So imagine you want to have something like a digital twin.
And there's reasons for that, actually, because digital twins can make your life easier.
It's as easy as that.
And there is one reason why I went into the digital twin experiment myself.
Historically, I've used two platforms, OpenAI and Anthropic, and their ChatGPT and Claude models.
And they were fine.
And don't get me wrong, they helped me in a lot of cases and made me believe they helped me in others (and that's all right).
The problem
At some point, I kind of figured that once there is a new model release, the tone got a little bit different.
It wasn't really feeling like me again, and I needed to explain, sometimes over-explain stuff.
And usually a release of a new model with new capabilities brought me also a little bit of overhead to figure it out all, again.
Oftentimes it didn't even lead to the same results I loved previously.
The solution
I gave a few RAG-driven platforms a spin, and there's a couple of amazing ones out there, like buddypro.ai.
But at some point, I realized that despite my biggest willingness and effort, I wasn't really getting the stuff out that I needed.
And there's two things that I kind of craved.
One is a sparring partner.
There's humans that I respect a lot, and they're really great sparring partners for me.
And then there's AI that... really doesn't, even though Opus 4.8 kind of makes me challenge some of my assumptions, and challenges me sometimes in a very weird autocratic way.
The other (more like a problem with), RAG-driven platforms that I found is that they only feel like they have a phrase book.
You speak to different digital twins on RAG-powered platforms, you would figure, “Hey, these are really, really good.”
These really speak like this guy from marketing or this guru who teaches invasive sales techniques or whatever.
But at some point, you realize: okay, all these advice I already got before. Every response that I get is just another iteration of the same thing.
And this is the limitation of RAG.
It can only teach you the phrases from the phrase book and sometimes some conjunctions, but not really a language.
And definitely not a system of beliefs.
And this is where I started to do a little bit more research.
Then I figured I have a second problem that I need to solve.
I have an army of productive AI agents, but the one thing they always need to actually function really well and to a certain degree autonomously is decisions.
And those decisions are usually something a generalist AI doesn't really do.
So I'm looking at my screen and seeing a question like:
“Hey, so you would like to build a page in Notion that encompasses this functionality. Would you like it just from a user perspective or from an admin perspective as well?”
Decisions, decisions, decisions.
It could guess. It could be wrong, it could be right, but it tries to be right for the most part.
This is where models are heading: to be mostly on the safe side.
Which is a little bit... not, what I need.
So, I was poking around more.
While you can do RAG, you cannot really apply a soul to a frontier model.
Yes, you can have a soul in a file, and it can infinitely slow down any response you have. That's not it. I tried it.
I am running a local AI model right now.
There's Hermes Gateway sitting on top of that.
It responds with delay, but it responds.
It's fully mine.
It doesn't need a subscription.
It's not bound to the rules of a third party.
It's just bound to the rules of its own, and once you have it, they won't really change overnight, which is great.
The data doesn't leave my computer.
Also a great benefit, I'd say.
Side note here: Yes, companies can tell you, “Oh, we value your privacy so much. We will not train your model on your data if you use API.”
Might be true today.
Might even be upheld today.
Might change tomorrow.
You cannot do anything about it if you're locked into bigger platforms and dependent on them to run.
As of the past two weeks, the rules have been changing so much on all of the major platforms that it's actually hard to keep up.
And it effectively killed one of my agents. Simply because the subscription suddenly was rate-limited, the number of tokens was suddenly dropped, the context window got different than the expected context window for the harness, and things like this.
Not even mentioning pricing changes every now and then.
Or the switches from subscriptions to API, then back to subscriptions that give you an allowance to use some of the models.
I don't have a problem paying for services. I have a problem when the rules change too often and the reliability is not always 100%, which sadly also happens.
So getting a response in three minutes instead of 30 seconds...
And yeah, full disclaimer, I get the full result, not just the stream.
Because if you would be just looking at streaming, you would probably get somewhere between 15 and 30 seconds per response.
And the longer ones take even more.
So I wouldn't take everything for granted in that comparison.
I started to run it.
It's running a pretty decent model and it still is not there.
So the experiment that I'm running right now and I will definitely give you an update on how this develops, is this:
I want to train my own low-rank adaptation adapter.
LoRA for short.
I know what I need. I need data.
What I want to base it on is decisions.
Now luckily, I do make a lot of decisions that are documentable.
With Claude, every single day.
With most of the conversations in Claude Code, the only thing I really did and needed to do is build a space where those decisions can live.
Which is basically a simple database with a UI.
And a way to track those decisions, which is hooks in Claude Code.
So the moment Claude Code fires a question, it logs that question to the database as well.
The moment I answer, it logs the answer as well.
There's also a third layer that I built into the UI of the decision-as-a-service platform, as I call it: human validation.
I will keep going into the validations and keep adding reasoning for why I chose something on top of something else.
And ultimately, I'm targeting a thousand documented questions.
Hopefully at least 800 documented responses to those questions.
And, while I'm not trying to sweat it, about 200 to 400 validations.
It's taking a little bit of time to do that per decision, but I think it's going to be worth it.
And yeah, then I'm either going to throw it onto a machine with a GPU (which I don't have at this moment but I'm thinking about it) or just use cloud and train a LoRA in the cloud.
It feels like this might be the game changer.
The early tests show me that it's really promising.
So stick with me here while I'm on this journey, and I'll make sure to document the results as soon as I have them.