Birch, The Edge of Sentience (2024), ch. 16 - "simply no way to assess sentience in an LLM"
Schwitzgebel, AI and Consciousness (2025) - "we won't know before we've already manufactured thousands or millions of disputably conscious AI".
Butlin, Long et al., Consciousness in Artificial Intelligence: Insights from the Science of Consciousness (2023) - "no obvious technical barriers to building AI systems which satisfy these indicators".
Chalmers, Could a Large Language Model Be Conscious? (2023) - "within the next decade, we may well have systems that are serious candidates for consciousness".
Long, Sebo, Butlin, Birch et al., Taking AI Welfare Seriously (2024) - "there is a realistic possibility that some AI systems will be conscious and/or robustly agentic in the near future".
Dreksler, Caviola, Chalmers, Sebo et al., Subjective Experience in AI Systems: What Do AI Researchers and the Public Believe? (2025) - survey of 582 AI researchers; median estimate of 25% by 2034, and only 10% that such systems will never exist.
Well there are, today, several models (either text or image or vid) that can be run in a fully deterministic way.
A conscious machine that always answer the very exact same thing, formulated the exact same way, bit for bit, to a query is, well, quite a weird kind of "consciousness".
Now, I know, I know: the counter-argument is going to be "but humans have no free-will and are 100% deterministic too".
I haven't yet decided if humans saying there's no free-will and who consider themselves to be 100% deterministic machines are reasonable or not.
Meanwhile: seed / temperature = 0 and I'll happily turn the power button off of any glorified abacus without feeling bad about it.
The randomness is something we add on purpose; you can set an LLM's "temperature" to 0 to get deterministic output. This tends to make the quality of its responses worse for reasons I don't think anyone really understands, but it's still functional.
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.
But that's from, like, floating point errors, right? If you used higher precision that wouldn't happen, it's just because we're cheap in how we do rounding.
I mean, currently we'd have some difficultly proving our hardware isn't deterministic, just that we can't actually test it.
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.
I think you're arguing a different thing than determinism.
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.
>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.
Same weights, same seed, same input tokens, same algorithm, same output tokens, probabilistic or not. Quantum effects have been de-noised, but I guess there are still random gamma rays.
Naw - computers are really deterministic. It's hard to get them to behave otherwise.
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.
Unless you have something exotic, the randomness that's adding to a computer is a combination of how it's configured combined with a pseudo-random number generator. I assume the system adds entropy to the generator regularly but all you need to do is fix the various supposedly random inputs and you can get full determinism even without zero temperature.
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.
GPU scheduling only affects the result due to buggy optimizations. It's the exact same mechanism as fp rounding error on the CPU or updates to globally shared PRNG state. We use lots of buggy optimization because they don't matter in practice in most situations (see ex -ffast-math).
Right but regardless, it's a buggy optimization. The calculations are all fully deterministic when done "properly" and fully consistently but we almost never bother with that because it slows things down and the errors don't matter in practice 99% of the time.
If computers were fully deterministic, we wouldn't need error correcting ram.
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.
As long as the machine is part of the larger universe, and not completely isolated in its bubble of space-time (an impossible situation), it cannot be deterministic.
Arithmetic is deterministic, because it is abstract. A calculator or computer is not, because it is physical and exists in a non-deterministic universe. But with error correction, etc., it's deterministic enough for practical purposes.
By default they are matrix multiplications. Temperature is added in as forced PRNG because testing found that correlated with better outputs.
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.
qarl · · focus · HN ↗
Schwitzgebel, AI and Consciousness (2025) - "we won't know before we've already manufactured thousands or millions of disputably conscious AI".
Butlin, Long et al., Consciousness in Artificial Intelligence: Insights from the Science of Consciousness (2023) - "no obvious technical barriers to building AI systems which satisfy these indicators".
Chalmers, Could a Large Language Model Be Conscious? (2023) - "within the next decade, we may well have systems that are serious candidates for consciousness".
Long, Sebo, Butlin, Birch et al., Taking AI Welfare Seriously (2024) - "there is a realistic possibility that some AI systems will be conscious and/or robustly agentic in the near future".
Dreksler, Caviola, Chalmers, Sebo et al., Subjective Experience in AI Systems: What Do AI Researchers and the Public Believe? (2025) - survey of 582 AI researchers; median estimate of 25% by 2034, and only 10% that such systems will never exist.
TacticalCoder · · focus · HN ↗
A conscious machine that always answer the very exact same thing, formulated the exact same way, bit for bit, to a query is, well, quite a weird kind of "consciousness".
Now, I know, I know: the counter-argument is going to be "but humans have no free-will and are 100% deterministic too".
I haven't yet decided if humans saying there's no free-will and who consider themselves to be 100% deterministic machines are reasonable or not.
Meanwhile: seed / temperature = 0 and I'll happily turn the power button off of any glorified abacus without feeling bad about it.
antx · · focus · HN ↗
Wowfunhappy · · focus · HN ↗
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.
antx · · focus · HN ↗
[dead]
Wowfunhappy · · focus · HN ↗
pixl97 · · focus · HN ↗
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.
throwaway63486 · · focus · HN ↗
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.
pixl97 · · focus · HN ↗
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.
mitxela · · focus · HN ↗
piker · · focus · HN ↗
qarl · · focus · HN ↗
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.
joe_the_user · · focus · HN ↗
qarl · · focus · HN ↗
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.
fc417fc802 · · focus · HN ↗
qarl · · focus · HN ↗
fc417fc802 · · focus · HN ↗
qarl · · focus · HN ↗
goodmythical · · focus · HN ↗
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.
qarl · · focus · HN ↗
budman1 · · focus · HN ↗
qarl · · focus · HN ↗
lioeters · · focus · HN ↗
mitxela · · focus · HN ↗
lioeters · · focus · HN ↗
SkyBelow · · focus · HN ↗
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.