Why Real-Time Control Changes Everything We Assumed About AI

What if the most interesting frontier in AI isn't the next giant language model, but the unglamorous, practical edge — the bit where the theory meets a motor, a pump, a drone, a power grid?

That's the thread running through this week's conversation with Matthew Carr, CEO and co-founder of Luffy AI. His argument is blunt and persuasive: while generative AI has been sprinting ahead in the cloud, the real-time control layer — the part of the stack that actually decides what physical systems do, in milliseconds, out in the real world — has barely moved. And that's where a lot of the real industrial value is quietly hiding.

Matthew's route into AI is unusual. His background is a PhD in plasma physics and nuclear fusion technology, a research fellowship, and time spent in renewable energy — the kind of world where control loops aren't a nice-to-have, they're the whole ballgame. The nagging realisation that kept pulling him back was this: generative models were getting cleverer by the month, but the industrial control systems running the equipment underneath them were using techniques that predated the deep-learning era entirely.

Luffy AI is his answer to that gap. The pitch: adaptive, neuroplastic neural networks that learn and re-tune on the fly, orders of magnitude more compute-efficient than conventional approaches, and small enough to sit directly on the equipment they're controlling. In plain English — AI that keeps learning on the job, runs on a chip at the edge, and makes decisions fast enough to actually change what's happening in the physical world.

The implication for project and asset delivery is significant. Better real-time control means less waste, fewer failures, and more energy-efficient operations. It also produces a genuinely different answer to the "AI and sustainability" question than the one you usually hear: instead of adding more data-centre load, you quietly remove energy and material waste from the systems AI is controlling.

One of the most useful framings from the episode is the distinction between AI that analyses and AI that controls. If your model can't respond inside the time window where the outcome is still changeable, it's analytics — useful, but not control. Matthew walks through why that window is shockingly short in industrial settings, and why the usual "send it to the cloud and wait" architecture breaks down the moment physics gets involved.

Neuroplasticity is the other big idea. A neuroplastic controller doesn't just ship with a fixed model; it keeps adapting as the equipment it's running on changes — wear, drift, temperature, new operating conditions. That matters because the traditional failure mode of deployed AI isn't dramatic error — it's slow decay as the world stops matching the training data.

Subscribe to the Project Flux newsletter to get the weekly catch-up in your inbox every Monday morning — and if you enjoyed this episode, the kindest thing you can do is share it with one person who'd find it useful.

AI keeps coming up at work, but you still don't get it?

That's exactly why 1M+ professionals working at Google, Meta, and OpenAI read Superhuman AI daily.

Here's what you get:

Join 1M+ pros

All content reflects our personal views and is not intended as professional advice or to represent any organisation.