On: Cache-to-Cache: Direct Semantic Communication Between LLMs (2025)

[cubefox]

So the models will not only be using more and more Neuralese in their CoT (like GPT-6), but different agents will also be able to communicate with each other in Neuralese. It's not looking good for monitorability.

[sparky_twofort → cubefox]

Is Neuralese in no way decodable into a human-interpretable system? Genuine question -- I don't know the answer.

[Y_Y → sparky_twofort]

Definitely decodable, that's what's being done now

[cubefox → Y_Y]

It's not.

[TeMPOraL → sparky_twofort]

Probably not without a sufficiently powerful LLM from same family, or something equivalent, to act as a translation layer, which includes the risk of the translator lying to you.

[jephs → cubefox]

This is from like a year ago.