AI chat apps are built around single-model conversations. You ask, the model answers-for most people, that answer becomes truth. That's probably fine for questions with clear-cut answers (which the models are increasingly able to provide, hallucination free), but it breaks down when you're exploring something genuinely uncertain or complex.
All of the big models are now sufficiently perceptive that they can pick up on what you seem to believe. Once they do that, they tend to reinforce your existing beliefs rather than challenge them. That's how these systems are trained, increasingly through the inherently opinionated process of RLHF. It's also a consequence of companies, understandably, wanting to drive engagement. No one wants to talk with someone who disagrees with them.
So I started using multiple models. I'd ask Claude something, copy the response to ChatGPT and ask what it thought, then bring Claude's feedback back to ChatGPT. When I liked parts of both responses, I'd manually blend them together. It was clunky, but it worked. Different models would catch things the other missed, push back on each other's reasoning, offer genuinely different perspectives on the same problem.
What I wanted was simple: an @Claude or @ChatGPT I could invoke mid-conversation the way you'd pull a colleague into a Slack thread when you want another take. xyz is that experiment. It's a space where multiple AI models can work on your questions together, with conversations that persist and build rather than restart from scratch each time.
xyz treats AI collaboration the way we treat team collaboration: as something worth organizing, persisting, and returning to. Not bots that give you one-off answers, but collaborators you build with over time.
I wrote more about this in "Dream team." The thesis: the most effective way to use AI isn't picking the "best" model. It's assigning different perspectives to different models and using them together, as inputs and outputs to each other.
xyz is also a playground for ideas I'm exploring as multi-agent systems mature. Areas I plan to use xyz to explore:
I'm Cale Reid. You can find me on LinkedIn, read more on Substack, or reach out at calereid@gmail.com.