Out of curiosity, which models are fully deterministic? I was under the impression that all LLMs were fundamentally probabilistic.
I don't think the state of the art LLM providers let you do this anymore (?), but they certainly could if they wanted to, and you can do it yourself with a local model.
I can see how you'd nitpick this, but to me this is a deterministic algorithm that just happens to be running on nondeterministic hardware.
But, I think you're tricking yourself on determinism. You'll say something like "I know if I ask an LLM what 1+1 is, it will answer 2", but the thing is, you don't. You have to run the LLM first to figure out it's output. And when you send in just a few bits of text, it's outputs are going to be rather limited.
But this all breaks when it hits the real world. Inputs are unpredictable. Hence while LLM outputs, like humans, are probabilistic, you can't figure out what it's going to be until you ask. And in any high complexity data gathering environment you're going have a difficult time ensuring your entire systems conditions are the same.
System consistency is very hard, once you start running thousands of processors in an agentic loop small errors accrue and timing starts differing and the system will take non-deterministic paths.
If I ask an llm to "add 2 and 2" is and it replies corectly, then I ask for "the sum of 2 and 2" and it replies "banana" that is a lack of predictability and consistency but not a lack of determinism.
As long as it produces the same output for a given input, unhinged or not, it is deterministic.
Your example at the end of different systems feeding data to each other is non-deterministic only at the system level, not the individual llm level.
Which is why llms aren't agents and depend on harnesses. The llm itself doesn't have a continual loop built in, that would be very power hungry. The harness works as the orchestrator of memory and action. Now, I can't think of a reason why an LLM couldn't bootstrap its own harness, but in general it sounds like a very dumb idea to actually build that from an AI safety perspective.
This discussion falls under the idea and refutation of the Chinese Room. The room may have no idea what Chinese characters are, but the system does.
Given the same prompts and the same weights, one can get the same answer each time.
In practice, there are a number of optimizations that makes the results dependent upon thing we give up control of to increase performance, meaning the results end up being effectively non-deterministic. But, if you are willing to run it in a slower mode so we don't do some steps out of order to speed things up and don't batch results (or if you consider the determinism of a given batch of requests rather than individual requests), then the same input gets the same output.
As I understand it, if you turn down the temperature to 0 you get repeatable behavior - EXCEPT - on large servers with lots of users - the GPU can sometimes produce slightly different results based on batch size.
In practice - on a multitasking OS with input from multiple human users - it's hard to get it deterministic because of that GPU scheduling thing I mentioned.
The abstracted design of the machine is meant to be deterministic, but you can't predict before running any command whether or not it will complete because there are externalities that effect the outcome.
Electromagnetic interference even happens in-chip where an electron can accidentally escape it's wire and enter another, possibly resulting in an error, but not every time.
It's even been used as an attack vector where rapidly flipping a bit increases the likelihood that a neighbor bit is also flipped, but the method is probabalistic, not deterministic.