# Mission: Open-Source / Open-Weight AI Models

## Why
To go from *consuming* LLMs through closed APIs to *owning the stack* — picking, running, serving, and fine-tuning open-weight models. Concretely: ship at least one product feature powered by an open model I run myself (or have fine-tuned), and build durable, career-grade understanding of the open-model ecosystem that isn't tied to a single project or vendor.

## Success looks like
- I can take any open model on Hugging Face and predict, before downloading it, whether it will fit on a given GPU — and at what quality.
- I can stand up a production-grade inference endpoint for an open model on a rented cloud GPU and call it from my app exactly like an API.
- I can decide correctly between three paths for a given use case: prompt/RAG, self-hosting, or fine-tuning — and justify the call.
- I can run a QLoRA fine-tune of an open model on my own data and deploy the result.
- I can reason out loud about what hardware a given workload needs, and what it would cost.

## Constraints
- **Hardware:** No dedicated NVIDIA GPU yet. Primary compute path is *rented cloud GPUs* (RunPod / Lambda / hyperscaler). Local machine is an Apple-Silicon Mac — good for light, quantized local inference, not heavy training. Part of the mission is learning *what hardware I'd even need*.
- **Starting point:** Comfortable building apps on Claude/OpenAI APIs; have never self-hosted a model, touched weights, or rented a GPU.
- **Style:** Wants both a shippable outcome *and* deep conceptual understanding — not just recipes. Likes to get hands-on.

## Out of scope (for now)
- Pre-training a model from scratch.
- Distributed multi-node training across GPU clusters.
- The deep mathematics of transformer internals (we use mental models, not derivations — revisit later if the mission demands it).