It is definitely an idea discussed quite a bit on X (and in the broader AI community) as i read around.

The core argument is that no single LLM is truly best-in-class at absolutely everything and never can it be (in the short/medium term).

Different models have relative strengths (e.g., one excels at reasoning, another at creative writing, coding, fact-checking, or domain-specific tasks like DeFi or medical analysis).

So, routing tasks to specialized LLMs (or combining/ensembling them) often beats forcing one generalist "do-it-all" model to handle every step.