Article

🦉 From Broken Models to Living Systems

I didn't start with a perfect AI setup.

No GPU. No funding. Just ideas, experiments, and a stubborn urge to build.

These are old projects now, but each taught me something.

🚀 Lynqbit — My Favorite Failure

Lynqbit was my first serious AI project.

A roughly 90M-parameter model built with ambitious goals and plenty of experimentation.

Eventually, my system couldn't properly support the training process.

Two months of work were gone.

But the failure taught me something:

Training shouldn't depend on one fragile setup.

🦉 Barn Owl AI

Next came Barn Owl AI.

I experimented with streamed training, cloud-hosted data, and incremental learning.

It didn't last long. Infrastructure problems and bugs eventually stopped the project.

The project failed.

The idea didn't.

🦉 Elf Owl AI

Then came Elf Owl AI.

This became my first AI experiment that actually reached a usable state.

• 25M parameters
• Custom training data
• Open-source
• Publicly released

It wasn't perfect.

It hallucinated, struggled with grammar, and had plenty of quirks.

But it worked.

That felt like a huge victory.

🦉 Xenoglaux AI

Later came Xenoglaux AI, another experiment focused on streamed and modular training.

The dataset was larger, but the same enemy remained:

Hardware.

Training that took hours on a GPU could become painfully slow on CPU.

🧠 What I Learned

• Failure is compressed knowledge.
• Small models can still be interesting.
• Hardware limitations force creativity.
• Every failed experiment can improve the next one.

These projects are finished chapters now.

But their lessons stayed with me.

Sometimes the model fails. The idea doesn't. 🦉

#AI #MachineLearning #Python #OpenSource #Programming #BuildInPublic