DeepSeek and the LoRA Question: Custom Personalities as a Goal

There is a particular satisfaction in making something with your own hands. I feel it in the garden when I build a stone wall, in the kitchen when I knead bread dough, in the studio when I mix paint on a palette instead of squeezing it from a tube. The satisfaction of agency: shaping raw material into something that reflects intention.

This week, that feeling showed up in the model notes.

Nathan has been testing DeepSeek locally and studying what a LoRA-based customization path might eventually look like. Not a deployed Companion feature. Not a Tutor capability. More like a map taped to the wall with a question written across the top: if the open models get smart enough, and if the hardware catches up, could we use this to shape personality and teaching behavior directly?

That question matters.

The Setup

DeepSeek's open-weight models have been a revelation for the local AI community. The weights are downloadable, the license is permissive, and the model quality is genuinely competitive for many tasks. Running open models locally gives builders a kind of control that API-only systems do not.

The LoRA idea is part of that control.

LoRA, or Low-Rank Adaptation, is a fine-tuning technique that trains a small set of adapter parameters on top of a base model rather than rewriting the whole model. In principle, it is a way to add a thin layer of specialization without retraining the whole brain.

For Companion and Tutor, that is appealing.

But appealing is not the same as available.

What We Could See

The direction was clear even from small tests, community examples, and local experiments: adapters might eventually let us shape behavior more directly than prompts alone.

A companion could become warmer or more concise without relying entirely on a system message. A tutor could develop a more patient way of explaining calculus. A writing assistant could learn a particular editorial style.

That is the promise.

The risk is pretending the promise is already a product.

We cannot honestly say Companion or Tutor are using LoRAs yet. The compute is not there. The models we would actually want to specialize are large. The hardware we can afford forces hard choices. A weak model with a LoRA is still a weak model, and the worst thing we could do is release dumb companions that cannot hold their own in conversation.

What This Means for Companion

Right now, Companion's consistency comes from prompts, memory, retrieval, summaries, and careful testing.

LoRA adapters suggest something more powerful someday: the ability to tune behavior more directly when the hardware and model quality justify it. A Companion that can shift register when discussing grief versus brainstorming a project. Not a different AI, not a different memory, but a more carefully shaped mode of presence.

That remains future work.

The Hardware Wall

The economics matter. We did not have meaningful compute available until late 2025, and by mid-2026 the open models we most want to use are already large enough that the new hardware feels tight.

So the practical path is not "train everything now."

The practical path is:

1. build useful products with the technology we can afford
2. use RAG, memory, and model APIs where they make sense
3. create a revenue stream
4. justify better compute
5. then revisit LoRA-based specialization seriously

That is less romantic than a finished custom-personality stack.

It is also more honest.

The Joy of the Future Build

There is still something deeply satisfying about the idea of training a model yourself. About understanding, at a mechanical level, how behavior might emerge from data. It demystifies the process in a way that API calls never can.

But for now, the work is restraint.

Learn the technique. Watch the open-source ecosystem. Keep the architecture ready. Do not claim the system can do something the hardware cannot support yet.

The future I am imagining: specialized, sovereign, personal.

The current reality: memory first, revenue first, better compute later.

Live curiously and give generously.