Comment by grey-area
6 months ago
On the contrary, anthropomorphism IMO is the main problem with narratives around LLMs - people are genuinely talking about them thinking and reasoning when they are doing nothing of that sort (actively encouraged by the companies selling them) and it is completely distorting discussions on their use and perceptions of their utility.
I kinda agree with both of you. It might be a required abstraction, but it's a leaky one.
Long before LLMs, I would talk about classes / functions / modules like "it then does this, decides the epsilon is too low, chops it up and adds it to the list".
The difference I guess it was only to a technical crowd and nobody would mistake this for anything it wasn't. Everybody know that "it" didn't "decide" anything.
With AI being so mainstream and the math being much more elusive than a simple if..then I guess it's just too easy to take this simple speaking convention at face value.
EDIT: some clarifications / wording
Agreeing with you, this is a "can a submarine swim" problem IMO. We need a new word for what LLMs are doing. Calling it "thinking" is stretching the word to breaking point, but "selecting the next word based on a complex statistical model" doesn't begin to capture what they're capable of.
Maybe it's cog-nition (emphasis on the cog).
What does a submarine do? Submarine? I suppose you "drive" a submarine which is getting to the idea: submarines don't swim because ultimately they are "driven"? I guess the issue is we don't make up a new word for what submarines do, we just don't use human words.
I think the above poster gets a little distracted by suggesting the models are creative which itself is disputed. Perhaps a better term, like above, would be to just use "model". They are models after all. We don't make up a new portmanteau for submarines. They float, or drive, or submarine around.
So maybe an LLM doesn't "write" a poem, but instead "models a poem" which maybe indeed take away a little of the sketchy magic and fake humanness they tend to be imbued with.
16 replies →
> this is a "can a submarine swim" problem IMO. We need a new word for what LLMs are doing.
Why?
A plane is not a fly and does not stay aloft like a fly, yet we describe what it does as flying despite the fact that it does not flap its wings. What are the downsides we encounter that are caused by using the word “fly” to describe a plane travelling through the air?
4 replies →
"predirence" -> prediction meets inference and it sounds a bit like preference
2 replies →
A machine that can imitate the products of thought is not the same as thinking.
All imitations require analogous mechanisms, but that is the extent of their similarities, in syntax. Thinking requires networks of billions of neurons, and then, not only that, but words can never exist on a plane because they do not belong to a plane. Words can only be stored on a plane, they are not useful on a plane.
Because of this LLMs have the potential to discover new aspects and implications of language that will be rarely useful to us because language is not useful within a computer, it is useful in the world.
Its like seeing loosely related patterns in a picture and keep derivating on those patterns that are real, but loosely related.
LLMs are not intelligence but its fine that we use that word to describe them.
It will help significantly, to realize that the only thinking happening is when the human looks at the output and attempts to verify if it is congruent with reality.
The rest of the time it’s generating content.
It's more like muscle memory than cognition. So maybe procedural memory but that isn't catchy.
1 reply →
This is a total non-problem that has been invented by people so they have something new and exciting to be pedantic about.
When we need to speak precisely about a model and how it works, we have a formal language (mathematics) which allows us to be absolutely specific. When we need to empirically observe how the model behaves, we have a completely precise method of doing this (running an eval).
Any other time, we use language in a purposefully intuitive and imprecise way, and that is a deliberate tradeoff which sacrifices precision for expressiveness.
It does some kind of automatic inference (AI), and that's it.
> "selecting the next word based on a complex statistical model" doesn't begin to capture what they're capable of.
I personally find that description perfect. If you want it shorter you could say that an LLM generates.
We can argue all day what "think" means and whether a LLM thinks (probably not IMO), but at least in my head the threshold for "decide" is much lower so I can perfectly accept that a LLM (or even a class) "decides". I don't have a conflict about that. Yeah, it might not be a decision in the human sense, but it's a decision in the mathematical sense so I have always meant "decide" literally when I was talking about a piece of code.
It's much more interesting when we are talking about... say... an ant... Does it "decide"? That I have no idea as it's probably somewhere in between, neither a sentient decision, nor a mathematical one.
Well, it outputs a chain of thoughts that later used to produce better prediction. It produces a chain of thoughts similar to how one would do thinking about a problem out loud. It's more verbose that what you would do, but you always have some ambient context that LLM lacks.
I mean you can boil anything down to it's building blocks and make it seem like it didn't 'decide' anything. When you as a human decide something, your brain and it's neurons just made some connections with an output signal sent to other parts that resulting in your body 'doing' something.
I don't think LLMs are sentient or any bullshit like that, but I do think people are too quick to write them off before really thinking about how a nn 'knows things' similar to how a human 'knows' things, it is trained and reacts to inputs and outputs. The body is just far more complex.
I wasn't talking about knowing (they clearly encode knowledge), I was talking about thinking/reasoning, which is something LLMs do not in fact do IMO.
These are very different and knowledge is not intelligence.
1 reply →
> EDIT: some clarifications / wording
This made me think, when will we see LLMs do the same; rereading what they just sent, and editing and correcting their output again :P
When I see these debates it's always the other way around - one person speaks colloquially about an LLM's behavior, and then somebody else jumps on them for supposedly believing the model is conscious, just because the speaker said "the model thinks.." or "the model knows.." or whatever.
To be honest the impression I've gotten is that some people are just very interested in talking about not anthropomorphizing AI, and less interested in talking about AI behaviors, so they see conversations about the latter as a chance to talk about the former.
As I write this, Claude Code is currently opening and closing various media files on my computer. Sometimes it plays the file for a few seconds before closing it, sometimes it starts playback and then seeks to a different position, sometimes it fast forwards or rewinds, etc.
I asked Claude to write a E-AC3 audio component so I can play videos with E-AC3 audio in the old version of QuickTime I really like using. Claude's decoder includes the ability to write debug output to a log file, so Claude is studying how QuickTime and the component interact, and it's controlling QuickTime via Applescript.
Sometimes QuickTime crashes, because this ancient API has its roots in the classic Mac OS days and is not exactly good. Claude reads the crash logs on its own—it knows where they are—and continues on its way. I'm just sitting back and trying to do other things while Claude works, although it's a little distracting that something else is using my computer at the same time.
I really don't want to anthropomorphize these programs, but it's just so hard when it's acting so much like a person...
Would it help you to know that trial and error is a common tactic by machines? Yes, humans do it too, but that doesn't mean the process isn't mechanical. In fact, in computing we might call this a "brute force" approach. You don't have to cover the entire search space to brute force something, and it certainly doesn't mean you can't have optimization strategies and need to grid search (e.g. you can use Bayesian methods, multi-armed bandit approaches, or a whole world of things).
I would call "fuck around and find out" a rather simple approach. It is why we use it! It is why lots of animals use it. Even very dumb animals use it. Though, we do notice more intelligent animals use more efficient optimization methods. All of this is technically hypothesis testing. Even a naive grid search. But that is still in the class of "fuck around and find out" or "brute force", right?
I should also mention two important things.
1) as a human we are biased to anthropomorphize. We see faces in clouds. We tell stories of mighty beings controlling the world in an effort to explain why things happen. This is anthropomorphization of the universe itself!
2) We design LLMs (and many other large ML systems) to optimize towards human preference. This reinforces an anthropomorphized interpretation.
The reason for doing this (2) is based on a naive assumption[0]: If it looks like a duck, swims like a duck, and quacks like a duck, then it *probably* is a duck. But the duck test doesn't rule out a highly sophisticated animatronic. It's a good rule of thumb, but wouldn't it also be incredibly naive to assume that it *is* a duck? Isn't the duck test itself entirely dependent on our own personal familiarity with ducks? I think this is important to remember and can help combat our own propensity for creating biases.
[0] It is not a bad strategy to build in that direction. When faced with many possible ways to go, this is a very reasonable approach. The naive part is if you assume that it will take you all the way to making a duck. It is also a perilous approach because you are explicitly making it harder for you to evaluate. It is, in the fullest sense of the phrase, "metric hacking."
8 replies →
Respectfully, that is a reflection of the places you hang out in (like HN) and not the reality of the population.
Outside the technical world it gets much worse. There are people who killed themselves because of LLMs, people who are in love with them, people who genuinely believe they have “awakened” their own private ChatGPT instance into AGI and are eschewing the real humans in their lives.
Naturally I'm aware of those things, but I don't think TFA or GGP were commenting on them so I wasn't either.
The other day a good friend of mine with mental health issues remarked that "his" chatgpt understands him better than most of his friends and gives him better advice than his therapist.
It's going to take a lot to get him out of that mindset and frankly I'm dreading trying to compare and contrast imperfect human behaviour and friendships with a sycophantic AI.
3 replies →
Most certainly the conversation is extremely political. There are not simply different points of view. There are competitive, gladiatorial opinions ready to ambush anyone not wearing the right colors. It's a situation where the technical conversation is drowning.
I suppose this war will be fought until people are out of energy, and if reason has no place, it is reasonable to let others tire themselves out reiterating statements that are not designed to bring anyone closer to the truth.
If this tech is going to be half as impactful as its proponents predict, then I'd say it's still under-politicized. Of course the politics around it doesn't have to be knee-jerk mudslinging, but it's no surprise that politics enters the picture when the tech can significantly transform society.
1 reply →
Wait until a conversation about “serverless” comes up and someone says there is no such thing because there are servers somewhere as if everyone - especially on HN -doesn’t already know that.
Why would everyone know that? Not everyone has experience in sysops, especially not beginners.
E.g. when I first started learning webdev, I didn’t think about ‘servers’. I just knew that if I uploaded my HTML/PHP files to my shared web host, then they appeared online.
It was only much later that I realized that shared webhosting is ‘just’ an abstraction over Linux/Apache (after all, I first had to learn about those topics).
11 replies →
Well "reasoning" refers to Chain-of-Thought and if you look at the generated prompts it's not hard to see why it's called that.
That said, it's fascinating to me that it works (and empirically, it does work; a reasoning model generating tens of thousands of tokens while working out the problem does produce better results). I wish I knew why. A priori I wouldn't have expected it, since there's no new input. That means it's all "in there" in the weights already. I don't see why it couldn't just one shot it without all the reasoning. And maybe the future will bring us more distilled models that can do that, or they can tease out all that reasoning with more generated training data, to move it from dispersed around the weights -> prompt -> more immediately accessible in the weights. But for now "reasoning" works.
But then, at the back of my mind is the easy answer: maybe you can't optimize it. Maybe the model has to "reason" to "organize its thoughts" and get the best results. After all, if you give me a complicated problem I'll write down hypotheses and outline approaches and double check results for consistency and all that. But now we're getting dangerously close to the "anthropomorphization" that this article is lamenting.
Using more tokens = more compute to use for a given problem. I think most of the benefit of CoT has more to do with autoregressive models being unable to “think ahead” and revise their output, and less to do with actual reasoning. The fact that an LLM can have incorrect reasoning in its CoT and still produce the right answer, or that it can “lie” in its CoT to avoid being detected as cheating on RL tasks, makes me believe that the semantic content of CoT is an illusion, and that the improved performance is from being able to explore and revise in some internal space using more compute before producing a final output.
> I don't see why it couldn't just one shot it without all the reasoning.
That's reminding me of deep neural networks where single layer networks could achieve the same results, but the layer would have to be excessively large. Maybe we're re-using the same kind of improvement, scaling in length instead of width because of our computation limitations ?
CoT gives the model more time to think and process the inputs it has. To give an extreme example, suppose you are using next token prediction to answer 'Is P==NP?' The tiny number of input tokens means that there's a tiny amount of compute to dedicate to producing an answer. A scratchpad allows us to break free of the short-inputs problem.
Meanwhile, things can happen in the latent representation which aren't reflected in the intermediate outputs. You could, instead of using CoT, say "Write a recipe for a vegetarian chile, along with a lengthy biographical story relating to the recipe. Afterwards, I will ask you again about my original question." And the latents can still help model the primary problem, yielding a better answer than you would have gotten with the short input alone.
Along these lines, I believe there are chain of thought studies which find that the content of the intermediate outputs don't actually matter all that much...
I like this mental-model, which rests heavily on the "be careful not to anthropomorphize" approach:
It was already common to use a document extender (LLM) against a hidden document, which resembles a movie or theater play where a character named User is interrogating a character named Bot.
Chain-of-thought switches the movie/script style to film noir, where the [Detective] Bot character has additional content which is not actually "spoken" at the User character. The extra words in the script add a certain kind of metaphorical inertia.
It looks superficially like reasoning, but is it reasoning?
"All models are wrong, but some models are useful," is the principle I have been using to decide when to go with an anthropomorphic explanation.
In other words, no, they never accurately describe what the LLM is actually doing. But sometimes drawing an analogy to human behavior is the most effective way to pump others' intuition about a particular LLM behavior. The trick is making sure that your audience understands that this is just an analogy, and that it has its limitations.
And it's not completely wrong. Mimicking human behavior is exactly what they're designed to do. You just need to keep reminding people that it's only doing so in a very superficial and spotty way. There's absolutely no basis for assuming that what's happening on the inside is the same.
Some models are useful in some contexts but wrong enough to be harmful in others.
All models are useful in some contexts but wrong enough to be harmful in others.
Relatedly, the alternative to pragmatism is analysis paralysis.
It's not just distorting discussions it's leading people to put a lot of faith in what LLMs are telling them. Was just on a zoom an hour ago where a guy working on a startup asked ChatGPT about his idea and then emailed us the result for discussion in the meeting. ChatGPT basically just told him what he wanted to hear - essentially that his idea was great and it would be successful ("if you implement it correctly" was doing a lot of work). It was a glowing endorsement of the idea that made the guy think that he must have a million dollar idea. I had to be "that guy" who said that maybe ChatGPT was telling him what he wanted to hear based on the way the question was formulated - tried to be very diplomatic about it and maybe I was a bit too diplomatic because it didn't shake his faith in what ChatGPT had told him.
LLMs directly exploit a human trust vuln. Our brains tend to engage with them relationally and create an unconscious functional belief that an agent on the other end is responding with their real thoughts, even when we know better.
AI apps ought to at minimum warn us that their responses are not anyone's (or anything's) real thoughts. But the illusion is so powerful that many people would ignore the warning.
> people are genuinely talking about them thinking and reasoning when they are doing nothing of that sort
Do you believe thinking/reasoning is a binary concept? If not, do you think the current top LLM are before or after the 50% mark? What % do you think they're at? What % range do you think humans exhibit?
> people are genuinely talking about them thinking and reasoning when they are doing nothing of that sort
With such strong wording, it should be rather easy to explain how our thinking differs from what LLMs do. The next step - showing that what LLMs do precludes any kind of sentience is probably much harder.
I think it's worth distinguishing between the use of anthropomorphism as a useful abstraction and the misuse by companies to fuel AI hype.
For example, I think "chain of thought" is a good name for what it denotes. It makes the concept easy to understand and discuss, and a non-antropomorphized name would be unnatural and unnecessarily complicate things. This doesn't mean that I support companies insisting that LLMs think just like humans or anything like that.
By the way, I would say actually anti-anthropomorphism has been a bigger problem for understanding LLMs than anthropomorphism itself. The main proponents of anti-anthropomorphism (e.g. Bender and the rest of "stochastic parrot" and related paper authors) came up with a lot of predictions about things that LLMs surely couldn't do (on account of just being predictors of the next word, etc.) which turned out to be spectacularly wrong.
I don't know about others, but I much prefer if some reductionist tries to conclude what's technically feasible and is proven wrong over time, than somebody yelling holistic analogies á la "it's sentient, it's intelligent, it thinks like us humans" for the sole dogmatic reason of being a futurist.
Tbh I also think your comparison that puts "UI events -> Bits -> Transistor Voltages" as analogy to "AI thinks -> token de-/encoding + MatMul" is certainly a stretch, as the part about "Bits -> Transistor Voltages" applies to both hierarchies as the foundational layer.
"chain of thought" could probably be called "progressive on-track-inference" and nobody would roll an eye.
I thought this too but then began to think about it from the perspective of the programmers trying to make it imitate human learning. That's what a nn is trying to do at the end of the day, and in the same way I train myself by reading problems and solutions, or learning vocab at a young age, it does so by tuning billions of parameters.
I think these models do learn similarly. What does it even mean to reason? Your brain knows certain things so it comes to certain conclusions, but it only knows those things because it was ''trained'' on those things.
I reason my car will crash if I go 120 mph on the other side of the road because previously I have 'seen' where the input is a car going 120mph has a high probability of producing a crash, and similarly have seen input where the car is going on the other side of the road, producing a crash. Combining the two would tell me it's a high probability.
how do you account for the success of reasoning models?
I agree these things don't think like we do, and that they have weird gaps, but to claim they can't reason at all doesn't feel grounded.
Serendipitous name...
In part I agree with the parent.
I agree that it is pointless to not anthropomorphize because we are humans and we will automatically do this. Willingly or unwillingly.
On the other hand, it generates bias. This bias can lead to errors.
So the real answer is (imo) that it is fine to anthropomorphise but recognize that while doing so can provide utility and help us understand, it is WRONG. Recognizing that it is not right and cannot be right provides us with a constant reminder to reevaluate. Use it, but double check, and keep checking making sure you understand the limitations of the analogy. Understanding when and where it applies, where it doesn't, and most importantly, where you don't know if it does or does not. The last is most important because it helps us form hypotheses that are likely to be testable (likely, not always. Also, much easier said than done).
So I pick a "grey area". Anthropomorphization is a tool that can be helpful. But like any tool, it isn't universal. There is no "one-size-fits-all" tool. Literally, one of the most important things for any scientist is to become an expert at the tools you use. It's one of the most critical skills of *any expert*. So while I agree with you that we should be careful of anthropomorphization, I disagree that it is useless and can never provide information. But I do agree that quite frequently, the wrong tool is used for the right job. Sometimes, hacking it just isn't good enough.
I don't agree. Most LLMs have been trained on human data, so it is best to talk about these models in a human way.
Anthropomorphising implicitly assumes motivation, goals and values. That's what the core of anthropomorphism is - attempting to explain behavior of a complex system in teleological terms. And prompt escapes make it clear LLMs doesn't have any teleological agency yet. Whenever their course of action is, it is to easy to steer them of. Try to do it with a sufficiently motivated human.
>. Try to do it with a sufficiently motivated human.
That's what they call marketing, propaganda or brain washing, acculturation , education depending on who you ask and at which scale you operate, apparently.
1 reply →
How will you know when an AI has teleological agency?
1 reply →
Even the verb 'trained' is contentious wrt anthropomorphism.
Somewhat true but rodents can also be trained ...
7 replies →
> On the contrary, anthropomorphism IMO is the main problem with narratives around LLMs
I hold a deep belief that anthropomorphism is a way the human mind words. If we take for granted the hypothesis of Franz de Waal, that human mind developed its capabilities due to political games, and then think about how it could later lead to solving engineering and technological problems, then the tendency of people to anthropomorphize becomes obvious. Political games need empathy or maybe some other kind of -pathy, that allows politicians to guess motives of others looking at their behaviors. Political games directed the evolution to develop mental instruments to uncover causality by watching at others and interacting with them. Now, to apply these instruments to inanimate world all you need is to anthropomorphize inanimate objects.
Of course, it leads sometimes to the invention of gods, or spirits, or other imaginary intelligences behinds things. And sometimes these entities get in the way of revealing the real causes of events. But I believe that to anthropomorphize LLMs (at the current stage of their development) is not just the natural thing for people but a good thing as well. Some behavior of LLMs is easily described in terms of psychology; some cannot be described or at least not so easy. People are seeking ways to do it. Projecting this process into the future, I can imagine how there will be a kind of consensual LLMs "theory" that explains some traits of LLMs in terms of human psychology and fails to explain other traits, so they are explained in some other terms... And then a revolution happens, when a few bright minds come and say that "anthropomorphism is bad, it cannot explain LLM" and they propose something different.
I'm sure it will happen at some point in the future, but not right now. And it will happen not like that: not just because someone said that anthropomorphism is bad, but because they proposed another way to talk about reasons behind LLMs behavior. It is like with scientific theories: they do not fail because they become obviously wrong, but because other, better theories replace them.
It doesn't mean, that there is no point to fight anthropomorphism right now, but this fight should be directed at searching for new ways to talk about LLMs, not to show at the deficiencies of anthropomorphism. To my mind it makes sense to start not with deficiencies of anthropomorphism but with its successes. What traits of LLMs it allows us to capture, which ideas about LLMs are impossible to wrap into words without thinking of LLMs as of people?