Comment by gortok
10 hours ago
When I use a computer to do work I want the computer to be right. I want to be able to trust the computer. With the inherit non-determinism and probabilistic nature of generative-AI, that fundamental reason why I engage a computer is lost.
If the spreadsheet is wrong, it’s because the math is wrong, it’s because I made a mistake. It’s not because all of a sudden the computer decided the nature of algebra should be different than it is.
Part of the reason why humans are rejecting AI is that we are putting it in places where it makes no sense, or places where humans prefer a human in the loop, there are plenty of places where machine learning algorithms make sense, but customer service is not one of them.
Heavily agree.
The computer should be a force for order, because being a living creature is chaos.
That said LLMs can be used in ways that promote order. People just got excited and wanted to believe they could be trusted in chaos mode.
For reference chaos mode would be prompting something like: "Look at my journal entries and tell me what I should do to fix my life". Versus using one to build a table of common themes and analyzing the resulting spreadsheet yourself.
I think we got excited / wanted to believe that we won't have to expend any effort whatsoever, and the AI could "do it all".
The reality (as far as I can comprehend it) seems to be that AI expanded the scope of what we can shove into one mouthful, and now it takes much more effort to chew. Metaphorically speaking.
Part of the reason coding agents are widely accepted and used is that human coders are chaos, and human coders have spent a lot of time building tooling to ameliorate chaos. Everything from git to language design, to lint, to profilers, etc. was built to keep the human chaos out. It's pretty good at keeping the LLM chaos out, and when it blows up anyway you can roll it back to the previous commit.
This is an interesting way of thinking about it. I generally agree. I especially agree that anti-AI sentiment partially comes from miss-using it. However:
Determinism isn't a requirement for 100% correctness.
A Las Vegas algorithm is randomized, non-deterministic and guarantees 100% correctness [0].
The execution can be different every time but the result will always be correct. determinism does not lose accuracy. It does lose time predictability.
So if your problem with AI is accuracy, then in theory your problem is just premature stopping.
[0] https://en.wikipedia.org/wiki/Las_Vegas_algorithm
A Las Vegas algorithm requires that you have a deterministic test that can definitively determine the correctness of an intermediate result. So what you're saying is that what it takes to make LLMs give 100% correct results is having a human between the LLM and the user, who's capable of re-prompting on incorrect answers from the LLM. Well, if the human is there, why not just ask the human? What value is the random number generator adding?
Like the GP said, the point of determinism is that you can trust the correctness of the results, without doing any checking. Solved problems stay solved.
> So what you're saying is that what it takes to make LLMs give 100% correct results is having a human between the LLM and the user, who's capable of re-prompting on incorrect answers from the LLM.
No, that's not what I'm saying. Im saying determinism isn't required for correctness. And Im not saying that models are already perfect other than prematurely stopping. What I am saying is that non-determinism doesn't mean they cant be 100% correct.
Besides, humans baby-sitting LLMs is not even implied by the misinterpretation of what I'm saying. What's implied by that is humans needing to give the LLM explicit success criteria from the outset. Which is totally reasonable.
> Like the GP said, the point of determinism is that you can trust the correctness of the results
Well, no. Now you are waaaay overstating determinism. Deterministic results might be incorrect.
1 reply →
The economics of it (token cost) means, however, that what will be chosen most often is the barely sustainable minimum level of quality, aka race to the bottom. AI is more cost-sensitive in that way than humans caring or not caring about making things robust and correct used to be.
For sure, that's why I say in theory. Also models aren't perfect. You can't plug in an explicit target and ensure it doesnt stop prematurely and expect 100 accuracy. But that's not because it's non-deterministic, since clearly non-deterministic systems can be 100% accurate.
I'm not sure I fully agree with this. We can use the computer to build a deterministic system.
People aren't fully deterministic either.
Correct.
Also, the better one gets at using AI, the better they can predict where AI might fail.
This allows you to both work 10x faster and prevent many mistakes, which puts you far ahead of not-using-AI.
Correctness can also be guaranteed by non-deterministic systems [0]. You do trade time predictability though. It will eventually give the correct answer, we just dont know when.
I keep thinking about the implications of this. So in some sense it's less about being inaccurate and more about prematurely stopping (or not having a well defined target, but that's a whole other mess).
In theory, if the target is well defined and it never prematurely stops, the question changes from "will the output be correct" to "when will it be done?"
[0] https://en.wikipedia.org/wiki/Las_Vegas_algorithm
> we just dont know when
Ah, so it's merely functionally useless, if not theoretically.
1 reply →
Yep. Exactly why we use machines. They do the deterministic bit quickly. We do the other bits.
> there are plenty of places where machine learning algorithms make sense, but customer service is not one of them.
I don't care about human vs AI, I just want my issue resolved. Whatever does that the best and fastest. Or even better, for there to not be an issue in the first place.
The root of all our evils, right here. We only want the fastest cheapest thing possible, without thinking about how it impacts literally everything else about our world.
They never said fastest or cheapest, they said it should resolve the issue. A fast cheap human who can't is not better than an AI that can.
Reality is non-deterministic (not actually, but in practice). We do strange things to balance this out all the time. Thinking that computer software must be exempt from that mess as a goal is just strange, and concerns itself too much with the ideal of a tool and not enough with the more important question: What if our lives get much, much better? Roughly everyone wants more out of life including the top one percent (and I don't mean the top 0.0001%, just your ordinary industrial nation doctors). How could we possibly morally justify denying everyone the best shot to get at least to that level, and I mean living people RIGHT NOW, ASAP not in a few hundred years? Who would be okay to say "yeah but you know, artists, copyright" and deny those who benefit most from this the opportunity?
How could we not take this shot? I understand perfectly well that a lot of things will need reconfiguration and that it's going to be painful, but dear lord, let's focus on making it go well instead of ending it.
> Thinking that computer software must be exempt from that mess as a goal is just strange
Software is deterministic, it has been since its inception[0]. Why go from something objective/provable to something that does "strange things", when we already have former? It's like making bricks out of paper and declaring "actually, this is logical next step for bricks because stuff waves in the wind".
> What if our lives get much, much better?
What if not?
[0] (Yes not really/actually if we're being pedantic)
> Why go from something objective/provable to something that does "strange things", when we already have former?
Because we like computers to feel like they are fast, mainly. Most compilers, for example, are non-deterministic because they can be made to run faster if they ignore things like thread execution order. Same goes for LLMs. Technically they are as deterministic as any other software, but we allow GPUs to play fast and loose with floating point numbers to speed things up, which gives the impression of "strange things".
Here is a simple and possibly clarifying thought experiment: Would you be willing to switch your standard of living with the standard of the median person living on this planet? (And I would urge you to look up what that looks like, in case you are unsure).
As long as you are not (and I sympathize) I have zero clue how to justify any delay in getting everyone at least to our current level.
I understand that there are risks and we should work hard to guard against them. But no society has seriously considered giving up driving while we figure out global warming. People want a good life, that's just the selfish fucks we are, and it's upon those with clout to will it for everyone.
> Why go from something objective/provable to something that does "strange things", when we already have former?
Because it does something different. It's not from/to. LLMs are subbing in for humans, not for deterministic computing. Replacing deterministic computing with LLMs for tasks that have be perfectly solved without LLMs would be wasteful and silly.
This is completely incoherent. Of course we need deterministic computer programs; much of society depends on it.
All software written will not disappear. There is nothing keeping us from using partly undeterministic software (see: humans) to write deterministic programs.
What does the N in NP stand for?
I'm a bit joking, but we've been working in deterministic computation for so long, we don't even think of there being another way.
But seriously, I do view AI as the input to a deterministic machine. Junior engineers (well all engineers) aren't deterministic, and we've made processes to direct their behavior towards making better software. AI agents do a better job of following my processes than engineers. We move up the stack towards testing and verification rather than writing. That doesn't make me sad, after 40 years of coding, I'm kind of tired of it. I have more ideas than I can code, so I'm happy to give AI my ideas and have it code for me.
I had a former manager tell me that all technology problems are really people problems, now maybe all technology problems are all agent problems and we just have to get comfortable with managing agents like we got comfortable with managing people.
> How could we possibly morally justify denying everyone the best shot to get at least to that level, and I mean living people RIGHT NOW, ASAP not in a few hundred years?
How exactly do you think this is going to happen?
By trying! How else is it going to happen? Are we going to deny the immense potential? Nobody needs to draw up a specific plan for that to hold true, we are good at figuring shit out and using tools.
I am not saying it's going to work but we are not getting much smarter right now, and we really need all the help we can get to accelerate more complicated stuff.
And it is accelerating! Will it be as useful as I hope it will be? That is entirely beside the point. This post was not at all about me assigning any chance of the good outcome. Just that there is no other ethical option.
1 reply →
> Reality is non-deterministic (not actually, but in practice).
At a quantum level, it is also actually non-deterministic.
That's a common misconception.
>What if our lives get much, much better? Roughly everyone wants more out of life including the top one percent (and I don't mean the top 0.0001%, just your ordinary industrial nation doctors). How could we possibly morally justify denying everyone the best shot to get at least to that level, and I mean living people RIGHT NOW, ASAP not in a few hundred years?
This is satire, right?
If you seriously think that AI is going to improve the lives of anyone except the robber barons who own the AI you are absolutely delusional.
Precisely! I don't think this is the whole problem but it's a part of it I keep coming back to. We took the one thing that computers were good at (getting the same answer every time, quickly and efficiently) and tossed it out the window.
Good AI system design can help somewhat, but even if you give the LLM a calculator tool, it's not guaranteed to use it every time, or to write the tool use correctly, or to copy out the answer correctly.
> there are plenty of places where machine learning algorithms make sense, but customer service is not one of them.
You're optimising for quality, where as companies optimise for some balance of quality and cost.
AI might not be quite as good as a skilled human, but it's often good enough and a lot cheaper, so companies use it.
I actually think customer service is one of the few places it makes sense to use AI – at least to some degree. AI can provide immediate support to customer queries, and can usually handle the majority of basic issues customers have. You might need to escalate to a human in edge cases, but that's how you balance quality and cost.
working with a non deterministic tool requires taste and judgement. if you want full determinism AI is not for you but it has a market for other people?