Skip to main content
MICHIEL HORSTMAN
← Back home

SELECTED WORK / INDEPENDENT PROJECTS

Projects that make an idea testable.

I use small projects to work through questions that are easier to answer in code than in a deck. One gives agents evidence from video. One studies refusal behavior in language models. One proves an identity claim without handing over the document. The last is a research result: changing what a small model can do by editing its weights.

Independent research · model editing04 / 04

Qwen3.5-0.8B-NKC

New abilities by editing existing weights, not by training.

About 3.28M of 873.4M weights changed (0.38%). No training run, no added parameters, no donor model at inference. The math gain repairs thinking-mode looping and default mode costs 2 to 4 points; the write-up lists the limits.

PUSH-T PHYSICS

50% → 89%

DEPTH JUDGMENTS

75% → 83%

MATH-500, THINKING

6% → 53%

Have a hard systems problem?

I am interested in work where software and applied AI have to survive outside a demo.

Surface. Thanks for diving.

Shark model by alextkd2003, CC BY, converted to ASCII.