Yanis Boucherk

Just retweeting and shitposting lmao

[Yanis Boucherk @ygboucherk]

@fuzziphy 6 days after warranty expiry is diabolical.
Good they still honored it though :)

[Yanis Boucherk @ygboucherk]

#keepandroidopen #OpenSource #OSS #AOSP

[Yanis Boucherk @ygboucherk]

Daily reminder that OneUI 9.0 applies the latest Google restrictions to apk sideloading.
These updates are automatically installed by the following packages: com.wssyncmldm, https://t.co/JQRUqC5J1p.soagent, https://t.co/CmwZyyF0pf.updatecenter

[Yanis Boucherk @ygboucherk]

You can disable them with the following commands (once the device is connected with ADB)
adb shell pm disable-user --user 0 com.wssyncmldm
adb shell pm disable-user --user 0 https://t.co/JQRUqC6gQX.soagent
adb shell pm disable-user --user 0 https://t.co/CmwZyyFyeN.updatecenter

[Leo · Building w/ AI @leo114119]

Deepseek v4.1 Flash 1017 发布以后，Dario 公开信

简洁大白话版：

1.我们致力于将自由从“平民”手中剥夺，以充实我们自己。
2.我们阻断了“小人物”接触生物学，这样我们的生物技术项目就能垄断AI。
3.我们启动了恐惧宣传和危险色情活动。
4.我的公关公司让我写了一封信。
5.不惜一切代价保护IPO。我的190亿美元很重要。
6.我们继续攻击中国，因为DeepSeek太便宜且太有帮助了。

他怕了他怕了。。。。。。

[DeepSeek @deepseek_ai]

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!

Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.

🔗 https://t.co/JGAik1kgk0

1/4

[Yanis Boucherk @ygboucherk]

@zane_os

[Yanis Boucherk @ygboucherk]

@teortaxesTex

[Yanis Boucherk @ygboucherk]

@ItsmeAjayKV @deepseek_ai

[Yanis Boucherk @ygboucherk]

@NomadsGalaxy @Keurig It's checking the quantum structure of every caffeine molecule for defects.
You wouldn't want a molecule to have a slightly different wave function and accidentally kill you :/

[Yanis Boucherk @ygboucherk]

Idk what @pidotdev is doing but it's GOOD!

[Yanis Boucherk @ygboucherk]

Nevermind, it exists (though it seems outdated)
https://t.co/oPletcY4Y8

[Yanis Boucherk @ygboucherk]

Why is every mainstream pareto frontier about capability vs. api cost instead of capability vs. RAM requirements?
Like, if we think about it, qwen3.8-27b is quite expensive on OpenRouter (dense model => more compute) but way easier to run locally than some comically sparse MoE.

[Yanis Boucherk @ygboucherk]

@22AlaaMuhammad @GeminiUpda0kkb Haiku?

[Yanis Boucherk @ygboucherk]

@deepfates If only there was models open enough to allow distillation.
Assuming they allow distillation, why wouldn't they directly publish their own weights, allowing US-based providers to host them (potentially charging API rates).
But those are only an utopia, are they?

[Yanis Boucherk @ygboucherk]

@AlexGodofsky Or we just get trillions of "tokens" just by existing in a physical world.
We tend to assume that training data = verbal corpus of stuff taught to us while forgetting that "overheard someone talking behind me" or "held a strange material" also teaches us something, verbal or not.

[Yanis Boucherk @ygboucherk]

@Pipe42uk @peter_szilagyi But it flagged 59.94Hz, which is a(n oddly specific) framerate in video formats inherited from NTSC, rather than 60Hz.
The camera/coils might make some noise when taking a frame.

[Yanis Boucherk @ygboucherk]

IQ2_XS quantization on stock llama.cpp v0.5.0, running on debian 13 (headless minimal install)

[Yanis Boucherk @ygboucherk]

ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO hitting ~100tok/s prefill and ~10tok/s decode on a 16GB Quadro P5000 (159k context), while retaining good agentic performance (tested in VSCode Copilot + OpenClaw).

[Yanis Boucherk @ygboucherk]

@PhotonPhalinx @Math_files There's a difference between "an infinite number of decimals" and "an absurdly high, finite, number of decimals".
Also, 0.99999... = 1 can be proven with no numerical approximations using geometric series.