Paul McGee

Atheist. Maths, Delphi, C++, Cat theory, FP ... in that order.
And now Rust I guess. Yes. Rust. Fine.

[Paul McGee @logical_canine]

@kirkwasgood @Victor_Rg31 @lauren_w67 Career would have been over???
You read the news at all?
Dudes are in power who've done a panoply of disgraceful shit.

[Anung Un Drama @spaceghost]

This is the coolest thing I’ve ever seen from a YouTube creator and it’s not even close. The quality of the animation is insane. I can’t believe the madlad @MoistCr1TiKaL actually pulled this off.

[Cruz @ViC305]

8GB RTX 4060 Laptop. ~16GB Ornith IQ3_S GGUF. 44.5 tok/s streaming locally.

FreeToken didn’t have the qwen35moe GGUF path needed for this model, so I implemented it and submitted PR #131 upstream. 👀

I also exposed the K/I GGUF quant types its CUDA kernels already supported through the Python loader.

This is the result.

Setup:

Ornith-1.5-35B-A3B
35B MoE / ~3B active
IQ3_S GGUF · ~16GB

RTX 4060 Laptop · 8GB VRAM
FreeToken · FlashInfer
MoE backend: offload

On the attached full-prompt run, I measured ~44.5 tok/s streaming on screen.

In the same runs, FreeToken’s server decode counter reported 46.7 → 50.1 tok/s.

VRAM stayed at 6,879 / 8,188 MiB, with 20GB host RAM and 98% GPU utilization during generation.

Load time: 65s.

My llama.cpp CPU reference on the same model was 11.1 tok/s.

The interesting part isn’t just “35B on an 8GB GPU.”

The code path itself was missing.

FreeToken’s CUDA side already supported more GGUF quant types than its Python loader exposed, including the K/I quants needed for this IQ3_S model.

And qwen35moe models like Ornith needed their own GGUF architecture/loading path.

I implemented both pieces in my fork and opened them upstream in PR #131.

Once loaded, FreeToken can use GPU + CPU + host RAM as an elastic MoE inference system instead of requiring the entire checkpoint to live in VRAM.

That fits Ornith well because only a subset of its experts are active per token.

Important caveat:

PR #131 is still open. This is not merged upstream yet.

I’m not calling this official Ornith support.

I’m showing a working implementation, the code is public, and the attached terminal video is the actual run.

PR #131:
https://t.co/uqZkjjaWAT

My fork:
https://t.co/Dcr4fruj4A

Next I’m going to test where this path holds up across more quants, longer context and real agent workloads.

[fintex @_yusufknl]

As someone who ships LLM systems in production, this vector calculus video is the closest thing to a "the actual math behind every gradient descent step in Claude" explainer I've ever seen released for free.

Everyone thinks "gradient descent" is a fancy AI term. It isn't. The gradient in gradient descent is literally the ∇f operator this video builds from scratch - the exact 1800s calculus equation OpenAI and Anthropic use to update every single weight in GPT-5 and Claude.

Bookmark this and watch tonight. Same nabla operator from 1800s calculus, now running billions of times per second inside every frontier LLM training cluster.

[MTS @MTSlive]

Google researcher @lauriewired warns the shift from C++ to Python in universities is creating programmers who don't understand how computers actually work:

"When I was in college getting my CS degree, my year was one of the last years that was using C++ as the core language for teaching students. Actually they were switching over to Python of all languages."

"One of the biggest issues that I have is that a lot of these higher level languages are really abstracting away what is happening under the hood. If you're not aware of how the computers work in general, then you're gonna have problems when it comes to debugging."

"Starting with a language like C++ gives you the ability to do these really high level abstractions, but it also gives you the ability to customize your memory management and get really close to the underlying machine. Especially if you're trying to work in high performance computing, game development, or high frequency trading."

"At C++ conferences, there's almost this fundamental anger response when it comes to mentioning something like C or C++. People start complaining about the complexity of it."
@Google

[Yan Holtz @R_Graph_Gallery]

I LOVE micro-interactions like this ❤️

No extra buttons, no clutter, so the risk of hiding the story is ~0.
Yet it invites exploration and makes the data feel alive.

With AI removing the tech barrier, shouldn't this be the norm?

Graph by J. King from https://t.co/25xAYT9jBU

[Massimo @Rainmaker1973]

Aidako Shisho is a Japanese artist who creates amazing sculptures out of cardboard.

[📹 aidakkoshisyo/nami_aidakko]

[Quanta Magazine @QuantaMagazine]

The week before Christmas 2025, five mathematicians were holed up at ETH Zurich with a singular goal: perfecting a solution to one of the biggest open problems in percolation theory. By Christmas, they had the proof. https://t.co/VGYpIliAEA

[Anung Un Drama @spaceghost]

The perfect woman doesn’t exis—

[Marko Denic @denicmarko]

Perfect weekend project doesn't exi...

[Bryce, the CUDA Colonel @blelbach]

These are just... algorithms. Why are we calling them AI algorithms?

[Mr. Turtle @TheMisterTurtle]

I won’t tell anyone when I make it, but there will be signs.

[Christoffer Lernö @nuoji]

Guess who just got rejected for a compiler job due to ”not enough production compiler experience”.

Man, that was one brutal criticism of the C3 compiler.

[Paul McGee @logical_canine]

@the_aiju Consider a not-so-arbitrary set of k+1 horses ...

[Paul McGee @logical_canine]

@WilliamPrevidi @JoJoFromJerz Adolescents might be a little old for them ... 😢

[Paul McGee @logical_canine]

@rastshawn @sarahbla3k @ContraPoints CameHereToSay PascalCase FTW

[X SHORTS 🎬 @thrillerhub]

This is what being a preschool teacher is really like. 😂

Ashley was fighting for her life; I’ve never seen her break soo hard. 😭

[Paul McGee @logical_canine]

He knew a pencil is a monoid ... 😄

[lova @CFC_lova]

Movie 🎥 📺📺📹