upvote
The craziest for me is companies that sticking stochastic agents into automated business processes and expecting stable/reliable outcomes. Businesses want deterministic processes in the vast majority of cases.
reply
People are stochastic. You build reliable processes out of unreliable parts with feedback and self-correcting mechanisms. AI is not actually magically special in this regard. It has higher variance and we're still figuring out how to get all the tradeoffs right.
reply
People live in very stochastic and volatile environments and they manage that in ways no LLMs currently ever can. (ie: imagine sending an LLM all the data - sensory/auditory/etc… - that a human receive)

People’s job is to partially reign in this volatile environment by creating processes with stable output.

reply
The big problem is that a person making a mistake can be taught to not make that mistake again. That's also not foolproof but at least it works a lot of the times. AI are unteachable, if you have given them a good prompt and they do something wrong 90% of the time you are shit out of luck.

That is to say I do agree that building reliable processes out of unreliable parts with feedback is the modus operandi. However AI cannot meaningfully handle feedback and learn. And that is a key unsolved problem.

reply
> AI are unteachable, if you have given them a good prompt and they do something wrong 90% of the time you are shit out of luck.

If the Model makes repeated mistakes on the same subject matter, you can update your agent.md file, or you can add skills to deal with specific prompts, or you provide a better default harness.

The whole idea of coding agents is their harness makes a big difference vs a pure raw model.

> However AI cannot meaningfully handle feedback and learn

How do you think models are created? They are trained on feedback and learn.

Its not cheap but you can post train models. This is how custom models are mode, that deal with specific tasks more efficiently and accurately.

Example ... Composer? Its base Kimi v2.5 model that has been post-trained 2 weeks, to create Composer 2.5, what is a much better coding model.

Its literally trained to make less mistakes by feeding it correct data. Hell, a lot of the models you are using, are often the same base model, where v2.0 was the initial released model but the model keeps training, so when they release v2.1, its still the same model, but with more training time on feedback provided to v2.0.

LLM Models are not a cake you cook one time and they are done, and you start from zero again. If you have the money, and a powerful server setup, you can take a model like GLM 5.2 and post-train it, to reduce specific errors. Sure, you need a ton of money because its a large model.

But people have been doing this with 5M, 100M, 1B, 5B models for a long time already. To the point that some of the small models can do specific tasks, almost or better then some of the huge more general trained models.

reply
> If the Model makes repeated mistakes on the same subject matter, you can update your agent.md file ...

That's all just prompting.

> How do you think models are created? They are trained on feedback and learn.

No one is post training models on a single mistake. At least I have not seen it. I also doubt it is effective. Post-training on a single failure will not meaningfully change the model. That even sidesteps the entire problem that you don't even have access to models if you use a provider like anthropic/openai

reply
> That's all just prompting

And telling someone not to repeat a mistake is… ?

reply
Are you really making the case that teaching a person how to work is equivalent to prompting an AI?
reply
I don't know of any modern workflows that rely on "we'll tell the person not to do it again", though. There's a reason that companies have adopted blameless postmortems, because if your response to the DB going down is "It's fine, Kevin learns and next time he won't misuse the prod credentials", you are guaranteeing prod will go down again in the same way at some point.
reply
Case law
reply
Literally the entirety of the worlds infrastructure relies on that. In the past we had (literally) had nuclear war hinging on a single person just deciding that some data point is an artifact.
reply
Every modern workflow implicitly relies on that. No infrastructure is fully robust. There's a senior DB person who has learned many things many times over who could bring down most of the US power grid.
reply
> The big problem is that a person making a mistake can be taught to not make that mistake again. That's also not foolproof but at least it works a lot of the times. AI are unteachable, if you have given them a good prompt and they do something wrong 90% of the time you are shit out of luck.

I feel like this line of thinking is kind of an unfair comparison. I'm not saying LLMs are magical beings that can suddenly learn by themselves after getting something wrong, but your "person making mistake then being corrected" assumes you do tell the person about the mistake and tell them to avoid doing the same mistake in the future, but for the "LLM making mistake" example you then intentionally avoid letting the prompt being changed in response to the mistake, which would be the "then being corrected" part on the LLM side of the comparison.

Similarly, if you just let a person make a mistake and don't let them know about the mistake, they might keep making that same mistake over and over again.

If you update how you use the LLM as you discover what mistakes it does, just like you'd correct a person, then you can use an LLM and also the LLM can "be taught to not make that mistake again".

reply
I'm not against the prompt being changed, the point I was making is that an LLM is prone to the exact same mistakes even if you change the prompt. A trivial example is the very basic character counting mistake, I just asked chatgpt:

> How many p are in strawperry?

> There are 0 “p”s in strawperry.

And I can trigger the same mistake with various words even when adjusting the prompt many times. So I cannot teach chatgpt to correctly count characters.

reply
"AI are unteachable, if you have given them a good prompt and they do something wrong 90% of the time you are shit out of luck."

please take a look at the error(s) made in the prior run. what could've been done better? create or modify an existing skill to emphasize this, or suggest additional language in AGENTS.md.

reply
It will return a bunch of relevant-sounding insight, modify skills and context files… Then do the same error again.

We’re not at the point where AI is capable of knowing what went wrong and self-aware enough to understand how it could reliably change its own behavior.

For months I’ve been trying to have the agents stop manually writing our auto-generated SQL migrations and run the command that generates them instead. SOTA models insist on occasionally getting it wrong.

reply
> However AI cannot meaningfully handle feedback and learn.

Well this is the central bet of AI coding isn't it? We, the humans-in-the-loop, get better at knowing ahead of time which patterns AI will handle better than others, all the while the models actually get better.

reply
Indeed. Any meaningful AGI/ASI will have to have a form of memory / continual learning. Sam Altman said last year that this will be the focus for GPT-6.

The whole "soul.md" stuff today is a poor approximation to that. But I wonder whether it will grow into it, like chain of thought prompting grew into reasoning models.

reply
LLM's as a technology - currently - are stateless. The memory layer is controlled by the agent. I m surprised with the gpt-6 reference unless it has to do with vertical integration between the agent and the layer.
reply
Yeah, I am aware of this statelessness.

This is what I was referring to:

https://www.cnbc.com/2025/08/19/sam-altman-on-gpt-6-people-w...

It's marketing speak, but the goal is clearly there, no idea how achievable.

reply
How it works under the hood is separate from selling the feature.
reply
Please. If you told a customer support rep that you are the former US president [0], they would not hand over the account straight away because you asked nicely.

These models are great tools, but putting them and people on the same level does a disservice to our species and also is simply incorrect to what we know these models to be and their capabilities/limitations.

[0] https://www.theguardian.com/technology/2026/jun/01/meta-ai-h...

reply
I didn't put them on the same level.

At the same time, one should acknowledge that not all tasks are on the same level.

reply
Most tasks we use computers for are deterministic and was coded for that specific quality. Introducing nom deterministic behavior is lowering the value of the app, especially for power users.
reply
[dead]
reply
I'm struggling with the assertion that these models cannot provide reasonably deterministic guarantees.

I am using gpt to populate JSON objects conforming to a list of natural language constraints for purposes of generating fake customers. I am finding that gpt5+ never fucks up. Not even a little bit. I've ran this test hundreds of times with 20+ constraints and it's been perfect every time.

Stable information yields stable control flow. Humans are much more likely to forget one of the many constraints during testing. This happy mistake may incidentally cover an edge but it also means we lose coverage elsewhere.

I think whether or not the LLM should be allowed to directly author deterministic control flow (code) is mostly the same thing. If you have a lot of constraints you want to satisfy all at the same time, this can give you a hit very close to the ideal target very quickly. Not knowing exactly what you want is when the LLM takes you for a ride.

reply
We could probably debate this ad nauseam, so I'll just give you my most compelling arguments.

1) Writing code the "old fashioned" way (i.e., a Python program that does X, Y, Z) allows you to arrive at a battle tested solution that will not change over time. From a risk assessment perspective, the behavior is essentially immutable, allowing a business to guarantee consistent behavior over long periods of time.

2) Just because something hasn't happened to you do, does not mean that it will not happen. LLM are opaque. If you stay on the "happy path", you may see consistent behavior for long periods of time, but there's always potential for an edge case where something goes catastrophically wrong. This is without even opening the can of worms regarding prompt injection and intentional sabotage of a working system.

3) There are plenty of real world examples of an LLM spontaneously deleting data from a DB (or the entire DB) or otherwise going completely off the rails. These might seem hyperbolic, but it happened at our company (to a test DB, not production). The severity of errors that occur can be existential to a business' survival without the proper guard rails.

4) There's no concrete way to truly confirm understanding between an LLM and a human. It can tell you that it completely understands what you want, and then it can do exactly the opposite. Followed by, "my bad" (Claude's new favorite catch phrase). Code can be audited and even proven to be correct given the appropriate level of time and energy.

My best results have been gleaned in using LLM to produce deterministic systems. I recognize everyone has different use cases and needs, but this seems to be the best use of the technology in my experience.

reply
> I'm struggling with the assertion that these models cannot provide reasonably deterministic guarantees.

LLMs are probabilistic by design so running the same prompt multiple times will give you different results.

Otherwise, we wouldn’t needed LLMs and could replace it all with Postgres

reply
> LLMs are probabilistic by design so running the same prompt multiple times will give you different results.

Reasonably deterministic is the phrase. If I can be sure the LLM is giving me back the same result 99% of the time I need it, that's reasonable for me. Maybe this is not reasonable for others.

ie How often will an LLM get 2+2 wrong? Now expand until you're uncomfortable.

reply
People really need to read Dijkstras Go to statement considered harmful letter [1]. If the obscurity of go to for static analysis of the code was too much, of course bringing in a literal ai black box is harmful for stable processes.

[1] https://homepages.cwi.nl/~storm/teaching/reader/Dijkstra68.p...

reply
I can argue that by applying multiple stochastic processes, with a human in the loop, that you will (may) converge on something that is deterministic. You use tests/test vectors to prove this.

We're no different to AI. The code we write to solve a particular problem can (and probably does) change from day to day, depending on your "mood", what you had for breakfast, if you've been fighting with your significant other, other problems/human emotions.

reply
the humans only there to take the blame. you arn't not goingto be a cyborg, nor is your idea of determinism ever going to last the cycle of the agents, no matter how many memorrry layers, skills and other context guardrails you place.

You're there for blame, not much else. The systems are still going to churn garbage, but because it's a business, that business will rather pay less for garbage they can sell than pay you a living wage, eventuially.

reply
Yeah because "works many times in a row" = "deterministic" to many people.
reply
"Business processes" can also mean "building power point decks" and other things.

But your point stands: for critical business processes that need predictability, we indeed need determinism.

reply