Understanding Reasoning LLMs
magazine.sebastianraschka.com
magazine.sebastianraschka.com
A lot of things that aren't well-defined require reasoning, and not just in a "SWE is ambiguous" kind of way - for example, thinking about how to present/teach something in a good way, iterating with the learner, thinking about what context they could be missing, etc.
I find that all of these reasoning models really will overfit and overthink if you attach some level of math problem to it but it will barely think for anything else. I had friends suggest to me (I can't tell if in jest or seriously) that other fields don't require thinking, but I dunno, a lot of these "soft things" I think about really hard and don't have great solutions to.
I've always been a fan of self-learning, for example - wouldn't it be great to have a conversation partner who can both infer and understand your misconceptions about complex topics when trying to learn, just from a few sentences, and then guide you for that?
It's not like it's fundamentally impossible. These LLMs definitely can solve harder coding problems when you make them think. It's just that, I'm pretty sure (and it's really noticable with deepseek) that they're overfit towards coding/math puzzles in particular.
It's really noticable with deepseek when you ask its reasoning model to just write some boilerplate code... you can tell it's completely overfit because it will just overthink and overthink and overthink. But it doesn't do that for example, with "soft" questions. In my opinion, this points to the idea that it's not really deciding for itself "how much thinking is enough thinking" and that it's just really overfit. Which I think can be solved, again, but I think it's more of a training decision issue.
These concepts will be shattered in the long run hopefully, because they are so small.
Algorithmic puzzles, on the other hand, both require reasoning and are easy to verify.
There are other things in coding that are both useful and easy to verify: checking that the generated code follows formatting standards or generating outputs with a specific data schema and so on.
Another issue is, how much data can you synthesize in such a way, so that you can construct both the problem and solution, so that you know the answer before using it as a sample.
Ie, some problems are easier to make knowing you can construct the problem yourself, but if you were to solve said problems, with no prior knowledge, they would be hard to solve, and could be used as a scoring signal?
Ie, you are the Oracle and whatever model is being trained doesn't know the answer, only if it is right or wrong. But I don't know if the reward function must be binary or on a scale.
Does that make sense or is it wrong?
Sorry, long answer incoming. It is far from complete too but I think it will help build strong intuition around your questions.
Will knowledge transfer? That entirely depends on the new problem. It also entirely depends on how related the problem is. But also, what information was used to solve the pre-transfer state. Take LLMs for example. There's lots of works that have shown them being difficult to train for solving calculations. Where they will do well on problems with the same number of digits but this will degrade rapidly as number of digits increase. It can be weird to read some of these papers as there will sometimes be periodic relationships with the number of digits but that should give us information about how they're encoding the problems. But that lack of transferability indicates that despite the problem solving and what we'd believe is actually just the same problem, doesn't mean it is. So you have to be really careful here, because us humans are really fucking good at generalization (yeah, we also suck, but a big part is our proficiency makes us recognize where we lack. But also, this is more a "humans can" more than "humans do" type of thing. So be careful when comparing). This generalization is really because we're focused around building causal relationships, while on the other hand the ML algorithms are build around compression (i.e. fitting data). Which, if you notice, is the same issue I was pointing to above.
> Ie, you are the Oracle and whatever model is being trained doesn't know the answer, only if it is right or wrong. But I don't know if the reward function must be binary or on a scale.
I picked this because there's a rather famous paper about emergent abilities (I fucking hate this term[0]) in ML models[1], and a far less famous counter to it[2]. There's a lot of problems with [1] that require a different discussion but [2] shows how a big part of the issue is how many of the loss landscapes are fairly flat and so when feedback is discrete the smaller models just wonder around that flat landscape needing to get lucky to find the optima (btw, this also shows that technically this can be done too! But that would require different training methods and optimizers). But when giving them continuous feedback (i.e. you're wrong, but closer than your last guess), they are able to actually optimize. A big criticism of the work is that it is an unfair comparison because there are "right and wrong" answers here, but it'd be naive to not recognize that some answers are more wrong than others. Plus, their work shows a clear testable way we can confirm or deny if this works or not. We schedule learning rates, there's no reason you cannot schedule labels. In fact, this does work.
But also look at the ways they tackled these problems. They are entirely different. [1] tries to do proof by evidence while [2] uses proof by contradiction. Granted, [2] has an easier problem since they only need to counter the claims of [1], but that's a discussion about how you formulate proofs.
So I'd be very careful when using the recent advancements in ML as a framework for modeling reasoning. The space is noisy. It is undeniable that we've made a lot of advancements but there is some issues with what work gets noticed and what doesn't. A lot does come down to this proof by evidence fallacy. Evidence can only bound confidence, it can unfortunately not prove things. But this is helpful and well, we can bound our confidence to limit the search space before we change strategies, right? I picked [1] and [2] for a reason ;) And to be clear, I'm not saying [1] shouldn't exist as a paper or that the researchers were dumb for doing it. Read back on this paragraph, because we've got multiple meta layers here. It's good to place a flag in the ground, even if it is wrong, because you gotta start somewhere, and science is much much better at ruling things out than ruling things in. We more focus on proving things don't work until there's not much left and then accept those things (limits here too, but this is too long already).
I'll leave with this, because now there should be a lot of context that makes this much more meaningful: https://www.youtube.com/watch?v=hV41QEKiMlM
[0] It significantly diverges from the terminology used in fields such as physics. ML models are de facto weakly emergent by nature of composition. But the ML definition can entirely be satisfied by "Information was passed to the model but I wasn't aware of it" (again, same problem: exhaustive testing)
[1] (2742 citations) https://arxiv.org/abs/2206.07682
[2] (447 citations) https://arxiv.org/abs/2304.15004
So knowledge transfer is something incredibly specific and much more narrow than what I thought. They don't transfer concepts by generalization, but they compress knowledge instead, which I assume the difference is, that generalization is much more fluid, while compression is much more static, like a dictionary where each key has a probability to be chosen, and all the relationships are frozen, and the only generalization that happens, is the generalization which is an expression of the training method used, since the training method freezes it's "model of the world" into the weights so to say? So if the training method itself cannot generalize, but only compress, why would the resulting model that the training method produces? Is that understood correctly?
Does there exist a computational model, which can be used to analyse a training method and put a bound on the expressiveness of the resulting model?
It's fascinating that the emergent ability of models disappear if you measure them differently. Guess the difference is that "emergent abilities" are kinda nonsensical, since they have no explanation of causality (i.e. it "just" happens), and just seeing the model getting linearly better with training fits into a much more sane framework. That is, like you said, when your success metric is measuring discretely, you also see the model itself as discrete, and it hides the continuous hill climbing you would otherwise see the model exhibit with a different non-discrete metric.
But the model still gets better over time, so would you expect the model to get progressively worse on a more generalized metric, or does it only relate to the spikes in the graph that they talk about? IE, they answer the question of "why" jumps in performance are not emergent, but they don't answer why the performance keeps increasing, even if it is linear, and whether it is detrimental to other less related tasks?
And if you wanted to test "emergent" wouldn't it be more interesting to test the model on tasks, which would be much more unrelated to the task at hand? That would be to test generalization, more so as we see humans see it? So it wouldn't really be emergence, but generalization of concepts?
It makes sense that it is more straightforward to refute a claim by using contradiction. Would it be good practice for papers, to try and refute their own claims by contradiction first? I guess that would save a lot of time.
It's interesting about the knowledge leakage, because I was thinking about the concept of world simulations and using models to learn about scenarios through simulations and consequence. But the act of creating a model to perceive the world, taints the model itself with bias, so the difficulty lies in creating a model which can rearrange itself to get rid of incorrect assumptions, while disconnecting its initial inherent bias. I thought about models which can create other models etc, but then how does the model itself measure success? If everything is changing, then so is the metric, so the model could decide to change what it measures as well. I thought about hard coding a metric into the model, but what if the metric I choose is bad, and we are then stuck with the same problem of bias as well. So it seems like there are only two options, it either converges towards total uncontrollability or it is inherently biased, there's doesn't seem to be any in-between?
I admit I'm trying to learn things about ML I just find general intelligence research fascinating (neuroscience as well), but the more I learn, the more I realize I should really go back to the fundamentals and build up. Because even things which seem like they make sense on a surface level, really has a lot of meaning behind them, and needs a well-built intuition not from a practical level, but from a theoretical level.
From the papers I've read which I find interesting, it's like there's always the right combination of creativity in thinking, which sometimes my intuition/curiosity about things proved right, but I lack the deeper understanding, which can lead to false confidence in results.
I'll come back and retype some of what I said but I need to do some other stuff right now. So I'll say that you're asking really good questions and I think you're mostly understanding things.
So give you very quick answers:
Yes, things are frozen. There's active/online learning but even that will not solve all the issues at hand.
Yes, we can put bounds. Causal models naturally do this but statistics is all about this too. Randomness is a measurement of uncertainty. Note that causal models are essentially perfect embeddings. Because if you've captured all causal relationships, you gain no more value from additional information, right?
Also note that we have to be very careful about assumptions. It is always important to uncover what assumptions have been made and what the implications are. This is useful in general problems solving and applies to anything in your life, not just AI/ML/coding. Unfortunately, assumptions are almost never explicitly stated, so you got to go hunting.
See how physics defines strong emergence and weak emergence. There are no known strongly emerging phenomena and we generally believe they do not exist. For weakly emerging, well it's rather naive to discuss this in the context of ML if we're dedicating so little time and effort to interpretation, right? That's kinda the point I was making previously about not being able to differentiate an emergent phenomena from not knowing we gave it information.
For the "getting better" it is about the spikes. See the first two figures and their captions in the response paper.
More parameters do help btw, but make sure you distinguish the difference between a problem being easier to solve and a problem not being solvable. The latter is rather hard to show. But the paper is providing strong evidence to the underlying issues being about the ease of problem solving rather than incapacity.
Proof is hard. There's nothing wrong with being empirical, but we need to understand that this is a crutch. It is evidence, not proof. We leaned on this because we needed to start somewhere. But as progress is made so too must all the metrics and evaluations. It gets exponentially harder to evaluate as progress is made.
I do not think it is best to put everyone in ML into the theory first and act like physicists. Rather we recognize the noise and do not lock out others from researching other ideas. The review process has been contaminated and we lost sight. I'd say that the problem is that we look at papers as if we are looking at products. But in reality, papers need to be designed with understanding the experimental framework. What question is being addressed, are variables being properly isolated, and do the results make a strong case for the conclusion? If we're benchmark chasing we aren't doing this and we're providing massive advantage to "gpu rich" as they can hyper-parameter tune their way to success. We're missing a lot of understanding because of this. You don't need state of the art to prove a hypothesis. Nor to make improvements on architectures or in our knowledge. Benchmarks are very lazy.
For information leakage, you can never remove the artist from the art, right? They always leave part of themselves. That's okay, but we must be aware of the fact so we can properly evaluate.
Take the passion, and dive deep. Don't worry about what others are doing, and pursue your interests. That won't make you successful in academia, but it is the necessary mindset of a researcher. Truth is no one knows where we're going and which rabbit holes are dead ends (or which look like dead ends but aren't). It is good to revisit because you table questions when learning, but then we forget to come back to them.
> needs a well-built intuition not from a practical level, but from a theoretical level.
It seems that if this was true that formal verification should be performed much more frequently. No doubt would this be cheaper than hiring pen testers, paying out bug bounties, or incurring the costs of getting hacked (even more so getting unknowingly hacked). It also seems to reason that the NSA would have a pretty straight forward job: grab source code, run verification, exploit flaws, repeat the process as momentum is in your favor.
That should be easy to reason through even if you don't really know the formal verification process. We are constantly bombarded with evidence that testing isn't sufficient. This is why it's been so weird for me, because it's talked about in schooling and you can't program without running into this. So why has it been such a difficult lesson to learn?
Tool use means you can click a button and make sure it transitioned to the next described UI screen verified again with multimodal as well.
Perhaps time spent thinking isn't a great metric, but just looking at deepseek's logs for example, it's chain of thought for many of these "softer" questions are basically just some aggregate wikipedia article. It'll brush on one concept, then move on, without critically thinking about it.
However, for coding problems, no matter how hard or simple, you can get it to just go around in circles, second guess itself, overthink it. And I think this is is kind of a good thing? The thinking at least feels human. But it doesn't even attempt to do any of that for any "softer" questions, even with a lot of my prompting. The highest I was able to get was 50 seconds, I believe (time isn't exactly the best metric, but I'd rate the intrinsic quality of the CoT lower IMO). Again, when I brought this up to people they suggested that math/logic/programming just intrinsically is harder... I don't buy it at all.
I totally agree that it's harder to train for though. And yes, they are next token predictors, shouldn't be hasty to anthropomorphize, etc. But like.... it actually feels like it's thinking when it's coding! It genuinely backtracks and explores the search space somewhat organically. But it won't afford the same luxury for softer questions is my point.
If it's not well defined then you can't do RL on it because without a clear cut reward function the model will learn to do some nonsense instead, simple as.
So, I question the hypothesis.
But as for machine RL in practice, you always need a reward model and once you go past things you can solidly verify like code that can be compiled/executed to check for errors or math that can be computed it becomes very easy to end up doing nonsense. If the reward model is a human judge (i.e. RLHF) then the results can be pretty good, but it doesn't scale and there's no accounting for taste even in humans.
I don’t know if vector spaces and transformers can encode that ability.
It’s a key skill in thinking and writing. I definitely tailor my writing for my audience in order to get a point across. Often the goal isn't simply an answer, it’s a convincing answer.
Update: forgot a word
I rolled out reasoning for my interactive reader app, and I tried to extract R1's reasoning traces to use with my existing models, but found its COT for writing wasn't particularly useful*.
Instead of leaning on R1 I came up with my own framework for getting the LLM to infer the reader's underlying frame of mind through long chains of thought, and with enough guidance and a some hand edited examples I was able to get reasoning traces that demonstrated real insight into reader behavior.
Obviously it's much easier in my case because it's an interactive experience: the reader is telling the AI what action they'd like the main character to try, and that in turn is an obvious hint into how they want things go otherwise. But readers don't want everything to go perfectly every time, so it matters that the LLMs are also getting very good picking up on non-obvious signals in reader behavior.
With COT the model infers the reader expectations and state of mind in its own way and then "thinks" itself into how to subvert their expectations, especially in ways that will have a meaningful payoff for the specific reader. That's a huge improvement over an LLM's typical attempts at subversion which tend to bounce between being too repetitive to feel surprising, or too unpredictable to feel rewarding.
(* I agree that current reasoning oriented post-training over-indexes on math and coding, mostly because the reward functions are easier. But I'm also very ok with that as someone trying to compete in the space)
Reasoning verification in the general case is harder - it seems "LLM as judge" (ask an LLM if it sounds right!) seems to be the general solution.
Meanwhile, the simpler, non-reasoning models got it: "Yup, test succeeded!" (Llama 3.2 was quite chipper about the test succeeding.)
Everyone's worried about the paperclip optimizers and I'm wondering if we're bringing forth Paranoia: https://en.wikipedia.org/wiki/Paranoia_(role-playing_game)
I know there is work being done with e.g. Lean integration with ChatGPT, but that's not what I mean exactly -- there's still this shakey natural-language-trained-LLM glue in the driver's seat
Like I'm envisioning something that has the creativity to try different things, but then JIT compile their chain of thought, and avoid bad paths
The reason why it seems the most reasonable path is because when you create restrictions like this you hamper search viability (and in a high multi-dimensional subspace, that's a massive loss because you can arrive at a result from many directions). It's like regular genetic programming vs typed-genetic programming. When you discard all your useful results, you can't go anywhere near as fast. There will be a threshold where constructivist, generative schemes (e.g. reasoning with automata and all kinds of fun we've neglected) will be the way forward, but I don't think we've hit that point yet. It seems to me that such a point does exist because if you have fast heuristics on when types unify, you no longer hamper the search speed but gain many benefits in soundness.
One of the greatest human achievements of all time is probably this latent embedding space -- one that we can actually interface with. It's a new lingua franca.
These are just my cloudy current thoughts.
As long as you're using something statistical like transformers, you're going to need deterministic bolt-ons like Lean.
You don't need to generate semantically valid reasoning 100% of time for such an approach to be useful. You just need to use semantic data to bias them to follow semantically valid paths more often than not (and sometimes consider using constraint solving on the spot, like offloading into a SMT solver or even incorporating it in the model somehow; it would be nice to have AI models that can combine the strengths of both GPUs and CPUs). And, what's more useful, verify that the reasoning is valid at the end of the train of thought, and if it is not, bail out and attempt something else.
If you see AI as solving an optimization problem (given a question, give a good answer) it's kind of evident that you need to probe the space of ideas in an exploratory fashion, sometimes making unfounded leaps (of the "it was revealed to me in a dream" sort), and in this sense it could even be useful that AI can sometimes hallucinate bullshit. But they need afterwards to come with a good justification for the end result, and if they can't find one they are forced to discard their result (even if it's true). Just like humans often come up with ideas in an irrational, subconscious way, and then proceed to rationalize them. One way to implement this kind of thing is to have the LLM generate code for a theorem prover like Coq or Lean, and then at the end run the code - if the prover rejects the code, the reasoning can't possibly be right, and the AI needs to get back to the drawing board
(Now, if the prover accepts the code, the answer may still be wrong, if the premises were encoded incorrectly - but it would still be a net improvement, specially if people can review the Coq code to spot mistakes)
I went into a little more detail here last week: https://news.ycombinator.com/item?id=42871894
> My personal insight is that "reasoning" is simply the application of a probabilistic reasoning manifold on an input in order to transform it into constrained output that serves the stability or evolution of a system.
> This manifold is constructed via learning a decontextualized pattern space on a given set of inputs. Given the inherent probabilistic nature of sampling, true reasoning is expressed in terms of probabilities, not axioms. It may be possible to discover axioms by locating fixed points or attractors on the manifold, but ultimately you're looking at a probabilistic manifold constructed from your input set.
I've been writing and working on this problem a lot over the last few months and hopefully will have something more formal and actionable to share eventually. Right now I'm at the, "okay, this is evident and internally consistent, but what can we actually do with it that other techniques can't already accomplish?" phase that a lot of these metacognitive theories get stuck on.
What? I'm sorry, but this is ridiculous. You can make plenty of sound logical arguments in an empirical vacuum. This is why we have proof by induction - some things can't be verified by taking samples.
https://deepmind.google/discover/blog/ai-solves-imo-problems...
But for a rigorously structured language with formal fixed meaning... Now the the LLM has no advantage anymore, only serious drawbacks and limitations. Save yourself millions of dollars and just write a normal parser, expression evaluator, SAT solver, etc.
You'll get answers faster, using fewer resources, with fewer fundamentally unfixable bugs, and it will actually be able to do math.
I say this because I am constantly annoyed by poor, opaque writing in other instances. In this case, DS doesn’t need to try to sound smart. The results speak for themselves.
I recommend anyone who is interested in the topic to read the R1 paper, their V3 paper, and DeepSeekMath paper. They’re all worth it.
>Whether and how an LLM actually "thinks" is a separate discussion.
The "whether" is hardly a discussion at all. Or, at least one that was settled long ago.
"The question of whether a computer can think is no more interesting than the question of whether a submarine can swim."
--Edsger Dijkstra
“[…] it tends to divert the research effort into directions in which science can not—and hence should not try to—contribute.” is a pretty myopic take.
It's not interesting if you're asking it to count to a billion.
In that regard, it's a bit of a false analogy, because submarines were never meant to mimic human swimming. But AI development often has that motivation. We could just say we're developing powerful intelligence amplification tools for use by humans, but for whatever reason, everyone prefers the scifi version. Augumented Intelligence is the forgotten meaning of AI.
Submarines never replaced human swimming (we're not whales), they enabled human movement under water in a way that wasn't possible before.
I do philosophy and it will take an exaggeration I give it, and call it fact.
The non reasoning models will call me out. lol
However, we know the base model is DeepSeek V3. From the DeepSeek V3 technical report, paragraph in 5.1. Supervised Fine-Tuning:
> Reasoning Data. For reasoning-related datasets, including those focused on mathematics, code competition problems, and logic puzzles, we generate the data by leveraging an internal DeepSeek-R1 model. Specifically, while the R1-generated data demonstrates strong accuracy, it suffers from issues such as overthinking, poor formatting, and excessive length. Our objective is to balance the high accuracy of R1-generated reasoning data and the clarity and conciseness of regularly formatted reasoning data.
In 5.4.1 they also talk about some ablation experiment by not using the "internal DeepSeek-R1" generated data.
While the "internal DeepSeek-R1" model is not explained, I would assume this is a DeepSeek V2 or V2.5 tuned for chain of thought. Therefore, it seems to me the "aha moment" is just promoting the behaviour that was already present in V3.
In the "Self-evolution Process of DeepSeek-R1-Zero"/ Figure 3 they claim reinforcement learning also leads to the model generating longer CoT sequences, but again, this comes from V3, they even mention the fine tuning with "internal R1" led to "excessive length".
None of the blogpost, news, articles I have read explaining or commenting on DeepSeek R1 takes this into account. The community is scrambling to re-implement the pipeline (see open-r1).
At this point, I feel like I took a crazy pill. Am I interpreting this completely wrong? Can someone shed some light on this?
I didn't say people did. But overcoming real challenges seems to be a big part of feeling alive, and I wonder if we really all would settle back into going for walks all day or whatever we could do that entertain us without needing others to work to provide the entertainment. Perhaps the WALL-E future, where we sit in chairs? But with AI-generated content?
My response a few months ago (Scroll down to my username and read that discussion):
https://news.ycombinator.com/item?id=41997727
If only I knew DeepSeek was going to tank the market with something as simple as that lol.
Note to self, take your intuition seriously.
You could use reward functions that do a lot more complicated stuff than "ground_truth == boxed_answer". You could, for example split the "CoT" in paragraphs, and count how many paragraphs match whatever you consider a "good answer" in whatever topic you're trying to improve. You can use embeddings, or fuzzy string matches, or even other LLMs / reward models.
I think math and coding were explored first because they're easier to "score", but you could attempt it with other things as well.
They are certainly capable of doing is a wide variety of computations that simulate reasoning, and maybe that's good enough for your use case. But it is unpredictably brittle unless you spend a lot on o1-pro (and even then...). Raschka has a line about "whether and how an LLM actually 'thinks' is a separate discussion" but this isn't about semantics. R1 clearly sucks at deductive reasoning and you will not understand "reasoning" LLMs if you take DeepSeek's claims at face value.
It seems especially incurious for him to copy-paste the "a-ha moment" from Deepseek's technical report without critically investigating it. DeepSeek's claims are unscientific, without real evidence, and seem focused on hype and investment:
This moment is not only an "aha moment" for the model but also for the researchers observing its behavior. It underscores the power and beauty of reinforcement learning: rather than explicitly teaching the model on how to solve a problem, we simply provide it with the right incentives, and it autonomously develops advanced problem-solving strategies.
The "aha moment" serves as a powerful reminder of the potential of RL to unlock new levels of intelligence in artificial systems, paving the way for more autonomous and adaptive models in the future.
Your link says that R1, not all models like R1, fails at generalization.
Of particular note:
> We expose DeepSeek R1 to the variations of AIW Friends problem and compare model behavior to o1-preview, o1-mini and Claude 3.5 Sonnet. o1-preview handles the problem robustly, DeepSeek R1 shows strong fluctuations across variations with distribution very similar to o1-mini.
They came into this with the assumption that LLMs are just a cheap trick. As a result, they deliberately searched for an example of failure, rather than trying to do an honest assessment of generalization capabilities.
And lo and behold, they still found a glaring failure. You can't fault them for not buying into the hype.
If given a few hundred tries, I bet I could find an example where you reason poorly too. Wikipedia has a whole list of common failure modes of human reasoning: https://en.wikipedia.org/wiki/List_of_fallacies
I experience LLM "reasoning" failure several times a day, yet I find them useful.
Perhaps the type of RL used to train them also has an effect on generalization, but choice of training data has to play a large part.
That's not actually what your link says. The tweet says that it solves the simple problem (that they originally designed to foil base LLMs) so they had to invent harder problems until they found one it could not reliably solve.
For example, many manipulative tactics (e.g. the fake sorry here, responding to something else than was said, ...) and lying can be considered insults (they literally assume the reader is not smart enough to notice, hence at least as severe as calling someone an idiot) but it's hard for a mod to notice without putting in a lot of effort to understand the situation.
Yet when people (very mildly) punish this behavior by calling it out, they are often noticed by the mod because the call out is more visible.
The simple argument is that when you write things like this:
> I am unwilling to invest any more time into arguing with someone unwilling to use reasoning
...you're bluntly breaking the rules, regardless of what another commenter is doing, be it subtly or blatantly abusive.
I agree that there are countless varieties of passive-aggressive swipe and they rub me the wrong way too, but the argument that those are "just as bad, merely less visible" is not accurate. Attacking someone else is not justified by a passive-aggressive "sorry", just as it is not ok to ram another vehicle when a driver cuts you off in traffic.
The issue is many rules are written as absolute statements which expect some kind of higher power (mods, police, ...) to be the only side to deal punishment. This obviously breaks in many situations - when the higher power is understaffed, when it's corrupt or when there is no higher power (war between nation states).
I would like to see attempts to make rules relative. Treat others how you want to be treated but somebody treating you badly gives you the right to also treat them badly (within reason - proportionally). It would probably lead to conflict being more visible (though not necessarily being more numerous) but it would allow communities to self-police without the learned helplessness of relying on a higher power. Aggressors would gain nothing by provoking others because others would be able to defend themselves.
Doing this is hard, especially at scale. Many people who behave poorly towards others back off when they are treated the same way but there also needs to be a way to deal with those who never back down. When a conflict doesn't resolve itself and mods step in, they should always take into account who started it, and especially if they have a pattern of starting conflict.
There's another related issue - there is a difference between fairness/justice and peace. Those in power often fight for the first on paper but have a much stronger incentive to protect the second.
I agree, it is a problem—but it is (almost by definition) less of a problem than aggression which does cross the threshold. If every user would give up being overtly abusive for being covertly abusive, that wouldn't be great—but it would be better, not least because we could then raise the bar to make that also unacceptable.
(I'm not sure this analogy is helpful, but to me it's comparable to the difference between physical violence and emotional abuse. Both are bad, but society can't treat them the same way—and that despite the fact emotional abuse can actually be worse in some situtations.)
> somebody treating you badly gives you the right to also treat them badly (within reason - proportionally)
I can tell you why that doesn't work (at least not in a context like HN where my experience is): because everyone overestimates the provocations and abuses done by the other, and underestimates the ones done by themselves. If you say the distortion is 10x in each case, that's a 100x skew in perception [1]
As a result, no matter how badly people are behaving, they always feel like the other person started it and did worse, and always feel justified.
In other words, to have that as a rule would amount to having no rule. In order to be even weakly effective, the rule needs to be: you can't be abusive in comments regardless of what other commenters are doing or you feel they are doing [2].
[1] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
[2] https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
Unless you also take into account scale (how often the person does it or how many other people do it) and second-order effects (people who fall for the manipulation and spread it further or act on it). For this reason, I very much prefer people who insult me honestly and overtly, at least I know where I stand with them and at least other people are less likely to get influenced by them.
> I'm not sure this analogy is helpful
This is actually a very rare occasion when an analogy is helpful. As you point out, the emotional abuse can (often?) be worse. TBH when it "escalates" to being physical, it often is a good thing because it finally 1) gives the target/victim "permission" to ask for help 2) it makes it visible to casual observers, increasing the likelihood of intervention 3) it can leave physical evidence and is easily spotted by witnesses.
(I witnessed a whole bunch of bullying and attempts at bullying at school and one thing that remained constant is that people who fought back (retaliated) were left alone (eventually). It is also an age where physical violence is acceptable and serious injuries were rare (actually I don't recall a single one from fighting). This is why I always encourage people to fight back, not only is it effective but it teaches them individual agency instead of waiting for someone in a position of power to save them.)
> I can tell you why that doesn't work
I appreciate this datapoint (and the fact you are open to discussing it, unlike many mods). I agree that it's often hard to distinguish between mistake and malice. For example I reacted to the individual instance because of similar comments I ran into in the past but I didn't check if the same person is making fallacious arguments regularly or if it was a one-off.
But I also have experiences with good outcomes. One example stands out - a guy used a fallacy when arguing with me, i asked him to not do that, he did it again so i did it twice to him as well _while explaining why I am doing it_. He got angry at first, trying to call me out for doing something I told him not to do, but when I asked him to read it again and pointed out that the justification was right after my message with the fallacy (not post-hoc after being "called out"), he understood and stopped doing it himself. It was as if he wasn't really reading my messages at first but reversing the situation made him pay actual attention.
I think the key is that it was a small enough community that 1) the same people interacted with each other repeatedly and that 2) I explained the justification as part of the retaliation.
Point 1 Will never be possible at the scale of HN, though I would like to see algorithmic approaches to truth and trust instead of upvotes/downvotes which just boil down to agree/disagree. Point 2 can be applied anywhere and if mods decide to step in, it IMO is something they should take into account.
Anyway, thanks for the links, I don't have time to go through other people's arguments rn but I will save it for later as it is good to know this comes up from time to time and I am not completely crazy when I see something wrong with the standard threshold-based approach.
Oh and you didn't say it explicitly but I feel like you understand the difference between rules and right/wrong given your phrasing. That is a very nice thing to see if I am correct (though I have no doubt your phrasing was refined by years or trial and error as to what is effective). In general, I believe it should always be made clear that rules exist for practical reasons, not pretend they are some kind of codification of morality.
Another way of putting it is that the rules aren't moral or ethical—they're just the rules of the game we're trying to play here. Different games naturally have different rules.
https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
But the link demonstrates the opposite- these models absolutely are able to reason out of distribution, just not with perfect fidelity. The fact that they can do better than random is itself really impressive. And o1-preview does impressively well, only vary rarely getting the wrong answer on variants of that Alice in Wonderland problem.
If you would listen to most of the people critical of LLMs saying they're a "stochastic parrot" - it should be impossible for them to do better than random on any out of distribution problem. Even just changing one number to create a novel math problem should totally stump them and result in entirely random outputs, but it does not.
Overall, poor reasoning that is better than random but frequently gives the wrong answer is fundamentally, categorically entirely different from being incapable of reasoning.
Thought I was about to be learn!
Instead, I just met an asshole.
There is also a much bigger group of people that haven't really tried anything beyond GPT-3.5, which was the best you could get without paying a monthly subscription for a long time. One of the biggest reasons for r1 hype, besides the geopolitical angle, was people could actually try a reasoning model for free for the first time.
Secondly, you're ignoring the fact that the community of voices with experience in data sciences, computer science and artificial intelligence themselves are split on the qualities or lack of them in current AI. GPT and LLM are very interesting but say little or nothing to me of new theory of mind, or display inductive logic and reasoning, or even meet the bar for a philosophers cave solution to problems. We've been here before so many, many times. "Just a bit more power captain" was very strong in connectionist theories of mind. fMRI brains activity analytics, you name it.
So yes. There are a lot of "us" who are pushing back on the hype, and no we're not a mini cult.
The simple fact they can generate language so well makes me think... maybe language itself carries more weight than we originally thought. LLMs can get to this point without personal experience and embodiment, it should not have been possible, but here we are.
I think philosophers are lagging science now. The RL paradigm of agent-environment-reward based learning seems to me a better one than what we have in philiosophy now. And if you look at how LLMs model language as high dimensional embedding spaces .. this could solve many intractable philosophical problems, like the infinite homunculus regress problem. Relational representations straddle the midpoint between 1st and 3rd person, offering a possible path over the hard problem "gap".
Parrots are self aware, complex reasoning brains which can solve problems in geometry, tell lies, and act socially or asocially. They also have complex vocal chords and can perform mimicry. Very few aspects of a parrots behaviour are stochastic but that also underplays how complex stochastic systems can be in their production. If we label LLM products as Stochastic Parrots it does not mean they like cuttlefish bones or are demonstrably modelled by Markov chains like Mark V Shaney.
The question to me seems: If we expand on this architecture (in some direction, compute, size etc.), will we get something much more powerful? Whereas if you give nature more time to iterate on the parrot, you'd probably still end up with a parrot.
There's a giant impedance mismatch here (time scaling being one). Unless people want to think of parrots being a subset of all animals, and so 'stochastic animal' is what they mean. But then it's really the difference of 'stochastic human' and 'human'. And I don't think people really want to face that particular distinction.
Like I said elsewhere in this overall thread, we've been here before. Yes, you do see improvements in larger datasets, weighted models over more inputs. I suggest, I guess I believe (to be more honest) that no amount of "bigger" here will magically produce AGI simply because of the scale effect.
There is no theory behind "more" and that means there is no constructed sense of why, and the absence of abstract inductive reasoning continues to say to me, this stuff isn't making a qualitative leap into emergent anything.
It's just better at being an LLM. Even "show your working " is pointing to complex causal chains, not actual inductive reasoning as I see it.
Induction is a hard problem, but humans can skip infinite compute time (I don't think we have any reason to believe humans have infinite compute) and still give valid answers. Because there's some (meta)-structure to be exploited.
Architecturally if machines / NN can exploit this same structure is a truer question.
The magical missing ingredient here is search. AlphaZero used search to surpass humans, and the whole Alpha family from DeepMind is surprisingly strong, but narrowly targeted. The AlphaProof model uses LLMs and LEAN to solve hard math problems. The same problem solving CoT data is being used by current reasoning models and they have much better results. The missing piece was search.
Or, more simply.. just hack into a bunch of aws accounts, spin up machines, boom.
You don't seem to understand how they work, they recurse their solution meaning if they have remembered components it parrots back sub solutions. Its a bit like a natural language computer, that way you can get them to do math etc, although the instruction set isn't of a turing language.
They can't recurse sub sub parts they haven't seen, but problems that has similar sub parts can of course be solved, anyone understands that.
I don't think anyone understands how they work- these type of explanations aren't very complete or accurate. Such explanations/models allow one to reason out what types of things they should be capable of vs incapable of in principle regardless of scale or algorithm tweaks, and those predictions and arguments never match reality and require constant goal post shifting as the models are scaled up.
We understand how we brought them about via setting up an optimization problem in a specific way, that isn't the same at all as knowing how they work.
I tend to think in the totally abstract philosophical sense, independent of the type of model, at the limit of an increasingly capable function approximator trained on an increasingly large and diverse set of real world cause/effect time series data, you eventually develop and increasingly accurate and general predictive model of reality organically within the model. Some model types do have fundamental limits in their ability to scale like this, but we haven't yet found one with these models.
It is more appropriate to objectively test what they can and cannot do, and avoid trying to infer what we expect from how we think they work.
What surprises us is the behaviors coming out of that process.
But surprise isn't magic, magic shouldn't even be on the list of explanations to consider.
This is similar to how understanding chemistry doesn’t imply understanding biology, or understanding how a brain works.
We also don't understand exactly how humans reason, so any claims that humans are capable of reasoning is also mostly an observation about abilities/capabilities.
Yes we do, we literally built them.
> We understand how we brought them about via setting up an optimization problem in a specific way, that isn't the same at all as knowing how they work.
You're mistaking "knowing how they work" with "understanding all of the emergent behaviors of them"
If I build a physics simulation, then I know how it works. But that's a separate question from whether I can mentally model and explain the precise way that a ball will bounce given a set of initial conditions within the physics simulation which is what you seem to be talking about.
By knowing how they work I specifically mean understanding the emergent capabilities and behaviors, but I don't see how it is a mistake. If you understood physics but knew nothing about cars, you can't claim to understand how a car works "simple, it's just atoms interacting according to the laws of physics." That would not let you, e.g. explain its engineering principles or capabilities and limitations in any meaningful way.
I am sure this is improperly worded, I apologise.
However, our brains, like language models, rely on implicit, distributed representations of concepts and rules.
So the intelligble representations of conceptual reasoning machines are maybe too strong a requirement for 'reasoning' unless you want to exclude humans too.
I mean, we already knew only a handful of companies with capital could train them, but at least the principles, algorithms, etc. were accessible to individuals who wanted to create their own - much simpler - models.
it seems that era is quickly ending, and we are entering the era of truly "magic" AI models that no one knows how they work because companies keep their secret sauces...
At the end of the day current O1-like reasoning models are still just fine-tuned LLMs, and don't even need RL if you have access to (or can generate) a suitable training set. The DeepSeek R1 paper outlined their bootstrapping process, and HuggingFace (and no doubt others) are trying to duplicate it.