←
Valentin Ignatev14,245 posts
Valentin Ignatev avatar

Valentin Ignatev

@valigo

programmer, speedrunner, father, immigrant alt: @valigo_ru youtube.com/@valigotech | github.com/valignatev | twitch.tv/valigo

◷ Joined June 2022▧ 1,680 Photos and videos
Valentin Ignatev avatar
Valentin Ignatev@valigo·Oct 26

There are many ways to kill performance in a multithreaded program. My favorite one is false sharing, because you can get 10x performance just by moving some bytes around. So let's look into it. 00:00 - When multiple threads is not even faster 04:23 - CPU L1 cache primer 09:40 - Getting ALL the performance 15:35 - Looking at disassembler 20:29 - How to find cache line address 22:38 - false sharing in disassembly 25:00 - closing words

26:06
Valentin Ignatev avatar
Valentin Ignatev@valigo·59m

A lot of people forgot that the main reason people were historically switching from Windows to Linux is because Linux is free. It was never about merit. Now Linux desktop is better than Windows, but only because Windows got so much worse. This was never the case before. Linux desktop was always built on layers of duct tape and prayers, and required a lot of niche knowledge to keep operational, and somehow people pretend that it's not.

Valentin Ignatev avatar
Valentin Ignatev@valigo·2h

let goyim play gta 6 at 15fps on PS5 I'm gonna enjoy it at 165fps 1440p in 2 weeks after it's fully decompiled

Valentin Ignatev avatar
Valentin Ignatev@valigo·5h

Slop on slop violence is something we've never seen in gamedev before 2026. What a year

Post attachment
Valentin Ignatev avatar
Valentin Ignatev@valigo·7h

I'm really tired from oneshot demos going viral if I be honest. They are very shallow and don't represent real work you have to do to ship something good and useful, and are not at all how you work with models in real life. I'm also worried if models are being "optimized" for oneshots and if it hurts their long-term capabilities. I don't have any proof that it happens, but it's just a concern on a back of my mind.

Valentin Ignatev avatar
Valentin Ignatev@valigo·9h

A lot of people dunk on this, but I gotta say this is some great target audience research by OpenAI marketing team.

OpenAI@OpenAINov 17

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
Valentin Ignatev avatar
Valentin Ignatev@valigo·19h

2026 is the best year for gaming since 2007 in my opinion. 2007 is the last year I remember to have such a high density of unfathomable bangers in such a volume you can get few years of sustenance off of it. And a lot of these games will be remembered decades after.

Valentin Ignatev avatar
Valentin Ignatev@valigo·21h

Astra 6.2 arrived. I didn't have early access to it and I haven't been testing it for a while. I think it's a model by OpenAI. I won't be sharing more.

Valentin Ignatev avatar
Valentin Ignatev@valigo·Nov 17

I'm experimenting with custom UI library and tried few approaches for not rebuilding UI tree every time (my UI library is immediate mode). I implemented SolidJS-like signals with Jai metaprogramming, so that the compiler literally inserts these signals when you write through a pointer and it's actually working. But then I started prototyping text layout with HarfBuzz shaping and doing text absolutely dwarfs any kind of convoluted UI layout optimization you can imagine, so at the end of the day I think rebuilding the tree every frame your UI code is reached is totally fine, even on mobile devices. Damage-based rendering is different, I'm pretty sure damage-based rendering is always good.

Valentin Ignatev avatar
Valentin Ignatev@valigo·Nov 17

If you work in a team - you really want all your team mates to you the same model (or family of models from the same vendor) for the project. Reason being - different models have different success and failure modes, which you should be aware of, and base your AGENTS.md around it. If you want appeal to authority - Andrej Karpathy talks a lot about needing to understand the model before building around it, and obviously he knows about LLMs more than you or me.

90% of success when coding with AI comes from the model. Harness really doesn't matter much and should mostly stay out of the way. They only time when harness starts to matter is when you code with dumb models (e.g. anything other than Opus 5.5/Fable 5.2) Ironically enough, this means that Opus and Fable are worse with Claude Code than without it. Because Claude does stay in the way A LOT.

Valentin Ignatev avatar
Valentin Ignatev@valigo·Nov 17

Of course, you still need to give model a high quality input data, but this is a separate topic that's a bit outside of model-harness axis.

Valentin Ignatev avatar
Valentin Ignatev@valigo·Nov 17

90% of success when coding with AI comes from the model. Harness really doesn't matter much and should mostly stay out of the way. They only time when harness starts to matter is when you code with dumb models (e.g. anything other than Opus 5.5/Fable 5.2) Ironically enough, this means that Opus and Fable are worse with Claude Code than without it. Because Claude does stay in the way A LOT.

Valentin Ignatev avatar
Valentin Ignatev@valigo·Nov 17

It's still crazy to me how "paste without formatting" is not the default. Surely I really mean to paste colorful multi-font abomination of a text into email body! Of course, who wouldn't want that!

Valentin Ignatev avatar
Valentin Ignatev@valigo·Nov 17

Nir is one of my favorite low level programming and tech content creator started actively posting here, so you should all follow him!

Fun fact: Windows 95 introduced a new shell called "Windows Explorer" and still included the Windows 3.1 shell "Program Manager" (progman.exe). In fact, it was only removed in XP SP2! betawiki.net/wiki/Program_Ma…

Post attachment