How could an LLM-based OS become reality in the light of non-determinism & hallucination?

[ChristianWeyer]

I love the idea!!!

But... how could an LLM-based OS become reality in the light of non-determinism & hallucination?

[dongyuanjushi]

Thanks a lot for your attention and recognition.

It's a great question. Actually, we are also working on some controllable mechanisms like control flow, automata to improve the determinism of LLM's each decision. And on the other hand, we imagine that such mechanism may be incorporated into the LLM training to enable LLM to follow instructions more precisely in the future. There is still a long way to go, but we will keep moving towards it.

Best,

[robinsonkwame]

The Concordia team has an open issue on a similar feature where they need repeatability in their LLM based social simulation framework, might be worth a chat with them.