[ai_fry_ur_brain]

I hate how seriously people take the output of an LLMs or how reliable they think it is.

Have Claude produce that spec 10 times, use the same prompt and same context. Identical requests, but you'll get 10 unique answers that wil contradict each other with each response seeming extermely confident.

Its scary how confident you people are in these outputs.

[CrazyStat]

If you ask 10 different humans to produce the spec with the same information (prompt and context) they will also produce 10 unique answers that will contradict each other and (depending on who you asked) may be just as confident.

There are real decisions to be made when going from a vague prompt to a spec. It's not surprising that an LLM would produce different specs for the same work on different runs. If the prompt already contained answers to all the decision points that come up when writing the spec then the prompt would already be the spec itself.

[b40d-48b2-979e]

LLMs aren't people. They don't reason. They're token generators, a black box. Your analogy falls on its face with any scrutiny.

[CrazyStat]

I didn’t claim that LLMs are people or that they reason.

If the behavior of the llm is the same as the behavior of reasonable people then the behavior of the llm is reasonable, regardless of how black of a box they generate tokens out of.

Reasonable people will generate divergent specs for the same prompt. Thus it is reasonable for an LLM to generate divergent specs out of the same prompt.

Edit: I use “reasonable” here in the legal sense of the “reasonable person” standard, not to imply any reasoning process.

[digitaltrees]

Aren’t people pattern matching neural networks as well? Why does being a token generator mean something is unreliable?

Further, why does that mean “it doesn’t reason”. Logic can be encoded in language, symbols or code. If I say “all apples are red” -> “all fruit in the bowl are apples” -> “therefor all the fruit are red”. It doesn’t really matter if I understand the logic or what red is or fruit/apples are, the logic is contained in the structure of the syntax. If an LLM can output the conclusion reliably from predictive operations it is able to have the effect of reason and we don’t need to know or care about whether it “understands” the reasoning.

[keeganpoppen]

no, brah, humans are TOTALLY different. just don’t think about it too hard. we are just special.

[jatora]

it's an analogy, it didnt fall on its face at all. it's just a comparison to highlight the point being made was nonsensical. example: you're just a next action generator controlled by trillions of cells and subconscious dna-based behavior. a black box.

[svieira]

> you're just a next action generator controlled by trillions of cells and subconscious dna-based behavior.

With moral agency and the ability to learn (even if we presume you are correct, which I don't think you are).

[jatora]

moral agency and the ability to learn are implicit in the description you quoted. this isn't some special superpower, all animals have the ability to learn, and many have moral agency. these aren't human specific traits

[b40d-48b2-979e]

Reductio ad absurdum.

[jatora]

exactly my point lol

[NobleLie]

It appears they don't need to reason or be intelligent to be able to produce working solutions for code. But sure let wild and unmonitored? I wrangle my LLMs like the code monkeys they are. They help materialize code and then you need to sculpt it (and test harness of varying sorts)

It really can be useful. It's very different from old world programming.

[keeganpoppen]

why do people insist on claiming that they don’t reason, when they clearly, for all intents and purposes, do. you can be vague; you can express your idea a thousand different ways, and you will get a unique blend of <your input bits> x <hidden reasoning layer> => semi-smoothed output. this is like some Searle Chinese Room bullshit that needs to just die. it is beyond clear that llms can interact with abstract concepts in an extremely meaningful way. this is like the “thought leader” version of the stupid-ass “it’s just smart autocomplete” argument. if you think that, it is user error— either a failure of creativity or a failure of perception or both. just because llms are not a panacea and are problematic for society and “overhyped” and whatever does not make it intellectually honest to claim that there is zero reasoning/creativity/cognition within the box.

[dnautics]

LLMs do reason (they just sometimes don't reason well).

I assure you I've met many devs and "engineers" that reason less than LLMs, and are black boxes, especially in terms of the code they write.

[claytongulick]

> LLMs do reason

No, they don't.

They are token predictors that use statistical techniques to emit the randomly weighted next most likely token given the previous token list.

The result is a strange mimic of human reasoning, because the tokens it predicts are trained on strings that were produced by humans that were reasoning, but that's not the same thing.

Human cognition is complex and poorly understood, and the nature of the mind is an area of study almost as old as consciousness itself. We don't know exactly how it works, or what its exact relationship to the brain is, but we do know that it is not a simple token predictor.

LLMs, by their very nature are constrained to the concept of language and the relationship between existing words in a corpus. This is a box they can not escape.

Modern neuroscience suggests that the human brain is much more vast than that, and in many ways looks like it is constrained by language, but certainly not limited to it.

[antonvs]

> They are token predictors that use statistical techniques to emit the randomly weighted next most likely token given the previous token list.

Sounds like an implementation detail. Now describe how human reasoning works and explain why that process of chemical and electrical signals results in "reasoning" whereas what LLMs do isn't.

The problem with being this reductive is you can do it to anything, including humans. You can’t be reductive about LLMs and refuse to be reductive about humans - that's poor reasoning, and an LLM would out-reason you on this point, further negating your case.

[claytongulick]

Human cognition is poorly understood and much more complex than it seems.

For an example, look at some of Julia Mossbridge's work.

If even a small part of her work is true and valid, it points to something far outside our current framework.

You don't need to go as far afield as Mossbridge, though - that's an extreme example. Pretty much any modern neuroscience will make you question a lot of assumptions, at least it did for me.

[dnautics]

You have moved goalposts from reasoning to "human cognition". I won't tolerate that sort of slippery wordplay.

Reasoning is making analogies between logical patterns found in conceptual space, with a direction of time (statements precede conclusions). For example. A => B and B => C. You may now deduce A => C. For something fuzzier, A~D and B~E, you may now deduce that D~=>E. This is the sort of thing that higher layer attention mechanism is capable of doing.

> This is a box they can not escape.

Would you say that Helen Keller was less capable of abstract reasoning because she had more constrained access to sensory input?

[claytongulick]

Reasoning requires cognition, otherwise there's nothing to reason about, no context or value system to use as a basis for reason.

Decision making can be done by trained machines following rules, but that's different that reasoning. A thermostat isn't reasoning when it decides to turn on the air conditioner, to argue otherwise expands the definition of "reason" to be so broad that it becomes useless.

LLMs are trained on human knowledge and reasoning that results from human cognition, and they are excellent at stochastic mimicry - if the argument is that they are actually reasoning, then some sort of equivalent to human cognition must be present for that to be true. Lacking that, they are nothing more than "token extrusion machines" with some potentially useful characteristics.

[NobleLie]

The problem with that is LLMs can output words or symbols that seen like it used "reason" to produce. But for everything the core algorithm does, it's simply nothing like the wetware reasoning to get to the same answer. So he didn't move goalposts. He always meant the reasoning that stems from human cognition.

Technically if it has that, it'd be singularity no? So basically the premise is they are doing nothing of the sort. Prove any LLM enough and it really does show it has no quarrels contradicting itself or being bossed around. Has no belief / no orientation etc. It's truly mindless but tricks our mind and soul (or whatever) probably.

[digitaltrees]

The structure of language encodes logic in many ways. So the models ability to reason may be an emergent property of the reasoning ability humanity has ejected an extracted from our neural networks and abstracted into language a symbols.

[keeganpoppen]

there is absolutely no line of demarcation between human reasoning and what you described

[IshKebab]

Wow, there are still people trying to claim they don't reason. What will they have to do before you'll admit that they can?

[esailija]

You are asking the wrong question. It's not about if you can do X which can be faked especially if you are given practically infinite tries and all failures are hidden.

The people who want to believe they actually reason just ignore all obvious evidence of contrary and cherry pick the times reasoning was faked well enough.

The people who don't want to believe will just take a second to understand how they work and then come up with ways to reveal they were faking all along. Like asking how many letters there are in a word lol.

It's only the people who don't want to believe that count because reality is what happens despite of what you believe.

[IshKebab]

You seem to believe that something is only "reasoning" if it works in a particular way. That it's not enough for it to observationally display reasoning skills; it has to be using a particular method to do that so it's not "faking" it. Is that correct?

[Jtarii]

It will be interesting to see the excuses people come up with when LLMs innevitably start solving millenium prize problems.

[Jtarii]

They very obviously reason.

[dnautics]

it's kind of crazy to think that the transformer architecture can't encode some primitive form of reasoning.

[johnnyanmac]

The issue is Lllms don't learn, despite the name. A human re-implementing a spec would strive to iterate towards what they feel is a better spec. They can take in their own input and self-correct. The work of implementing the spec gives insight into pain points and strengths, even if they never actually test the spec (they 100% should, but this is to emphasize that struggle for humans is in itself iteration, even before external feedback comes in).

An LLM is isn't deterministic but also isn't iterative without an existing human. You give it the same spec 10 times and it produces 10 results that aren't far off itself but vastly different when you go into the weeds. And not different in a way of improvement. |

[olafmol]

An LLM should not "generate specs", a human should. The LLM can work from the specs. It can never infer meaning from a vague prompt. If so, it will start guessing. Every human that ever did functional specification or information analysis at some point knows this. Or has learned the hard way, something with assumptions and asses ;)

[dist-epoch]

The guessing of a LLM for a vague prompt is better than the one of your average developer.

A prompt like "write these two files on disk" will very likely make the LLM do some sort of an atomic write/swap operation, unlike the average developer which will just write the two files and maybe later encounter a race condition bug. You can argue the LLM output is overkill, but it will also be more robust on average.

[rixed]

What kind of race condition do you have in mind?

[skydhash]

So what’s most important is knowing those parameters and the ranges of values, not having the final result. A human, after producing a specs, can the provide the mental model of how he created the specs. Where the inflection points are and what the range of valid results.

What has always mattered is how you decide the specs, not the specs in themselves.

[claytongulick]

> If you ask 10 different humans to produce the spec with the same information (prompt and context) they will also produce 10 unique answers

But they didn't ask humans, they asked a machine. We expect our machines to behave in predictable ways.

> If the prompt already contained answers to all the decision points that come up when writing the spec then the prompt would already be the spec itself.

This is one of the best arguments against using LLMs I've seen.

It reduces to the classic argument- at the point where you've described a problem and solution in sufficient detail to be confident in the results, you've invented a programming language.

[CrazyStat]

> We expect our machines to behave in predictable ways.

I expect LLMs to produce randomly varying output. Maybe it's the thousands of hours I spent doing monte carlo simulations for my PhD.

> This is one of the best arguments against using LLMs I've seen.

> It reduces to the classic argument- at the point where you've described a problem and solution in sufficient detail to be confident in the results, you've invented a programming language.

I'm not an LLM true believer, but I use codex for various small tasks and it often (not always) does a thoroughly decent job. Yesterday I gave it a pretty vague request to set up a new Home Assistant dashboard and it handled it just fine--I told it what I wanted to see but it figured out itself which helper variables it would need to set up to realize that vision and wrote all the config for it.

I probably could have done it in 15 minutes if I was familiar with Home Assistant's yaml configuration schema and all, but I'm not so it probably would have taken me closer to an hour. Asking codex took me 30 seconds and it did just fine.

I am skeptical that LLM's are going to kill all white collar jobs or whatever anytime soon. Not being able to truly learn things is an issue. Reality has a surprising amount of detail[1], and while codex does well at things like writing Home Assistant configs and setting up a Minecraft server, where there are thousands of examples online of how to do it, when I've asked it to do some more esoteric things it has sometimes failed spectacularly. I don't think having the LLM keep notes and then read them back (filling up the context window) is a real solution here.

[1] http://johnsalvatier.org/blog/2017/reality-has-a-surprising-...

[claytongulick]

I haven't made the argument that LLMs aren't useful, I can see cases where they are.

I don't think they include areas where correctness, determinism or human reasoning are important.

At least, not in isolation.

[dxxvi]

> It's not surprising that an LLM would produce different specs for the same work on different runs This is what I don't understand: AI is a computer program with its own data. If we give the same input to that computer program every time, why does it produce different outputs every time? Or does the input include LLM data + our prompt + some random data that computer program picks from its Internet search?

[CrazyStat]

LLMs have a temperature parameter. At zero temperature they are deterministic: they always choose the most likely next token at each step based on what came before and the model weights, and they will always generate the same output given the same input.

As you raise the temperature they will start (pseudo)randomly choosing tokens other than the single most likely token (though that one will still be the most likely to be chosen). It turns out this is almost always better than zero temperature, which has a tendency to get caught in repetitive loops. I imagine all the frontier labs have spent thousands (millions?) of CPU hours tuning the temperature parameters on their models for optimal performance.

[thesz]

  > LLMs have a temperature parameter. At zero temperature they are deterministic: they always choose the most likely next token at each step based on what came before and the model weights, and they will always generate the same output given the same input.

https://en.wikipedia.org/wiki/Softmax_function

"A value proportional to the reciprocal of β is sometimes referred to as the temperature: β = 1/kT, where k is typically 1 or the Boltzmann constant and T is the temperature. A higher temperature results in a more uniform output distribution (i.e. with higher entropy; it is "more random"), while a lower temperature results in a sharper output distribution, with one value dominating."

"Temperature" in the context of softmax does not change a "winning" token, it changes how much probable (in the sense of softmax distribution) winning token will be. If the winning token is "New York", it will be a winner with temperature close to 0 and with temperature of 1e9.

The actual selection of the random token is done separately by using inputs outside of the softmax distribution, for example, by using random number generator. I believe most of LLM configs have a seed for the random number generator.

More than that, generation of code in most programming languages is done with the more guardrails such as beam search guided by schema, syntax and semantics.

[NobleLie]

Nah. Even with zero temperature this is still variation.

[digitaltrees]

But those differences fall within a band of generally accepted results don’t they? And the cost to throw the code away and reimplement is low now. So maybe it doesn’t really matter if the implementation is perfect or identical.

That being said I agree people trust AI too much. Especially people with less experience. It’s easy to forget the models are mirrors of we are as the drivers of the input context not mentors that will guide us to best practices reliably.

[Robdel12]

Imagine making this your entire identity