
@__int32
Building next-gen QA tools for the Cursor age @qalti_ai
Katılım Aralık 2010
130 Takip Edilen1K Takipçiler
Found a shader named cubemap_copy.wgsl. Deleted it on reaction as I thought I accidentally copied the cubemap.wgsl...
Turns out it was a shader for copying a cubemap :)
English
@FrameworkPuter @AIatAMD Mature? Yeah, sorry, no. Spent 2 weeks trying to port Qwen3-TTS to RyzenAI NPU runtime with every possible trick I know to no avail. @LisaSu can you at least bump the open-sourced version with new kernels?
English
We have 128GB Framework Desktop configurations in stock, and the inference software stack on @AIatAMD is now mature.
Basic Apple Guy@BasicAppleGuy
Apple removed the 256GB memory option for the M3 Ultra Mac Studio. Only configuration is now 96GB.
English
@gleb_alexandrov it's not a synthetic data, they collected their own dataset using sodium light and sodium light filter in a prism: youtube.com/watch?v=UQuIVs…
YouTube
English
That is brilliant and fantastic. A hard problem that can't be solved via different means; the machine learning trained on synthetic data; EXR output; open-source; presentation - I like everything about it.
GitHub Projects Community@GithubProjects
Green screen keying, solved at the pixel level. Corridor's neural keyer unmixes edges into true foreground color and linear alpha, EXR out.
English
@anemll Amazing! I was hoping to squeeze at least 1tps from M1 Ultra 64Gb, but hit a wall at 0.1tps due to preloading going into swap and messing things up. Will try to do a direct read from SSD this weekend.
English
~ 6.5 - 6.7 t/s for GLM 5.1 on M5 Max 128GB
Added “Dense” model export, now model load is only 5s ! Experts are streaming from SSD, so we do not pre-load it.
Added direct SSD->Slot memory path, removed prefetch... Many dead end experiments.
See Export a “dense-only GGUF” and “Fast path ” in tools/flashmob-sidecar/README.md
WIP branch for Flash-MoE-SSD
github.com/Anemll/anemll-…
English
@eugenebokhan And two chips that support cooperative tensor properly are m4 and m5, yes? Other chips ‘support it’, but only two latest apple chips gain actual acceleration with cooperative tensors?
English
4/ But buried in the Metal Shading Language Spec there's a function on cooperative_tensor called get_multidimensional_index(i). It returns the logical (row, col) for element i. It works on every Apple Silicon generation. Here' the recipe for running MPP on any chip.
English
@styropyro_ @Winterrose Nobody: …
Styropyro: boy, do I love the smell of fresh Helium-3 Z-pinched plasma in the morning!
English
@Winterrose for what 4000 car batteries would cost, i think there is more i could do with just 2000, and then use the leftover funds to buy copper, SCRs,....
...and deuterium
English
I want to find a way to get this kid 4000 car batteries
English
@ManyATrueNerd This is exactly how the current AI works. It spews out perfectly written code, paragraphs of in-depth on any topic, but when you look inside, it’s just a bunch of math, and nobody reeeally knows why and how it works.
English
If you literally have a machine that's calculating the end of the world to the minute, and you believe this machine, then you know the factors that end the world are data in the machine.
So... maybe look into the machine? Like run some tests? Pull some data out of it? Anything?
English
Ehh, I don't really love how House is written in Fallout Season 2.
Writing 'the cleverest person in the world' is always difficult as it's asking writers to write a character who is vastly smarter than they are, but even then I find this take on House as a futurist weird.
English
Slava Gilevich retweetledi
We’re excited about our upcoming partnership with Razer! The companions are now physically on your desktop! They can see, hear, and speak with you, and even see your screen.
R Λ Z Ξ R@Razer
Meet Razer Project AVA, your all-in-one AI companion: rzr.to/AVA From planning your day to analyzing spreadsheets and game starts, Razer Project AVA leverages advanced AI inferencing and reasoning that dynamically evolves based on your personal interactions. Select your 5.5" companion from an expanding library of characters from esports legends to custom anime-inspired Razer designs. Be among the first to sign up for Razer Project AVA at release. #CES2026Razer #RazerCES2026 #CES2026
English
@Luckyballa My solution in the MLS-MPM implementation is to use sort + atomic. I have a case when I need to scatter over 3x3x3 cube of voxels. The sort improves data locality a lot, and atomics outperform 27 scatters with later reduction over 27x data
English
I don’t know why, but I find the scatter-reduce problem very entertaining
The naive solution is simple, but it costs you massive atomic contention
Traditional warp reduction doesn’t work since your data is scattered with an irregular pattern
But there is always a way
English
Every time the government vows to make something affordable it’s a sure bet its price is going to go up significantly. Tesla is probably the only exception where a subsidy didn’t make cars more expensive in the end
English
@Nick_Davidov There is another factor to this. With extremely high housing prices and 30-50 year (!) mortgages, you are almost forced to think in terms of stable income, and you have to have at least half a year in savings to afford any experiments of making something yourself.
English
I will never understand why people obsess over Elon controlling $750B net worth and are ok with politicians controlling bigger amounts which they did not create.
English
@Nick_Davidov School/uni makes a huge bias towards avoiding failure at all costs, and making things is risky.
At least, it made in my case, and I’m still beating this crap out of me.
English
@__int32 Maybe just being angry at the world is easier than trying to make things and try to break out. Even though making things and persevering is highly satisfying and anger is just self destructive
English
@RoszykAdam Did tou get access to camera stream, or images are loaded from the iPhone?
English
As far as I'm aware this is the very first gaussian Splat trained locally on Vision Pro
English
@lisyarus You want to add earth rotation effect, where atmosphere has a slight movement from west to east , and it would generate vortices after land masses, aka hurricanes:)
English
Improved my precipitation algorithm - now it doesn't always rain, but only when water vapour content is above threshold - which means I can visualize clouds! Honestly I didn't expect it to look THAT realistic 🤩
#indiedev #gamedev #indiegames
English
@Luckyballa @ionthedev I’ve always thought that you can just convert Gaussian splats into SDF with zero being some variable threshold, so that the big splats result in larger volume?
English
@ionthedev A couple of newer methods (without gaussians) can produce mesh or it’s easy to extract
But even with gaussians you can generate a proxy meshes too, multiple papers did it, it’s coarse and ugly, but works for some case
MrNeRF@janusch_patas
MeshSplatting: Differentiable Rendering with Opaque Meshes Contributions: (i) An end-to-end optimization of mesh-based scene representations retains visual quality while training 2× faster than current state-of-the-art methods. (ii) Rather than a polygon soup, we generate a connected mesh by refining the vertex locations of a restricted Delaunay triangulation. (iii) Triangles are naturally connected to each other, and quantities stored within vertices are smoothly interpolated across each triangle. (iv) The optimization is aware that the triangles should be opaque, allowing direct high-quality rendering in standard game engines (see Fig. 1), opening the door for classical techniques like the use of depth buffers and occlusion culling [1, 22].
English
@twominutepapers Here’s implementation of the similar approach (bar the whirlpools), a full blown MLS-MPM running 60 fps at 1M particles on my iPhone!
English
Previous simulators saw whirlpools vanish, but one genius fix made them 30X more stable! 🤯 These particles remember every twist. Cleaner weather predictions, quieter jets - what a time to be alive! Full video: youtu.be/rRMlhHDCNr0
YouTube
English