@simonw

–

29 July 2025 21h32

GLM 4.5 Air produced code that worked first time, but it also uses ~48GB RAM where the new Qwen used ~30GB

ChatGPT Study Mode System Prompt Extraction Analysis

By

–

29 July 2025 21h29

The new ChatGPT "study mode" feature appears to be entirely implemented as a carefully crafted system prompt – thankfully OpenAI mostly don't take measures to protect those these days so it's easy to extract it and see how it works

Running Quantized LLM Locally on Mac with LM Studio

By

–

29 July 2025 20h59

This one drew me a pretty cute pelican, it even has a smile! I got it working on my Mac too, using an 8bit quantized model via LM Studio – notes here: https://
simonwillison.net/2025/Jul/29/qw
en3-30b-a3b-instruct-2507/
…

Space Invaders Clone Generated by 3-Bit MLX GLM-4.5 Air

By

–

29 July 2025 16h08

Here's the Space Invaders clone my 2.5 year old laptop wrote itself using 3 bit MLX GLM-4.5 Air, from the single prompt "Write an HTML and JavaScript page implementing space invaders" https://t.co/DVDLcLnoTI pic.twitter.com/53ildnkObA
— Simon Willison (@simonw) 29 juillet 2025

Here's the Space Invaders clone my 2.5 year old laptop wrote itself using 3 bit MLX GLM-4.5 Air, from the single prompt "Write an HTML and JavaScript page implementing space invaders" https://
tools.simonwillison.net/space-invaders
-GLM-4.5-Air-3bit
…

MLX Builds for GLM-4.5 Air 3bit Quantization

By

–

29 July 2025 15h40

Great work @ivanfioravanti on the MLX builds of GLM-4.5 Air – I used the 3bit one for this

GLM 4.5 Air 3bit MLX Running Impressively on Mac

By

–

29 July 2025 15h08

I got an MLX 3bit version of GLM 4.5 Air running on my 64GB Mac and WOW this is an impressive local model!

GLM-4.5 Air 3bit MLX runs smoothly on MacBook M2

By

–

29 July 2025 7h40

I've run the GLM-4.5 Air 3bit MLX build on my 64GB MacBook M2, it's impressed me – it drew me a simpler pelican https://
huggingface.co/mlx-community/
GLM-4.5-Air-3bit
…

Context Rot and Token Efficiency in Large Language Models

By

–

29 July 2025 4h53

Yeah context rot is about bad tokens (mistakes etc) getting into the context and causing poor performance later on – what Theo is describing is more models straight up wasting tokens

Dedicated Image Video Generating Models Show Strong Performance

By