Where I think AI is heading

The industry is pouring its energy into making models bigger. I think that's optimizing the wrong variable. The bottleneck in front of us isn't raw intelligence — it's the distance between what a human wants and what actually gets done.

I've spent twenty years shipping systems where the hard part was never the cleverest algorithm — it was the integration, the context, the coordination between people and machines that had to hold up in the real world. That lens shapes how I read the current moment in AI. We have models that can reason, write, and code at a remarkable level. And yet most of that capability never converts into outcomes, because the model has no memory of you, no awareness of the room, and no stake in the goal.

Close that gap and the value of every model you already have goes up. That's where I'm placing my bet.

The bottleneck is context, memory, and coordination

A model invoked through a chat box starts from zero every time. It doesn't remember the decision you made last Tuesday, the constraint your customer gave you in March, or the commitment your colleague made in this morning's meeting. Each interaction is brilliant and amnesiac. We've been compensating with longer prompts and bigger context windows, but that's a workaround, not an architecture.

The teams and tools that win the next decade won't be the ones with marginally larger models. They'll be the ones that build a persistent layer between human intent and intelligent execution — something that remembers, prioritizes, tracks commitments, and carries context across people and time.

The next leap isn't a smarter model. It's a memory that outlives the conversation.

One cognitive architecture, three scales

The bet I'm prototyping — I call it MIRA / C.OS — is that the same cognitive architecture should operate at three scales, because intent, memory, and coordination are the same problem whether you're one person or a hundred:

• MIRA — the individual. A personal cognitive engine that remembers, prioritizes, learns, and acts on your behalf across life and work — with persistent private memory, a morning brief, and the ability to participate rather than just respond.
• Meeting Intelligence — the room. A live companion that listens, structures the discussion, and turns meetings into searchable operational memory — extracting the decisions and follow-ups instead of letting them evaporate.
• C.OS — the group. A coordination layer where many MIRAs cooperate on shared goals, with governance, role contracts, accountability, and coalition-level execution.

One underlying engine spans all three — personal assistance, meeting intelligence, group coordination, and eventually embodied agents and robots. The same architecture, scaling from one person to many.

Why one architecture matters

If each layer is a separate product, you get integrations. If they share one cognitive engine, you get compounding. The memory MIRA builds for you makes the meeting smarter; the meeting's decisions feed the coalition; the coalition's outcomes teach every MIRA in it. That's a network effect built on memory and context, not just data — and it's the same reason a great engineering organization outperforms a collection of great engineers.

It's also the most credible path I see to physical AI. A robot is just an agent with a body. If you've already solved persistent memory, intent, and coordination in software, embodiment becomes the next endpoint of the architecture — not a separate moonshot.

Why I'm building it, not just writing about it

I don't trust a thesis I can't ship. Every strong opinion I hold about systems came from putting one into production and watching where it broke. So I'm building MIRA / C.OS as a working prototype to pressure-test the idea — to find out where the architecture holds and where reality pushes back. That's the same instinct that's served me from firmware to vehicle platforms: have a point of view, then earn it by making it real.

If you're building in this space — or betting on a team that is — the question I'd be asking isn't "whose model is biggest." It's "who is building the layer that turns intelligence into outcomes." That's where the next decade gets decided.