Fine-tuning and steering on a desk
Training and steering small open AI models on my desk instead of renting a data center.
- Status
- Ongoing experiments
- Year
- 2026
- Role
- Solo
- Area
- Machine learning
In plain English
Big AI models need big hardware, yet small open models can learn narrow skills on a single desktop machine. These are my running experiments in doing exactly that on Apple silicon.
How it works
- Teach Phi-4-mini new skills with LoRA adapters (small add-on layers that are cheap to train).
- Nudge Qwen3-4B's behavior by editing its internal activations (the model's working signals).
- Fine-tune Qwen to write SEO content (text written to rank well in search engines).
- Send heavier runs to RunPod (rented cloud computers) with a spending limit per experiment.
What I did
These are my solo experiments, run with MLX, llama.cpp and Ollama on my own hardware.
What I learned
Small models are far more capable than their reputation once the task is narrow and the data is clean. When one machine runs several jobs, memory-aware scheduling matters more than raw speed.