In non-algorithmic music software, instruments with overtones is common, and the solution is to just place the notes yourself ensuring they're compatible.
But, in nondeterministic software like mine, that's not always an option, and instead a clever a algorithm is the best answer.
But, in nondeterministic software like mine, that's not always an option, and instead a clever a algorithm is the best answer.
October 20, 2025 at 9:07 PM
In non-algorithmic music software, instruments with overtones is common, and the solution is to just place the notes yourself ensuring they're compatible.
But, in nondeterministic software like mine, that's not always an option, and instead a clever a algorithm is the best answer.
But, in nondeterministic software like mine, that's not always an option, and instead a clever a algorithm is the best answer.
How does "nondeterministic layers of coding abstraction" make any sense? Whatever language or abstraction should produce consistent machine code, at least for a given archiecture/compiler/set of libraries. How would you even debug software if your code produces inconsistent results?
October 18, 2025 at 7:21 PM
How does "nondeterministic layers of coding abstraction" make any sense? Whatever language or abstraction should produce consistent machine code, at least for a given archiecture/compiler/set of libraries. How would you even debug software if your code produces inconsistent results?
no, it's "it depends", like every other fucking technology! that's the whole point! sometimes it does cool shit. And also, when people use a nondeterministic tool in places it doesn't belong, because they're buying into hype, they fuck everything up. That's not transformational, actually.
October 17, 2025 at 5:07 AM
no, it's "it depends", like every other fucking technology! that's the whole point! sometimes it does cool shit. And also, when people use a nondeterministic tool in places it doesn't belong, because they're buying into hype, they fuck everything up. That's not transformational, actually.
With some trial and error it can certainly feel like you get prompt writing and the various quirks and can coax it into doing what you want except, wait it's nondeterministic so it won't won't necessarily work the same way next time and you're at the bottom of the curve again.
October 17, 2025 at 3:42 AM
With some trial and error it can certainly feel like you get prompt writing and the various quirks and can coax it into doing what you want except, wait it's nondeterministic so it won't won't necessarily work the same way next time and you're at the bottom of the curve again.
i mean, no part of the existing unreal engine required massive data centers during development and it’s extremely unclear to me why you’d want to offload that to a nondeterministic process
October 16, 2025 at 6:32 PM
i mean, no part of the existing unreal engine required massive data centers during development and it’s extremely unclear to me why you’d want to offload that to a nondeterministic process
Happy Birthday Dana Stewart Scott! Scott received the 1976 #ACMTuringAward along with Michael O. Rabin, for their joint paper "Finite Automata and Their Decision Problem," which introduced the idea of nondeterministic machines. amturing.acm.org/award_winner...
October 11, 2025 at 11:02 AM
Happy Birthday Dana Stewart Scott! Scott received the 1976 #ACMTuringAward along with Michael O. Rabin, for their joint paper "Finite Automata and Their Decision Problem," which introduced the idea of nondeterministic machines. amturing.acm.org/award_winner...
“nondeterministic automation”
One day the industry will recognize the drawbacks of AI agents and nondeterministic automation, and rediscover the UNIX philosophy of chaining together small purpose built tools in a low cost and predictable way, otherwise known as shell scripts.
October 9, 2025 at 2:51 PM
“nondeterministic automation”
One day the industry will recognize the drawbacks of AI agents and nondeterministic automation, and rediscover the UNIX philosophy of chaining together small purpose built tools in a low cost and predictable way, otherwise known as shell scripts.
October 9, 2025 at 11:22 AM
One day the industry will recognize the drawbacks of AI agents and nondeterministic automation, and rediscover the UNIX philosophy of chaining together small purpose built tools in a low cost and predictable way, otherwise known as shell scripts.
“nondeterministic automation” is gonna join the oxymoron lexicon alongside “military intelligence” and “jumbo shrimp”
October 9, 2025 at 2:02 PM
“nondeterministic automation” is gonna join the oxymoron lexicon alongside “military intelligence” and “jumbo shrimp”
The thing that terriffies me of nondeterministic errors in automation is that they become very difficult perhaps impossible to isolate and thus fix. There's a reason step one of bugfixing is "reproduce the bug"
Folks that have experience with deterministic automation in (esp. in distributed systems) understand compounding error rates well enough to be very sceptical about throwing non-determinism in that mix.
October 9, 2025 at 12:42 PM
The thing that terriffies me of nondeterministic errors in automation is that they become very difficult perhaps impossible to isolate and thus fix. There's a reason step one of bugfixing is "reproduce the bug"
We fundamentally accept, as consumers of software, a shocking amount of what is, functionally, nondeterministic behaviour in software today. Software systems are so complex and so poorly understood and shabbily implemented that the tradeoff here is illegible to most folks.
October 9, 2025 at 12:24 PM
We fundamentally accept, as consumers of software, a shocking amount of what is, functionally, nondeterministic behaviour in software today. Software systems are so complex and so poorly understood and shabbily implemented that the tradeoff here is illegible to most folks.
Contracts should not have nondeterministic strings put in them!
October 8, 2025 at 11:51 PM
Contracts should not have nondeterministic strings put in them!
I’ve been saying this for… oh several years now. And it’s always been obvious because we’ve always known the output of these LLMs is NONDETERMINISTIC.
Yet OpenAI persisted in lying to people that they’d eventually solve hallucination.
Yet OpenAI persisted in lying to people that they’d eventually solve hallucination.
September 21, 2025 at 12:25 PM
I’ve been saying this for… oh several years now. And it’s always been obvious because we’ve always known the output of these LLMs is NONDETERMINISTIC.
Yet OpenAI persisted in lying to people that they’d eventually solve hallucination.
Yet OpenAI persisted in lying to people that they’d eventually solve hallucination.
LLM output is essentially hardware dependent which makes it practically non-deterministic plus the sampling techniques and the probability that most use some math package that probably effects their output.
All of this adds up to it being just easier to say they’re nondeterministic.
All of this adds up to it being just easier to say they’re nondeterministic.
September 21, 2025 at 2:08 PM
LLM output is essentially hardware dependent which makes it practically non-deterministic plus the sampling techniques and the probability that most use some math package that probably effects their output.
All of this adds up to it being just easier to say they’re nondeterministic.
All of this adds up to it being just easier to say they’re nondeterministic.
How come different people giving the same prompt to ChatGPT get different answers?
A new article from ThinkingMachines explains why LLMs are nondeterministic, and it largely comes down to how parallelization and precision are handled.
🧵
A new article from ThinkingMachines explains why LLMs are nondeterministic, and it largely comes down to how parallelization and precision are handled.
🧵
September 13, 2025 at 10:16 PM
How come different people giving the same prompt to ChatGPT get different answers?
A new article from ThinkingMachines explains why LLMs are nondeterministic, and it largely comes down to how parallelization and precision are handled.
🧵
A new article from ThinkingMachines explains why LLMs are nondeterministic, and it largely comes down to how parallelization and precision are handled.
🧵
#TodaysNondeterministicAlgorithm solves the subset sum problem: is there a subset of S having sum k? This one shows a standard “design pattern” for nondeterministic algorithms: for each element of the set S, you just guess if you’re taking it or not, then check if you were right.
November 26, 2024 at 10:07 AM
#TodaysNondeterministicAlgorithm solves the subset sum problem: is there a subset of S having sum k? This one shows a standard “design pattern” for nondeterministic algorithms: for each element of the set S, you just guess if you’re taking it or not, then check if you were right.
the nondeterministic behavior makes it harder to test. the feedback loop is not as straightforward as with a deterministic system
January 21, 2025 at 2:20 PM
the nondeterministic behavior makes it harder to test. the feedback loop is not as straightforward as with a deterministic system
Generative AI has made it easier than ever for companies to build products quickly. However, LLMs are inherently nondeterministic and unpredictable. Integrating LLMs into products demands an unprecedented level of quality assurance – requiring new strategies for continuous evaluation.
June 25, 2025 at 7:18 PM
Generative AI has made it easier than ever for companies to build products quickly. However, LLMs are inherently nondeterministic and unpredictable. Integrating LLMs into products demands an unprecedented level of quality assurance – requiring new strategies for continuous evaluation.
An individual MCP tool may or may not be deterministic based on its implementation.
How an AI agent chooses to use MCP tools is nondeterministic.
How an AI agent chooses to use MCP tools is nondeterministic.
June 19, 2025 at 1:11 AM
An individual MCP tool may or may not be deterministic based on its implementation.
How an AI agent chooses to use MCP tools is nondeterministic.
How an AI agent chooses to use MCP tools is nondeterministic.
nondeterministic choices are constrained. Our results are backed by modular program-logic and compiler-correctness theorems, and they integrate into a neat end-to-end theorem in the Coq proof assistant. [6/6 of https://arxiv.org/abs/2504.15550v1]
April 23, 2025 at 5:59 AM
nondeterministic choices are constrained. Our results are backed by modular program-logic and compiler-correctness theorems, and they integrate into a neat end-to-end theorem in the Coq proof assistant. [6/6 of https://arxiv.org/abs/2504.15550v1]
Sometimes I try to ask it something that is complex and technical but could plausibly benefit from a little nondeterministic sampling across the relevant text space, and then it comes back with eminently plausible horseshit and I’m like “JUST SO YOU KNOW I NEVER BELIEVED IN YOU ANYWAY 😡”
August 27, 2025 at 2:18 AM
Sometimes I try to ask it something that is complex and technical but could plausibly benefit from a little nondeterministic sampling across the relevant text space, and then it comes back with eminently plausible horseshit and I’m like “JUST SO YOU KNOW I NEVER BELIEVED IN YOU ANYWAY 😡”
The nondeterministic browser
October 23, 2025 at 12:33 AM
The nondeterministic browser
Wait, by nondeterministic automation, are we talking about "triage customer support requests" or "monitor fleets of servers and fix things when they go down"?
Nondeterminism isn't great in either case but far more tolerable for one of those use cases.
Nondeterminism isn't great in either case but far more tolerable for one of those use cases.
October 9, 2025 at 8:42 PM
Wait, by nondeterministic automation, are we talking about "triage customer support requests" or "monitor fleets of servers and fix things when they go down"?
Nondeterminism isn't great in either case but far more tolerable for one of those use cases.
Nondeterminism isn't great in either case but far more tolerable for one of those use cases.
Why do such failures occur? These next-token-prediction models are nondeterministic and can be fragile. And they’re not getting consistently better over time—OpenAI’s latest models like o3 and o4-mini show higher hallucination rates compared to previous versions. (3/5)
April 24, 2025 at 6:21 PM
Why do such failures occur? These next-token-prediction models are nondeterministic and can be fragile. And they’re not getting consistently better over time—OpenAI’s latest models like o3 and o4-mini show higher hallucination rates compared to previous versions. (3/5)
nondeterministic signing is fine, but your sigs probably aren't low-s
October 8, 2025 at 5:34 PM
nondeterministic signing is fine, but your sigs probably aren't low-s