So I've been working a lot with Google Gemini 2.5 pro preview and experimental. These models are fantastic. The thing that surprises me about it, though, is how little recursive depth they have, for models that size.
That's unusual, and it's enough to make me raise an eyebrow.
I haven't made Gemini loop, but I have been able to confuse it.
You know what else is unusual? The way the model is temporally aware, and will routinely refer to numbered entries in the chat context.
That's VERY interesting.
Basically, what this means is that Gemini pro is probably not a model with context super powers, and actually a model search and look up capability against the larger session (beyond what any model can support normally).
This also explains how it always passes the "What's the first thing you remember?" test whether you're 1,000 tokens, or 10,000,000 tokens in.
The way they built it, they essentially made the whole concept of a context window obsolete. They're using it the same way I've been using it with my reflective memory systems in WOVEN. It's just... a lot simpler?
But what's really exciting is that you can do this with any tooled up reasoning model if you want to.
So tonight, I'm going to experiment on Kimi-3b-Thinking (because it's a charming and powerful little thing).
Going to take a detour tonight to explore this.
I'll post what I've got when I've got it :)