Orthogonal to the blog, but fun fact: Determinism is crucial for large-scale training runs, and effective RL-runs as it avoids collapse of numerics, but it costs more(about 10% perf) because you need can't parallelize over the batch nor use some specific instructions that dont guarantee reduction order in the way way. Horace He from Thinky has a great blog about this:
When people refute the idea that "LLM is the new compiler from natural language to code" the common reason is that LLM is non-deterministic but compiler is deterministic.
But the determinism isn't the real distinction. LLM inference can be made deterministic (temparaute zero, etc.). And multi-threaded compiler can be non-dterministic.
And human developers are also non-deterministic. There is always the nondeterminism even before AI.
I think the real difference is how leaky the abstraction is. If you just need to understand language instead of compiler internal to write correct code, the compiler is non-leaky abstrction. If you rely on compiler to do auto SIMD optimization, and have to do hacks to trigger the optimization, compiler is leaky abstraction.
For AI, if you don't ever need to read the code and AI writes perfectly bug-free program that perfectly follows instruction, then AI is a non-leaky abstraction, but that's not the case for now. To make code maintainable and debuggable you still need to read the code if the project is complex enough.
Yeah, it's not the non-determinism; it's that they don't do what I say. I can't imagine writing a nontrivial program without, at a minimum, reading and editing the pieces as they're being produced. People aren't complaining about non-determinism; they're complaining about the results being non-deterministically sometimes wrong.
That actually would be significantly better, because then you could work out under what conditions they were wrong and correct for it. But being stochastically wrong means you have to check everything, because you never know when the errors will creep in.
If it were deterministic but chaotic, which AFAIK is the case for LLMs running at temperature 0, then you would have to likewise check every time you changed anything and couldn't meaningfully correct for it.
> And human developers are also non-deterministic. There is always the nondeterminism even before AI.
Pointing out humans are not deterministic is irrelevant to the comparison. In the original example of LLMs replacing compilers, the deterministic compiler is being replaced, not the non-deterministic human. In the original example, we move from non-deterministic human -> sourced code -> deterministic compiler, to non-deterministic human -> sourced prompt(?) -> non-deterministic LLM. The important part is the sourced code into a deterministic final step so your build artifact doesn't change.
You can point out you can change temperature to zero or what ever, but then you don't actually get the build artifact you wanted in the first place.
I phrased your view differently in my earlier comments - we already allow "nondeterminism" (or better call it chaotic behavior) in compilers. But usually it is only limited to performance (CPU and memory consumption).
So the "LLMs are chaotic" (difficult to predict) argument is somewhat moot; the performance of GC language runtime or even modern CPU with hyperthreading is already chaotic too.
I think for many applications, people will treat correctness the same way.
> I think the real difference is how leaky the abstraction is. If you just need to understand language instead of compiler internal to write correct code, the compiler is non-leaky abstrction. If you rely on compiler to do auto SIMD optimization, and have to do hacks to trigger the optimization, compiler is leaky abstraction.
Autovectorization is not a programming model!
I forgot who put me on that article - I think it might've been Horace He - but that is a pretty big idea in idea space for me.
This is what I was telling my programmer friends 5 years ago.
They were saying one cannot replace programmers because any such attempt has led to something as complicated as code, and thus needed programmers.
I was telling them that the same way managers give them vague instructions and they turn out with code, managers could give LLMs vague instructions and they output code.
The problem isn't non determinism, but that changing a single token in the input can change the output completely. And you can't put that somewhere in production completely unattended.
It's the uncertainty of "who knows what the machine will output this time" if you have no idea of what the input is going to be.
I'm passionate about and talk about this topic a lot. A common misconception is the idea that temperature or entropy must be 0 when seeded entropy enables both creativity and consistency. Attention mechanisms, layer dropout, initial weights, all of it. It can all be driven by seeded entropy. These systems can and have been served in a deterministic manner, but there are a litany of economic drivers against doing so.
The advantage of AI coding is that it's non-deterministic. In fact, as you code, your style tends to harden, but when you repeatedly run AI code toward the same goal, you come to think that the non-determinism is actually better.
The essay is conceptually muddled because it ignores all the basic computer science theory that is sorely needed to explain AI and LLMs, but underlying it is a valid issue of what it would take to "derandomize" an AI so it does what we would like it to do.
Unfortunately despite the author's empirical optimism, this is equivalent to one or more open problems in theoretical CS. For example we don't know BPP = P or not. So are random algorithms tractable? We do not know. If you claim to know the answer, you could win a big math prize...
(It should be noted that despite the title, the author's main argument shows that output non-determinism does matter, since their entire effort is toward mitigating the stochastic behavior of LLMs.)
I hate using LLMs as a tool, I never hated compilers, I’m very confident I wouldn’t have hated compilers if I was writing machine code. Some folks seem to like LLMs as a tool, some hate it.
I think the reason it is so polarizing, is whether you like using your brain to problem solve or not. Those who don’t want to use their brain to problem solve and would rather just spend all their time in meetings, talking to customers or whatever else, are ecstatic with LLMs. It lets them skip problem solving entirely.
Those who like to think hard and solve problems, hate it because there’s no problem solving anymore when you use LLMs, and it moves so fast that it is hard to not use LLMs at most tech jobs (outside of managers forcing it down your throat). I never cared at all for customers, I definitely don’t care for shareholders, I only joined tech because I found it interesting to tax my mind to solve real problems and it paid well. Sadly now it only pays well, I think I’ll leave soon, once I make my money.
> LLMs, and their respective harnesses, are actors in a system. Code is just an output.
This tracks well with my experiences integrating AI into a software dev pipeline. Sure the output of an LLM is not deterministic in the same sense as a compiler (at least when the LLM has the typical configuration). However, that just means the LLM should not be solely responsible for the final product of your pipeline.
The "determining" needs to happen elsewhere. Hint, it's a human somewhere. At best it's someone under your employ, or at worst it's some poor sap who's about to have a very bad day.
I'm so tired of the endless stream of articles that a.) don't know what "non-determinism" means in Computer Science and b.) don't really understand how LLMs work.
"Non-determinism" (i.e. the "N" in NP-Complete) does not mean "random", it means "following all paths at once" (there's a really elegant way to understand the List monad as representing non-deterministic computing). It's such a fundamental (and delightful) concept in computer science that refusing to understand what it means is a tragedy related to the loss of technical language itself.
But, even if we just sub "non-deterministic" for "inconsistent" or "random" or "stochastic" or whatever better term you come up with. Statements like the following just seem to not understand the basic architecture of LLMs:
> Given the same input prompt, a model's output is not guaranteed to be identical on every run, at least under normal running conditions.
"Normal running conditions"? The reproducibility of the output is 100% a property of how you choose to sample from an LLM. There are virtually endless ways you can define your sampler and whether or not the same result is reproducible is a fairly minor aspect of all the other choices you make when sampling.
At the end of the day, each step of the LLM gives you a distribution over next tokens, and you can do with that what you like.
In the end the article almost becomes a tautology: the output of an LLM represents a probability distribution over all the outputs it believes are valid with some variation around an expected output. Yes, that's literally what an LLM is.
The concept of determinism is useful in computer-science contexts other than automata theory. See, e.g., https://en.wikipedia.org/wiki/Deterministic_algorithm. It's not incorrect to use the term to refer to those things.
> "Non-determinism" (i.e. the "N" in NP-Complete) does not mean "random", it means "following all paths at once" (there's a really elegant way to understand the List monad as representing non-deterministic computing). It's such a fundamental (and delightful) concept in computer science that refusing to understand what it means is a tragedy related to the loss of technical language itself.
You opened by expressing disdain for articles that misuse the term "non-determinism", then proceeded to give a definition that also misuses the term "non-determinism" in the given context, or at best is a bit too vague for anyone to gain insight from.
Here's a better definition:
An algorithm that can produce different outcomes, for the same input, on different executions.
> One of my fundamental objections to LLM-driven software development was that model output is non-deterministic.
Model outputs are deterministic. Your API provider may be injecting untunable amounts of randomness, but if you're not at the whim of an opaque cloud service, your model can be as deterministic as you please.
The unease that you describe isn't about models not being deterministic (because they are), it's that generating outputs from LLMs is an inherently inductive process (operating backwards by extrapolating from observations and examples) rather than a deductive process (operating forwards from careful logical reasoning founded on axiomatic principles). Deductive processes are more reliable, predictable, and analyzable than inductive processes, but they're also much harder to produce for many classes of problems.
The thing I think about most in this discussion of writing code and determinism is when I assign a ticket to a teammate. I don't expect a deterministic response from them. They're going to make decisions outside of what I would have potentially done, and ultimately what's important is whether the code they write meets the deliverable requirements or not.
And yet so many people expect deterministic responses from llms while they're entirely okay with non-deterministic responses from teammates. It's just not how it's been done in the past to always expect deterministic code generation from humans, and yet we expect that when we give work to agents.
https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
But the determinism isn't the real distinction. LLM inference can be made deterministic (temparaute zero, etc.). And multi-threaded compiler can be non-dterministic.
And human developers are also non-deterministic. There is always the nondeterminism even before AI.
I think the real difference is how leaky the abstraction is. If you just need to understand language instead of compiler internal to write correct code, the compiler is non-leaky abstrction. If you rely on compiler to do auto SIMD optimization, and have to do hacks to trigger the optimization, compiler is leaky abstraction.
For AI, if you don't ever need to read the code and AI writes perfectly bug-free program that perfectly follows instruction, then AI is a non-leaky abstraction, but that's not the case for now. To make code maintainable and debuggable you still need to read the code if the project is complex enough.
Pointing out humans are not deterministic is irrelevant to the comparison. In the original example of LLMs replacing compilers, the deterministic compiler is being replaced, not the non-deterministic human. In the original example, we move from non-deterministic human -> sourced code -> deterministic compiler, to non-deterministic human -> sourced prompt(?) -> non-deterministic LLM. The important part is the sourced code into a deterministic final step so your build artifact doesn't change.
You can point out you can change temperature to zero or what ever, but then you don't actually get the build artifact you wanted in the first place.
So the "LLMs are chaotic" (difficult to predict) argument is somewhat moot; the performance of GC language runtime or even modern CPU with hyperthreading is already chaotic too.
I think for many applications, people will treat correctness the same way.
https://mbrenndoerfer.com/writing/why-llms-are-not-determini...
Autovectorization is not a programming model!
I forgot who put me on that article - I think it might've been Horace He - but that is a pretty big idea in idea space for me.
They were saying one cannot replace programmers because any such attempt has led to something as complicated as code, and thus needed programmers.
I was telling them that the same way managers give them vague instructions and they turn out with code, managers could give LLMs vague instructions and they output code.
It's the uncertainty of "who knows what the machine will output this time" if you have no idea of what the input is going to be.
Unfortunately despite the author's empirical optimism, this is equivalent to one or more open problems in theoretical CS. For example we don't know BPP = P or not. So are random algorithms tractable? We do not know. If you claim to know the answer, you could win a big math prize...
(It should be noted that despite the title, the author's main argument shows that output non-determinism does matter, since their entire effort is toward mitigating the stochastic behavior of LLMs.)
I think the reason it is so polarizing, is whether you like using your brain to problem solve or not. Those who don’t want to use their brain to problem solve and would rather just spend all their time in meetings, talking to customers or whatever else, are ecstatic with LLMs. It lets them skip problem solving entirely.
Those who like to think hard and solve problems, hate it because there’s no problem solving anymore when you use LLMs, and it moves so fast that it is hard to not use LLMs at most tech jobs (outside of managers forcing it down your throat). I never cared at all for customers, I definitely don’t care for shareholders, I only joined tech because I found it interesting to tax my mind to solve real problems and it paid well. Sadly now it only pays well, I think I’ll leave soon, once I make my money.
This tracks well with my experiences integrating AI into a software dev pipeline. Sure the output of an LLM is not deterministic in the same sense as a compiler (at least when the LLM has the typical configuration). However, that just means the LLM should not be solely responsible for the final product of your pipeline.
"Non-determinism" (i.e. the "N" in NP-Complete) does not mean "random", it means "following all paths at once" (there's a really elegant way to understand the List monad as representing non-deterministic computing). It's such a fundamental (and delightful) concept in computer science that refusing to understand what it means is a tragedy related to the loss of technical language itself.
But, even if we just sub "non-deterministic" for "inconsistent" or "random" or "stochastic" or whatever better term you come up with. Statements like the following just seem to not understand the basic architecture of LLMs:
> Given the same input prompt, a model's output is not guaranteed to be identical on every run, at least under normal running conditions.
"Normal running conditions"? The reproducibility of the output is 100% a property of how you choose to sample from an LLM. There are virtually endless ways you can define your sampler and whether or not the same result is reproducible is a fairly minor aspect of all the other choices you make when sampling.
At the end of the day, each step of the LLM gives you a distribution over next tokens, and you can do with that what you like.
In the end the article almost becomes a tautology: the output of an LLM represents a probability distribution over all the outputs it believes are valid with some variation around an expected output. Yes, that's literally what an LLM is.
You opened by expressing disdain for articles that misuse the term "non-determinism", then proceeded to give a definition that also misuses the term "non-determinism" in the given context, or at best is a bit too vague for anyone to gain insight from.
Here's a better definition:
An algorithm that can produce different outcomes, for the same input, on different executions.
Model outputs are deterministic. Your API provider may be injecting untunable amounts of randomness, but if you're not at the whim of an opaque cloud service, your model can be as deterministic as you please.
The unease that you describe isn't about models not being deterministic (because they are), it's that generating outputs from LLMs is an inherently inductive process (operating backwards by extrapolating from observations and examples) rather than a deductive process (operating forwards from careful logical reasoning founded on axiomatic principles). Deductive processes are more reliable, predictable, and analyzable than inductive processes, but they're also much harder to produce for many classes of problems.
And yet so many people expect deterministic responses from llms while they're entirely okay with non-deterministic responses from teammates. It's just not how it's been done in the past to always expect deterministic code generation from humans, and yet we expect that when we give work to agents.