Harris Oldroyd

I like taking AI apart to see how it works.

I run models on my own hardware, test ideas, and write up what I find.

Explore the writingSee the experiments
Inside my local AI workstation, with two RTX 3090 graphics cards

My local AI workstation

Two RTX 3090s, 48 GB of VRAM.

See the build

Explore

How AI works

Tokens, attention and language models.

Run models locally

Hardware, software and getting started.

Performance benchmarks

Measured performance on real hardware.

Selected work

Quantization in pictures

Change the precision of a photo and see what happens.

Try the explainer3 min read

Tracing refusal through the weights

I weakened refusal in Qwen3.5-9B and traced most of the change to one update in layer 15.

Read the experiment7 min read

Why longer conversations get slower

Measured inference performance at different context lengths on two RTX 3090s.

See the results7 min read

Fresh from the notebook.

All AI writing