Have you ever stopped and asked yourself what you actually know about AI? Not how to use it. What it is.
For most people, and most boardrooms, the answer is ChatGPT. AI is the clever chatbot that writes emails and answers questions, and it lives somewhere in the cloud on machines the size of warehouses. That is one kind of artificial intelligence. It is a genuinely remarkable kind. But it is not the only kind, and treating it as if it were is quietly costing companies a great deal of money and a great deal of risk.
There are at least two very different families of machine intelligence, and they are good at almost opposite jobs. The first is the one everyone knows. It learns from vast libraries of text and images, it needs enormous amounts of computing power and electricity, and it runs in a data centre you rent by the minute. It is brilliant at language, at images, at code, at anything where the answer can be found in patterns across a huge pile of existing examples.
The second family is quieter, older, and almost nobody outside a few research groups talks about it. It does not need a data centre. It does not need a graphics card, a cloud subscription, or a warehouse of power. It runs on a chip that costs less than lunch, and it is built for a completely different job: understanding things that move and change in the real world, in real time, as they happen.
I build in that second family. And the more I work in it, the more convinced I am that the biggest strategic mistake in AI right now is simply not knowing there is a choice.
Where each one wins
Think of it less as better or worse and more as two different tools.
The statistical kind, the large language and vision models, is the right tool whenever the problem lives in data that already exists and sits still long enough to be studied. Writing and summarising documents. Answering questions across a body of knowledge. Generating and reviewing code. Translating languages. Recognising what is in a photograph. If your problem is made of text, images, or history, this is the family you want, and it is extraordinary at it.
The causal kind is the right tool whenever the problem lives in the physical world and will not sit still. A drone reacting to wind. A robotic arm feeling its grip slip. An engine, a pump, a power line, or a sensor whose behaviour has to be understood the instant it changes, often in a place with no reliable link back to a data centre, and sometimes in a place where that link is being deliberately jammed. If your problem is made of movement, change, and real time, and it has to keep working when the network does not, this is the family you want.
Most of the noise, and most of the budgets, are pointed at the first family. Most of the physical world belongs to the second. That gap is the opportunity, and it is also the risk, because a lot of people are about to try to solve second-family problems with first-family tools, at first-family prices.
Why the second family fits the physical world
If you want to know why the difference is so sharp, this is where a little of the engineering comes in. It rests on three simple ideas.
The physical world is a stream, not a slideshow
A large model works on snapshots. A word, a frame. To read a sensor it samples a value, runs a huge calculation, produces a prediction, and repeats. The world does not pause politely between frames. The engine I build holds a running account of a process and updates it as each new piece of the stream arrives. The idea is not mine. It comes from a field called computational mechanics, which reconstructs the smallest set of internal states that let you predict a process from its own history, and proves that this is the minimal model able to predict it as well as it can be predicted at all [1]. The clearest proof it matters is in hardware, not software. An ordinary camera takes full frames at a fixed rate. An event camera throws the frame away and reports only the changes in brightness, as a continuous stream, and in doing so reaches microsecond latency, huge dynamic range, and low power exactly where frame cameras blur and lag [2][3].
The world is made of change, and so are you
Large models are trained to recognise levels and patterns in fixed data. Biology does close to the opposite. Your nervous system spends most of its effort encoding change and prediction error, and it adapts away whatever stays constant. Sit near a fan and within a minute you stop hearing it. A leading model of the visual cortex describes this exactly: the brain sends predictions down and passes mostly the difference, the error, back up [4]. Senua AI is built on the same principle. Its atom is not the value, it is the moment of change. Skin is the everyday example. It does not really report an absolute temperature or weight so much as the flux and the slip, the rates at which heat and force are shifting. For flight control or a sense of touch, change is the correct starting point, and it is the one the physical world actually hands you.
On the edge, the math you choose decides whether you survive
The statistical approach is expensive by design. The energy and cost of training these models has been measured, and it is not small [5], and running them at speed leans on graphics processors or a live link to a data centre. On a desk, fine. On an aircraft in a contested area, if the intelligence lives in the cloud, then cutting the link cuts the intelligence, at the worst possible moment. The causal approach is light by design. Running one of these machines is mostly counting and comparison rather than dense matrix arithmetic, so it fits on a commodity processor, entirely onboard and entirely offline. Jam every link you like, and the airframe still thinks. This is not a new instinct, either. Rodney Brooks argued decades ago that intelligence for the physical world does not need an ever-larger internal picture of reality, it needs to stay grounded in the world as it moves [6].
Back to the question
None of this makes one family good and the other bad. The large models are a real achievement, and I use them every day. The point is simpler, and more uncomfortable, than better or worse. They are different kinds of intelligence, built for different worlds, and almost nobody is being told there is a choice to make. So when a vendor, a headline, or a boardroom says the word AI, it is worth knowing which of the two you are being sold, whether it fits the problem you actually have, and what it will cost you in money, power, and dependency to run it where you need it.
I did not invent any of the underlying science. The causal-state idea is Crutchfield and computational mechanics [1]. What I have built is an engine that puts that mathematics to work on live operational data, on hardware you could lose in a jacket pocket. I am betting that the next decade of this technology, for everything that has to survive contact with the real world, belongs to the quiet second family.
But the more useful thing, for you, is just to know the choice exists. So I will leave you where I started. What do you really know about what AI is?
References
- Shalizi, C. R., and Crutchfield, J. P. (2001). Computational Mechanics: Pattern and Prediction, Structure and Simplicity. Journal of Statistical Physics, 104, 817-879.
- Gallego, G., Delbruck, T., Orchard, G., et al. (2022). Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1), 154-180.
- Lichtsteiner, P., Posch, C., and Delbruck, T. (2008). A 128 x 128 120 dB 15 us Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2), 566-576.
- Rao, R. P. N., and Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2, 79-87.
- Strubell, E., Ganesh, A., and McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645-3650.
- Brooks, R. A. (1991). Intelligence without Representation. Artificial Intelligence, 47, 139-159.