Sjors Provoost on Nostr: Then there is the secondary problem that these intermediate "thoughts", especially if ...

Then there is the secondary problem that these intermediate "thoughts", especially if only the model itself needs to understand it, will gradually drift away from readable English.

I suppose we can always just use another LLM to translate it?

But this reminded me another trend, namely that humans are starting to use LLM'ism more often. And we get better at reading Sloplish, e.g. the sentence I quoted elsewhere:

> Let’s be precise about attribution, because breach coverage collapses when speculation hardens into fact.

If over time it develops a less verbose internal "thinking" language, it might then also output less verbose language. A new language might then evolve in the interaction with humans: Compact Sloplish