Keith Duke

I feel like this photo of masked, armed men pepper spraying a pastor protecting his community is going to be a defining picture of this moment in America for a long, long time.

I've spent so long in the "web" world I haven't really looked at what exists elsewhere in the software ecosystem.

As someone now immersed in LLMs and language, and a grammar nerd, this made me happy

www.nltk.org

He lost me shortly after he said "I'll be quick. I promise." and then went on and on about it. He toiled on that piece for sure...

what if I told you that taking accountability for your actions was traditionally considered a load-bearing feature of masculinity?

That's so cool! Thanks for sharing!

“Terrible things are happening outside. Poor helpless people are being dragged out of their homes. Families are torn apart. Men, women, and children are separated. Children come home from school to find that their parents have disappeared.”

Diary of Anne Frank
January 13, 1943

Good to know, I haven’t had a use for it but sounded interesting

IBM Granite-Docling: End-to-end document understanding IBM released Granite-Docling-258M, an ultra-compact, open-source vision-language model (VLM) for converting documents to machine-readable formats while fully preserving layout, tables, equations, code...

Tried Granite Docling?

www.ibm.com/new/announce...

Holding a copy of Gödel, Esher, Bach outside Green Apple Books in San Francisco I purchased today.

Score

Don't get me started.

@anthropic.com many people here would LOVE if you posted this stuff here too. Amazing work!

Why aren’t they here?

Waffles?

By, roundtrip a prompt, read, reprompt, response, I mean tools/function calling

Wow, this thing is fast! I've been using Qwen 3 4B Thinking, granted a thinker, but to roundtrip a prompt, read, reprompt, response took a while on my machine. But not Granite, this thing is fast. I mean way faster in use. Going to keep tuning around this version for a bit. Really like it so far.

They convert to MLX for Apple silicon with the latest mlx, I have the Micro version strapped in to a framework. It’s faster than Qwen 3 4b, may be my new go to.

Dropped it into the framework I had used for granite 3.

First thoughts, it's fast, much faster than 3. Responses are better, working out tool calling as its changed a bit.

So far really liking it, once I get it all strapped in will really test it out, so far looks good.

Update MLX, latest handles conversion no problem. Stats from running 4.0-h-micro convert to 8 bit using same prompt as before, answer was much more verbose. Testing it in a framework next

Prompt: 14 tokens, 12.835 tokens-per-sec
Generation: 256 tokens, 26.793 tokens-per-sec
Peak memory: 6.636 GB

A screenshot showing the following text. % python g4test.py 7:20:27 Fetching 12 files: 100%|██████████████████████████████████████████████████████████████████████████████████████| 12/12 [00:00<00:00, 133505.70it/s] ========== The sky appears blue because of a process called Rayleigh scattering. As sunlight reaches Earth's atmosphere, it is scattered, or redirected, in all directions by the gases and particles in the air. Blue light is scattered more than other colors because it travels in shorter, smaller waves. This scattered blue light is what we see when we look up at the sky. ========== Prompt: 65 tokens, 64.747 tokens-per-sec Generation: 82 tokens, 79.240 tokens-per-sec Peak memory: 7.248 GB

Won't convert to MLX format, it's also MoE and MLX doesn't support those layers yet.

There is a tiny-preview that's older on HF that has been converted. Pulled it and kicking the tires a little. Prompt given was "Why is the sky blue"

Granite 4.0 Language Models - a ibm-granite Collection We’re on a journey to advance and democratize artificial intelligence through open source and open science.

IBM released a new Granite model yesterday, going to convert to mlx here in a few and try it out. Its mamba based.

huggingface.co/collections/...

Why limit yourself? There are several options. Variety is the spice of life or some such

X training grok on its own version of wikipedia might help...

If they push it hard enough, all the numbers will point to 88

(solemnly) You've had mail.

TextWorld - Microsoft Research TextWorld is text-based game and is an open-source, extensible engine that generates and simulates text games-use it to train reinforcement learning agents.TextWorld is an open-source text-based game ...

Oh nice! Definitely will check it out. I read an interesting article about playing text games with agents. Led me to this from MS, playing Zork sounds cooler lol. www.microsoft.com/en-us/resear...

I find bsky locks up safari. The tab becomes unresponsive.

It has been linked to cancer as well, since the 80s. There is no safe amount to drink.

I used to drink daily, but stopped a few years back. Was one of the best things I did for myself.

granite-docling-258M demo - a Hugging Face Space by ibm-granite This app converts images of documents into structured formats like Markdown or LaTeX and answers questions about the content. Users can upload images or select from sample images and ask questions ...

258M, this blew my mind
huggingface.co/spaces/ibm-g...

Hahahaha that was a good one! Omg thanks, made my morning