Entirely Foreseeable AWS Outageshttps://rys.io/en/182.html
-
@hans I agree people are precious and unique - and also super non-deterministic.
I think AI can be most effectively used to stimulate new thinking, and make bold forays into new-to-oneself domains of knowledge. The LLMs can reveal lots of new terminology, and possibilities one didn't think to search for before. But then one has to take one's time, verifying the claims (often wrong), and synthesize a truth for oneself, going beyond the hallucination of the LLMs. So I think of LLMs/agents more like *discovery* tools, at best, but much less something to build solid foundations with.
@Owl Eyes That's my experience so far, indeed. But the new knowledge that my tool has brought me, is more often than not wrong.
It has no problem telling me to configure settings that don't exist, even after I tell it they don't exist.
So far, it has been an interesting journey, but hasn't saved me any time. In fact, I have spent a lot more time, because I had to explain to my agent what I wanted, and go on a wild goose chase afterwards to check its solutions.
But I kind of expect that in a few years, these tools will become good enough to actually help. -
> On the other hand, people aren't deterministic either, and they are the most valuable tools
Whoa, okay, maybe let's start by not calling people "tools".
You have a massively complex system like AWS infrastructure. You have engineers who are not "deterministic" in the sense that software is deterministic, managing it.
Why on Earth would you want to complicate your life and take on loads of risk by adding another layer of random non-determinism in there? Makes no sense.
@Michał "rysiek" Woźniak ·
I'm a sysadmin myself, so I can call them tools 
But agree: at the moment these agentic tools aren't good enough to be trusted with massive, complex tasks. But I would be surprised if that would remain the situation for long. -
@Michał "rysiek" Woźniak · I'm a sysadmin myself, so I can call them tools
But agree: at the moment these agentic tools aren't good enough to be trusted with massive, complex tasks. But I would be surprised if that would remain the situation for long.@hans I would be surprised if it ever meaningfully changes.
-
Entirely Foreseeable AWS Outages
https://rys.io/en/182.htmlOnce you strip away the marketing hype, agentic systems like Kiro AI are just automation tools.
The difference between Kiro and regular infrastructure management tools is that the latter are deterministic. They can be tested, analyzed, and bugs can be reliably, provably fixed.
That's just not the case with agentic tools. They are by their very nature non-deterministic. And that's the last thing a systems engineer should want.
Pfft, please. Engineers. They're so unreasonable.
What are engineers looking for? Precision?
-
Entirely Foreseeable AWS Outages
https://rys.io/en/182.htmlOnce you strip away the marketing hype, agentic systems like Kiro AI are just automation tools.
The difference between Kiro and regular infrastructure management tools is that the latter are deterministic. They can be tested, analyzed, and bugs can be reliably, provably fixed.
That's just not the case with agentic tools. They are by their very nature non-deterministic. And that's the last thing a systems engineer should want.
@rysiek It’s not that they are non-deterministic - they actually aren’t. Same input will generate the same output as long as you configure it not to perform random sampling or bind the random number generator to a stable input.
The problem with these tools is that they are unpredictable. You cannot reason about their output beforehand. Nor can you reason about the effect changes to inputs are gonna have on the outputs.
That’s not non-determinism, that’s chaos.
-
@rysiek It’s not that they are non-deterministic - they actually aren’t. Same input will generate the same output as long as you configure it not to perform random sampling or bind the random number generator to a stable input.
The problem with these tools is that they are unpredictable. You cannot reason about their output beforehand. Nor can you reason about the effect changes to inputs are gonna have on the outputs.
That’s not non-determinism, that’s chaos.
@slotos if we want to be nit picky, sure why not – these models use random seeds while generating their output.
So while *technically* you are correct (the best kind of correct!) that if all inputs are exactly the same, the outputs will be the same as well, from the perspective of these systems as they are being used bye people using them, they are non-deterministic, because these users have no control over the random seed.
-
R relay@relay.an.exchange shared this topic
-
@slotos if we want to be nit picky, sure why not – these models use random seeds while generating their output.
So while *technically* you are correct (the best kind of correct!) that if all inputs are exactly the same, the outputs will be the same as well, from the perspective of these systems as they are being used bye people using them, they are non-deterministic, because these users have no control over the random seed.
@rysiek I’m not saying this to sound correct, I’m saying this to point out a deep design issue with these tools that gets ignored in the public discourse.
Non-determinism is not where the actual issue lies. If it was, tech bros advocating for adoption of local LLMs would have a leg to stand on.
There are useful [pseudo-]non-deterministic tools in IT. I cannot name a single useful chaotic one.
-
@rysiek I’m not saying this to sound correct, I’m saying this to point out a deep design issue with these tools that gets ignored in the public discourse.
Non-determinism is not where the actual issue lies. If it was, tech bros advocating for adoption of local LLMs would have a leg to stand on.
There are useful [pseudo-]non-deterministic tools in IT. I cannot name a single useful chaotic one.
@slotos well, Chaos Monkey:
https://netflix.github.io/chaosmonkey/ -
Entirely Foreseeable AWS Outages
https://rys.io/en/182.htmlOnce you strip away the marketing hype, agentic systems like Kiro AI are just automation tools.
The difference between Kiro and regular infrastructure management tools is that the latter are deterministic. They can be tested, analyzed, and bugs can be reliably, provably fixed.
That's just not the case with agentic tools. They are by their very nature non-deterministic. And that's the last thing a systems engineer should want.
@rysiek and the non-determinism is a feature, not a bug. LLMs are just a set of transformations performed on an input and the same input should/would result in the same output but they intentionally add a randomness factor to the input so the output seems more "natural" and therefore also more error-prone and inscrutable.
anyone using LLMs in situ for performing tasks, especially automation tasks, are playing russian roulette in a literal sense.
-
@slotos well, Chaos Monkey:
https://netflix.github.io/chaosmonkey/@rysiek And how is that chaotic? In a mathematical sense, please.
-
@rysiek And how is that chaotic? In a mathematical sense, please.
@slotos wow, I had no clue I'm taking an exam.
-
R relay@relay.infosec.exchange shared this topic
