AGI won't be one big brain
We keep picturing AGI as one enormous model that finally got big enough. The systems that are actually getting good are collections of narrower parts that have learned to call each other.
The film version of artificial general intelligence is one enormous all-knowing brain. It makes for a good villain and a bad prediction. I expect AGI to arrive as a lot of narrower systems that learned to call each other, rather than as a language model that finally got big enough.
Your brain works this way already
Your brain is a set of specialised regions. The visual cortex handles sight, the motor cortex handles movement, and they pass information between them constantly. What you experience as a single mind is those parts coordinating well enough that you never notice the joins.
Software went the same way. We stopped building one huge application and moved to smaller services that call each other, because that’s easier to fix, easier to scale, and easier to hold in your head. There’s no obvious reason intelligence should be the exception.
Hiding the complexity is the hard part
Building the specialists is the easy half. The hard part is the layer that works out which ones to call, in what order, and stitches their answers back together, while the person using it just states a goal and gets a result. That layer is what makes the system useful, and it has very little to do with how big any individual model is.
What’s missing
People talk about AI becoming smarter than us as though it’s a future event. It already is, in narrow slices. It has beaten the best Go players for years and it reads some scans with remarkable accuracy. Raw capability in a slice was never the thing holding it back.
What’s missing is context and reach. A model doesn’t have the enormous pile of ordinary real-world knowledge you accumulated by being alive, and mostly it can’t act on the physical world. A toddler learns by touching things and seeing what happens. AI needs the equivalent: better sensors, better actuators, more varied data, more ways to actually do something and find out whether it worked. Systems get better at real-world tasks as they get better connected to the world. No individual model has to get bigger for that to happen.
So I’d stop watching parameter counts. The interesting work is in the wiring.