mkmccarty3

I built a solution that automatically optimizes your website over time by reviewing your traffic data and A/B testing different landing pages.

I'd love any feedback. Let me know what you like or don't like about the solution. I'm just getting started, so any thoughts would be helpful.

Most AI models stop learning once training ends. They can generate text, but they can’t evolve. I started wondering what would happen if they could.

Humans work differently. We replay experiences in our sleep, strengthening what matters and pruning what doesn’t. That cycle of reflection is what makes our learning continuous.

I wanted to see if a machine could do the same.

So I built Dream ML, a framework that gives AI models a kind of sleep cycle. During downtime, the model replays its own experiences with added entropy. It connects patterns that were never seen together, fine-tunes itself locally through LoRA adapters, and wakes up slightly changed.

Each dream cycle helps the model reorganize and deepen its internal representations.

This started as part of my project to build a local AI clone that learns from my data and runs fully on-device. It observes how I write, reason, and communicate, then improves through Dream ML cycles without ever sending data to the cloud.

Write-up and details in the linked essay.

Over the weekend I built a model that could run offline and navigate me around the city of Chicago.

It started with an idea to see if I could get a model to learn turn by turn directions between any points in the city by only training on a subset of the street data.

It turns out it's more practical than expected. With only a 360M parameter model, I was able to train it to navigate around the city with 500,000 examples.

Has anyone tried anything similar? What other models should I try building?

Rostra published a short study where they forced models to collaborate in order to see if they perform better on factual questions. Interestingly, the results showed that they were able to outperform all of the best models, even with last generation models or "mini" versions of larger models.

Is anyone else building something like this where they are working on model collaboration? I'm interested in learning more.