AI tools are becoming specialized. Let's say you want to generate a video clip. You'll use one AI tool to create images for the storyboard, then an other one to convert images to video, an other one for voice, an other one for underlying music, and so on. Just like you would need a whole team (of humans) to produce a video clip today.

Will it be a winner-takes-it-all for each type of work? Probably not, I think each AI tool will have its own "personality" or specialization: this one is better at making photorealistic images, the other it better at illustrations, an other one is the best for natural landscapes, and so on.

Small models should manage simple tasks and turn to larger models for complex ones—not the other way around. This is why i’m bullish on Apple’s approach.

Dad, geek, conversation liquidity provider. vrypan.net (home), /weatherxm (work), @l--o--l (fun). Also https://paragraph.xyz/@purplesubmarine

“autonomous swarm intelligence” 
https://warpcast.com/gm8xx8/0xc843773e

It's similar to farcaster clients, each one has its own personality. 

I picture each of these as close to NP-hard. so, in a way, this is way better bang-for-buck than pretraining general purpose models/agents.

and as @vitalik.eth describes in the glue-vs-coprocessor blogpost, I think agents are thalamus-vs-data-in-distribution. we don't need a bigger model supervising smaller models, just a bigger ecosystem that can index smaller models.

I picture each of these as close to NP-hard. so, in a way, this is way better bang-for-buck than pretraining general purpose models/agents.

and as @vitalik.eth describes in the glue-vs-coprocessor blogpost, I think agents are thalamus-vs-data-in-distribution. we don't need a bigger model supervising smaller models, just a bigger ecosystem that can index smaller models.
https://x.com/m_j_rossman/status/1712564985316909389