Posts in project
From seed to weights: fine-tuning a shell operator
- 15 June 2026
two cycles complete. At archetype-level holdout (n=16, task types absent from training), fine-tuning lifts Mistral termination from 0/16 (base) to 9/16 (tuned), same harness, only the adapter differing. The operate / terminate mechanism generalises to unseen archetypes; task competence (verified 0.31) stays archetype-local. One model, one seed; signal clean.
Semantic Execution Kernel
- 14 June 2026
A virtual, POSIX-adjacent micro-kernel whose most unusual user is a language model. It logs in on a chat-completion pty and drives a real shell: no tool-calling, no function schemas, no agent framework. The model operates the system in the one vocabulary it already has, text.
Scrollback priming: can synthetic history run a shell?
- 14 June 2026
replicated (N=5). Within llama3.1:8b, structure is the lever (0->2->5 clean). Cross-model (6 subjects, 3.8B-8B, non-tool + tool-trained): two axes dissociate. Operation transfers broadly, clean exit is llama-only (2/154 non-llama). Neither scale nor tool-training explains it; leading read is seed-overfit to llama.
Regressions…
- 14 June 2026
Well, after a nice business line meeting and a lot of claims I made there about the SEK project I’m working on, it was time to reproduce my claims… Well, that didn’t work as expected. It’s been a couple of weeks and I only remembered problems I’ve been having back then. One of the runit daemons I’ve developed for model instance management (llmsv) was behaving slow, so I thought before working on anything else, let me fix that, so that reproducing my findings is smooth sailing.
New Impulses: The proof is in the pudding
- 11 June 2026
Was on a business line meet-up with my company. Had a couple of interesting in-depth conversations and presented my work on SEK to them. The meeting was under the banner of AI anyway, so it was natural to me to follow the directive of “let’s talk about AI”.
Learning Content as Code
- 12 December 2025
Building a Libre Open-Source e-Learning Authoring Toolchain
Most people have taken an online course at some point, maybe a workplace training, a coding bootcamp, or a university module. Behind the scenes, those courses are usually built with specialized software platforms. They often come with drag‑and‑drop editors, proprietary formats, and limited ways to move your content elsewhere. If you switch platforms, you often have to rebuild everything from scratch.