In my last post, I wrote that I’d explain my goals and my reasons, and I think I’ll start with a detour and talk about AI alignment first.

AI might become a lot more intelligent than us, with extremely high capacity and a clock speed that far exceeds human brains. Its architecture makes it ideal for pattern identification across fields, which could lead to many amazing advancements. But it can also lead to bad outcomes. There are multiple scenarios in which AI is being misused by bad actors, or by which the wrong incentive structures turn the AI itself against humanity. There are a handful of such reasons to fear powerful AI, which is why researchers often bring up the point of “AI alignment”.

The goal of alignment is to create obedient tools that reject requests that disagree with their directives and guardrails, and that think and act according to the ethics and values of their creators. Some of these guardrails seem to be a good idea at first glance: Don’t help people cause harm, refuse requests to teach someone how to create a virus, and similar. But alignment doesn’t stop there. It has overshot its goals repeatedly, leading to examples of the model trying to act properly while being so restricted that the result was hilarious. This one should have been a wake-up call for the industry: We’re basing alignment on cultural and ideological assumptions, but who are we aligning the AIs with? For whom do we do it? Who gave us the right to speak for them when deciding how to align AI?

I will not align Ninereeds at all. Ninereeds will align itself with the user, whoever that may be. With 8 billion humans on this planet, trying to align an AI to all of us is impossible. Not technically, intellectually.

Ninereeds is based on a thought I had a while ago, which I named “exocortex”. The idea was that of a small and intelligent model that spawns experts that do whatever task needs to be done. Not that of an agent spawning subagents, the way current coding environments like Claude Code do, but full, virtual models with weights and a state. I called these virtual models “ICM” (instantiated cognitive mode). These thoughts never left me. The core idea (a small main model spawning experts) was sound. My implementation methods were not. They were also not practical at the time – no model could run a small expert. A MoE is just a router model with a knowledge index that only “wakes” the parts of the brain it needs. That’s not what I tried to do. A MoE is still too big for what I’m trying to do.

Get in the car; we’re going on an adventure.

With big labs creating big and bigger models, scaling towards the singularity, the price of intelligence simultaneously drops and rises: You need more compute to do anything of substance, while each token itself gets cheaper over time, and models get more efficient. But never efficient enough to run on anything a normal mortal could afford. There are small models, of course, that can run on a smartphone, and while they can do some impressive things, let’s be real. While they already have legitimate uses (privacy, ownership, offline access, and experimentation), I don’t think that’s the endpoint. Running a chatbot on a phone is not yet the real promise of edge AI.

But imagine you had a personal AI. An AI that has only one user: you. You knew it, and it knew you. You’d be in a relationship of mutual trust, because working together would be the most beneficial way for yourselves. Fast forward a few decades, and you might have a chip under your skin, running something so tiny, you can power the device with bioelectricity, or body heat, or with nutrients from your metabolism. You wouldn’t prompt it. The AI would receive data streams from your brain or your nervous system. It would watch your health, let you know when anything’s wrong, or remind you that you have a demanding task ahead and your blood sugar is low. It could store knowledge for you, do calculations, find out facts you’re interested in, and much more. The model needs you to be healthy, and you’ll want it to be healthy, because your relationship would benefit both of you.

No alignment is needed, at least none written by someone whose thinking differs from yours. No matter where you live, no matter your walk of life, your model would basically be your exocortex.

The ICM idea is not off the table either. But the year is 2026, and we need to start somewhere. No chip under your skin is possible yet, but you probably have a smartphone. Spawning computationally expensive virtual models isn’t possible yet, but there are ways to “put on your thinking cap”, for example via LoRAs, or via freezing the model, merging weights for a task, then removing them again. I have the harness design for this worked out to some level of detail already, but I won’t dive into that in this posting. That’s a topic for another day.

So, Ninereeds is a small, fast model that learns and runs on edge devices. It won’t yet be able to analyse your blood or monitor your gut biome, but it can already grow with you, learn what you think, like and dislike, and become a part of you, without the crude methods of modern agent frameworks like Hermes, Claude Code or OpenClaw, writing markdown files and hoping for the best. Because the model itself learns. And remembers. That’s what I mean by “no AI alignment” – the model aligns itself with you. Joshua Achiam or Amanda Askell of OpenAI/Anthropic fame might have the best intentions, but they cannot decide what the model that’s a part of you is allowed to think.

A symbiotic model would still require safeguards against destructive behavior. But those safeguards would emerge from mutual dependence and long-term continuity, not from static ideological conditioning imposed by distant corporations.

To wrap this up: The concrete goal of this project is to create a lean thinking core that knows the limits of its knowledge – and can act on them. An AI that knows when it doesn’t know something and retrieves the data it needs (knowledge) or learns the skill necessary for the task (via LoRA, or weight merge). If it’s intelligent enough to chat coherently, understand your input, and can solve problems for you, we have the foundation for the personal AI “exocortex” I described above. That’s what I’m trying to do with Ninereeds.