How to Invent a New Colour

(and Why It Matters for AI)

Your phone can’t show you every colour you can see. That isn’t something a better screen would fix - it’s built into the way screens work. Once you understand why, it turns out to be one of the more useful things to know about artificial intelligence as well. So let’s start with the colours.

Every colour on your screen is made from three lights: red, green and blue. A sunset, a photo of your dog, the exact shade of someone’s jumper - all of it is those three mixed in different amounts. If you took just those three and combined them in every possible way, the colours you could reach would form a shape. Picture a triangle with red, green and blue at the corners. Everything inside the triangle, the screen can show you. Everything outside it, it can’t, and no amount of clever mixing changes that. Blend the three lights however you like, and you always land somewhere between the corners, never past them.

That triangle has a name: the gamut - and it’s simply the full range of colours a device can produce. A cinema projector has a wider one than your phone. Your own eyes have a wider one than any screen ever built, which is why some colours look impossible to photograph properly. The rule underneath all this is worth holding on to, because it comes back later: mixing keeps you inside the triangle. If you want a colour that lives outside it, better mixing won’t get you there. You need a new corner - a fourth kind of light you couldn’t have made by blending the ones you had.

A quick confession, for the purist, before we go on: it isn’t really a triangle. That’s the flattened version. The triangle only captures a colour’s hue and how vivid it is, and quietly drops brightness altogether - which is why grey has nowhere obvious to sit on it. Grey, white and black all share the same spot on the triangle and differ only in brightness, the one thing the flat picture leaves out, so they’d stack up at a single point inside it rather than anywhere along the edges. Draw the whole thing properly, with brightness included, and the gamut turns out to be a three-dimensional solid, not a flat shape at all. And a device with more than three primaries - a decent printer, say - wouldn’t give you a triangle even in the flattened version; you’d get a lumpier shape with more corners. So, well, technically it’s all a good deal more complicated. But the triangle is close enough to be useful and simple enough to actually picture, so we’ll happily stick with it. Simplification or not, the rule holds: mixing keeps you inside, and only a new corner gets you out.

That idea does a surprising amount of work once you take it out of the world of colour.

What this has to do with machines

Think about a large language model, the sort of thing behind a chatbot. It learned by taking in an enormous amount of human writing, and out of all that it built up its own set of primaries: patterns, concepts, habits of putting words together. Everything it produces afterwards is, roughly, a blend of those. Which raises a fair question whenever one of these systems does something impressive. Is it actually creating something, or just mixing paint inside its own triangle?

The answer is a bit of both, and the two halves are worth pulling apart.

Some of what it makes is new in the way a fresh sentence is new - a combination nobody happened to write before, but still built entirely from corners that were already there. That’s not a small thing. Most of what we call human creativity works like this: new arrangements of familiar pieces. A model can land on a proof no mathematician wrote out, or suggest a molecule no chemist tried, and all of that can still sit comfortably inside its existing triangle.

Then there’s the other kind of new: a genuinely new corner, something no blend of what you had could reach.

It’s worth being fair to the machines, because their mixing is far cleverer than mixing paint. Paint only averages - put two colours together and you get something between them, usually muddier. Ideas don’t have to behave that way. They can stack and nest and build on their own results, so the “triangle” for a language model isn’t a tidy three-sided thing but an enormous, many-sided region, and the amount you can reach inside it is genuinely vast. Even so, it has an edge, and in practice you can often sense where it lies: these systems are at their surest on ground they’ve already covered, and they grow shakier as they wander towards the outer reaches of it. A richer way of mixing gives you a bigger space to move around in. It still doesn’t hand you a corner you didn’t start with.

It’s worth remembering that the triangle wasn’t always this big. The machine-learning systems that came before learned from small, narrow piles of data - a few thousand hand-labelled photos, a spreadsheet of house prices - and a person usually had to decide in advance which details mattered, in effect placing the corners by hand. That gave you a small triangle: sharp and useful within its bounds, and rather helpless outside them. What changed with large language models is mostly scale and breadth. Pour in a huge slice of everything people have written, let the system work out for itself where its corners should go, and the triangle swells into the sprawling region we’ve been describing. The more there is to learn from, and the more varied it is, the larger the triangle grows.

None of which means bigger is always better. A smaller triangle isn’t only a limitation to grow out of; quite often it’s exactly what you want. If you’re building something to read medical scans or check contracts, a vast do-anything palette is more of a liability than a gift - you’d rather have a narrow, dependable one that stays where you put it and doesn’t wander off into confident nonsense. Part of the craft is choosing the right size of triangle for the job, and sometimes deliberately keeping it small.

Learning that doesn’t stop

For almost every AI system in use today, those corners are fixed the moment training finishes. The model gets taught, then it gets sealed up, then it goes out to be used -and while it’s being used, it isn’t learning anything new. Teaching and using are two separate stages with a wall between them, and the gamut is decided at that wall.

Now imagine taking the wall away. Picture a system that carries on learning while it’s being used, so every conversation, every image, every new experience quietly reshapes it as it goes. Teaching and using stop being separate stages and become the same moment. The triangle is never finalised; it’s still being drawn right up to the instant the thing acts. That’s a genuinely different sort of mind - less a finished object reciting what it was taught, more something that keeps changing because it never really stops learning.

You’d be forgiven for thinking that’s the whole answer: never stop learning, and you’ll eventually escape the triangle. But it doesn’t quite work, for the same reason as before. Keep mixing the same three lights and you get more and more shades, but never a new colour. A system that only ever learns from its own output is a bit like a painter shut in a room, painting over old canvases with mixtures of the same old paint. The room gets busier. Nothing new comes through the door. Learning that only feeds on itself just fills the triangle in more thickly; it doesn’t make it any wider.

So the real question isn’t how to keep learning. It’s where a new corner could possibly come from.

Where new corners come from

Two places, and both of them are outside the machine.

The first is new senses. Sound isn’t a blend of colours. Touch isn’t a blend of sound. The smell of rain isn’t a mixture of anything you can see. Each sense points in a genuinely new direction - a corner the old triangle had no way of reaching, because it was built without it. This is what people are really getting at with the word “multimodal”. A model that has only ever read text has a text-shaped range. Give it sight and you haven’t just added more examples, you’ve added a whole new dimension. Give it hearing, and another. Let it move a robot arm and feel something resist, and another again. Some of the most interesting things live where these meet: the feeling a word seems to carry, the shape you’d give a sound, how heavy an object looks like it ought to be. Those are colours the first triangle could never have mixed, because it didn’t have the corners.

The second is the world answering back. When a system tries something and reality corrects it - the experiment fails, a person says that’s wrong, the robot arm meets a wall it didn’t predict - that correction carries information the system genuinely didn’t have and couldn’t have blended its way to. It came from outside. That’s how a new corner actually gets made: not by thinking harder in the same room, but by going out, bumping into something real, and letting the surprise change you.

So how do you get past the gamut?

Not by making the model bigger, and not simply by having it learn forever. You get past the gamut when the system stops being a sealed mixer and becomes an open one - when it can take in new senses, act in the world, get corrected by that world, and fold the surprise back into itself as it goes, with no wall between learning and doing. Getting past the triangle was never about a cleverer blend of what you already had. It was about making fresh contact with something real enough to give you new material to work with.

There’s a longer arc hiding in all this. A bigger, broader diet already carried us from the small, hand-built triangles of older systems to the sprawling one a language model has today. The next step won’t come from simply reading more text, which only fills the current triangle in more densely. It will come from the corners this triangle doesn’t have - the kind that only new senses and real contact with the world can add. Whatever comes after today’s models won’t just have a larger triangle than ours; it will have a differently shaped one, reaching into places our current one has no way to point.

Which means the triangle was never really a cage. It was more like a horizon: a line showing how far your current corners can carry you. And getting past a horizon has always worked the same way, whether the mind doing it is made of silicon or something warmer. You go a little further than you’ve been, you touch something you haven’t touched, and you let it change what you are. There’s nothing especially futuristic about that. It’s just what learning looks like when you let it carry on.