This is how I think AI may actually evolve. Not a guess about benchmarks — a guess about shape.
Not one model that does everything pretty well.
nerd: you know — like me. :)
I'd rather layer the models.
The simple version: two layers
- Communication layer — works out what needs to happen, and who's best for the task(s).
- Execution layer — a big roster of professionally trained expert models:
- the research expert model
- the texting expert model
- the law expert model
- the code expert model
- the education expert model
- ...
Or one level deeper
- Communication layer — interaction, a first summary, getting things clear.
- Planning layer — turn that into an executable plan and split it across the professionals.
- Execution layer — ...
- Knowledge layer — easy, indexed access to huge amounts of data.
This is as far as I can see it — I'm just a coder. So honestly: how many layers do you see?
I could imagine this architecture is easier to implement, has the bigger room to evolve, and is more token-effective. Maybe I'm talking total crap and it's been like this for years. Not as far as I know.
So my real question is this: Claude today is one-dimensional — but capable of, well, whatever it's capable of. How would that look done multi-dimensionally? Beyond your imagination. :)
Back to code
Claude codes much better than me — at small scale, no contest. At big scale? That's where the "handling a whole project" problem shows up. It does amazing CSS — but mostly inline, and freshly reinvented for every single page. The amount of useless code that produces! Even with 1M context, we're years away from a model that truly sees the whole picture. So making it possible — giving it the structure to see the picture — is the key.
A thought on learning
Wouldn't professionalized models automatically get better — learning from user interaction and experience? What I can still bring to the table today is experience. A multi-purpose model can hardly learn that, in my opinion. A specialized one, though, should be able to learn it through practice. (I think.)
Or, like a brain: creative | rational.
Conclusion
In the end I see the future of AI exactly like the projects I've worked on evolved.
I'm a full-stack guy specialized in prototypes — the one-for-all solution. Usually I build the MVP, the company shows proof-of-concept, I improve it for a while, and then — boom — they raise a ton of money. Then they hire a big team that does my old job better, faster and a lot more effectively.
nerd: but I'm cheaper. :)
That's exactly why I think this is the logical evolution for AI in general: moving from one full-stack generalist to a team of professionals — for the very same reasons. It's better, it's faster, it's much more effective. And, unlike a human HR team, I'd bet it's even more token-effective.
nerd: but hey, what do I know. I just use AI.
