All Things AI · Season 2 · The Models · Episode 7
LoRA is not a chatbot. It is a small file trained on your own writing that leans another model toward your cadence. You never open it. It sits underneath whatever you are already using and changes what comes out of it.
Why this is in a season about models
Every other episode in Season 2 is something you open and use. This one is not, and it belongs here anyway because it changes the output of all the others.
Think of it as infrastructure rather than a tool. A model is the engine. An adapter is a modification to that engine that makes it produce your work rather than generic work, and it applies to everything the engine does afterwards.
What the technique actually is
Low-rank adaptation. The mechanism is that instead of retraining an entire model, which is enormously expensive, you train a small additional set of values that sit alongside it and shift its behaviour.
The result is a file, usually a few dozen megabytes rather than the many gigabytes of the model itself. That size difference is the whole reason this is available to individuals rather than to companies with data centres.
It is trainable on your own hardware
A few hours on a capable laptop, unattended. This is the fact that most people find hardest to believe and it is the one that matters, because it means training a model on your own work is a Saturday rather than a project.
It attaches to a base model
The adapter is not useful alone. It modifies something. That means you can move it to a better base model later, which is a real advantage as the underlying models keep improving.
It carries style, not knowledge
This distinction is worth holding firmly. It learns how you write. It does not learn what you know, and it should never be relied on for facts. Voice from the adapter, facts from your files.
What it changes about the work
The output stops being something you rewrite and becomes something you correct.
That sounds like a small distinction and it is the entire economics of the thing. Rewriting a generic draft takes about as long as writing from nothing, which is why so many agents try AI writing, find it faster in theory, and quietly stop.
Correcting one beat in a draft that already sounds like you takes a minute. The difference between those two experiences is what determines whether any of this actually saves time, and it is almost entirely down to whether the model was trained on your work or told about it.
What you feed it
The quality of the adapter is determined almost entirely by what goes in, and there are only a few rules.
- Thirty to fifty pieces minimum, of your own writing, actually written by you.
- One kind of writing per adapter. Scripts or long posts or client correspondence, not a mixture, because a mixture averages into a register you do not have.
- Your better work rather than everything you have. An adapter trained on your average output produces your average output.
- Nothing a tool generated, ever, because that teaches it to sound like the tool.
What to try in the first week
Two exercises, and the first one is a decision rather than a task.
Go and find your thirty pieces. This is the actual work and it is the step where most attempts stop, because it is unglamorous and it takes an evening. Open the folders, read what you have, and choose the ones that sound most like you at your best. That selection is the single largest determinant of the outcome.
Then, before training anything, run a test: take one of those pieces, give a model the same brief that produced it, and read what comes back next to what you actually wrote. That comparison shows you exactly what an adapter is going to have to close, and it makes the result far easier to judge honestly afterwards.
Where this goes wrong
Four failures.
Training on tool-generated text. The most common and the most quietly damaging, because the material looks fine and the result is a machine reproducing the exact voice you were trying to escape.
Mixing registers. Produces an adapter that writes everything adequately and nothing well.
Expecting it to know things. It will produce confident specifics in your voice, which is more dangerous than confident specifics in a generic one, because you are less likely to check something that sounds like you wrote it.
Rewriting the output wholesale. If you do that you have trained a model in order to keep writing everything yourself, and you have started editing your own voice back out.
How it sits next to the others
It is not an alternative to anything in this season. It sits underneath them.
The natural pairing is with a strong long-form open model as the base, running locally, because that combination gives you your cadence, at length, free, and private. That is a genuinely capable writing setup and none of it is rented.
It also pairs with memory, and the two do different jobs that people conflate. The adapter carries how you write. Memory carries your rules and your past decisions. Neither substitutes for the other and both are needed for output you can actually send.
Choosing which register to train first
You will end up with more than one adapter eventually, so the question is which to build first rather than which to build.
The right answer is whichever kind of writing you produce most often and dislike producing most. That combination is where an adapter pays back fastest, and it is usually not the writing that feels most important.
For most agents that is the repeated explanatory writing: the same things said again to different people, which is high volume, genuinely useful, and the thing that gets skipped when the week is bad. Training that register first means the adapter is earning within days rather than waiting for the occasion it was built for.
Build the second one only after the first has been in use for a month. By then you will know what it does well and what it misses, and the second is considerably better for having learned that.
How you know it worked
Judging your own voice is harder than it sounds, so it is worth having a test rather than an impression.
The useful one is blind comparison. Take three pieces you wrote and three the adapter produced from similar briefs, strip anything identifying, leave them a few days, and read all six without knowing which is which. If you cannot reliably tell, it worked. If you can tell instantly, it needs more material.
The second test is the correction count. Track how many beats you change per piece over a couple of weeks. One or two is the target and means the thing is working. Five or more every time means the adapter is producing a draft rather than your draft, and the answer is more training material rather than more editing.
Both tests take almost no effort and they replace the vague sense of whether it seems better, which is not something anyone can judge about their own writing.
Why this is the most underused thing here
Almost nobody in this industry has done it, and the reason is not difficulty.
It is that the phrase sounds technical, the acronym is off-putting, and the entire category is described in language written for engineers. The actual process is gather your writing, run a command, wait a few hours, and use the result.
The gap between how hard this sounds and how hard it is may be the largest of anything on this site, and it happens to sit on the capability with the most direct effect on how your work reads.
What it costs
Nothing. Training runs on hardware you own, the adapter is a file on your disk, and using it afterwards is free. The cost is an evening of gathering your writing and a few unattended hours, paid once.
Questions agents actually type
What is LoRA in AI?
Low-rank adaptation: a technique that trains a small file to sit alongside a model and shift its behaviour, rather than retraining the whole model. The file is small enough to train on your own laptop, which is why this is available to individuals at all.
Can I train AI on my own writing?
Yes, with thirty to fifty pieces of your own work, one consistent kind of writing, on hardware you already own, in a few unattended hours. The result is a file you keep and can move to a better model later.
Does an adapter make AI know about my business?
No. It carries how you write, not what you know, and it should never be relied on for facts. It will produce confident specifics in your voice, which is more dangerous than in a generic one because you are less likely to check them.
Do I use LoRA instead of ChatGPT?
No, it sits underneath a model rather than replacing one. The common setup is an adapter carrying your cadence on top of a strong open model running locally, which gives you your voice, at length, free and private.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.