> LLMs are PhD-level reasoners in math and science, yet they fail at children's puzzles. How is this possible?
Because they are not.
Pattern matching questions on a contrived test is not the same thing as understanding or reasoning.
It’s the same reason why most of the people who pass your leetcode tests don’t actually know how to build anything real. They are taught to the test not taught to reality.
gwd 5 hours ago [-]
> Pattern matching questions on a contrived test is not the same thing as understanding or reasoning.
Do submarines swim? I don't really care if it gets me where I want to go. The fact is that just two days ago, I asked Claude to look at some reasonably complicated concurrent code to which I had added a new feature, and asked it to list what tests needed to be added; and then when I asked GPT-5 to add them, it one-shot nailed the implementations. I've written a gist of it here:
Seriously just even read the description of the test it's trying to write.
In order to one-shot that code, it had to understand:
- How the cache was supposed to work
- How conceptually to set up the scenario described
- How to assemble golang's concurrency primitives (channels, goroutines, and waitgroups), in the correct order, to achieve the goal.
Did it have a library of concurrency testing patterns in its head? Probably -- so do I. Had it ever seen my exact package before in its training? Never.
I just don't see how you can argue with a straight face that this is "pattern matching". If that's pattern matching, then pattern matching is not an insult.
If anything, the examples in this article are the opposite. Take the second example, which is basically 'assemble these assorted pieces into a rectangle'. Nearly every adult has assembled a minimum of dozens of things in their lives; many have assembled thousands of things. So it's humans in this case who are simply "pattern matching questions on a contrived test", and the LLMs, which almost certainly didn't have a lot of "assemble these items" in their training data, that are reasoning out what's going on from first principles.
HarHarVeryFunny 57 minutes ago [-]
> Do submarines swim?
It doesn't matter HOW LLMs "swim" as long as they can, but the point being raised is whether they actually can.
It's as if LLMs can swim in the ocean, in rough surf, but fail to swim in rivers or swimming pools, because they don't have a generalized ability to swim - they've just been RL-trained on the solution steps to swimming in surf, but since those exact conditions don't exist in a river (which might seem like a less challenging environment), they fail there.
So, the question that might be asked is when LLMs are trained to perform well in these vertical domains like math and programming, where it's easy to verify results and provide outcome- or process-based RL rewards, are they really learning to reason, or are they just learning to pattern match to steer generation in the direction of problem-specific reasoning steps that they had been trained on?
Does the LLM have the capability to reason/swim, or is it really just an expert system that has been given the rules to reason/swim in certain cases, but would need to be similarly hand fed the reasoning steps to be successful in other cases?
I think the answer is pretty obvious given that LLM's can't learn at runtime - can't try out some reasoning generalization they may have arrived at, find that it doesn't work in a specific case, then explore the problem and figure it out for next time.
Given that it's Demis Hassabis who it pointing out this deficiency of LLMs (and has a 5-10 year plan/timeline to fix it - AGI), not some ill-informed LLM critic, it seems silly to deny it.
gwd 13 minutes ago [-]
> I think the answer is pretty obvious given that LLM's can't learn at runtime - can't try out some reasoning generalization they may have arrived at, find that it doesn't work in a specific case, then explore the problem and figure it out for next time.
This is just a problem of memory. Supposing that an LLM did generate a genuinely novel insight, it could in theory they could write a note for itself so that next time they come online, they can read through a summary of the things they learned. And it could also write synthetic training data for itself so that the next time they're trained, that gets incorporated into its general knowledge.
OpenAI allows you to fine-tune GPT models, I believe. You could imagine a GPT system working for 8 hours in a day, then spending a bunch of time looking over all its conversation looking for patterns or insights or things to learn, and then modifying its own fine-tuning data (adding, removing, or modifying as appropriate), which it then used to train itself overnight, waking up the next morning having synthesized the previous day's experience.
naasking 42 minutes ago [-]
> are they really learning to reason, or are they just learning to pattern match to steer generation in the direction of problem-specific reasoning steps that they had been trained on?
Are you sure there's a real difference? Do you have a definition of "reasoning" that excludes this?
HarHarVeryFunny 14 minutes ago [-]
I define intelligence as prediction (degree of ability to use past experience to correctly predict future action outcomes), and reasoning/planning as multi-step what-if prediction.
Certainly if a human (or some AI) has learned to predict/reason over some domain, then what they will be doing it pattern matching to determine the generalizations and exceptions that apply in a given context (including a hypothetical context in a what-if reasoning chain), in order to be able to select a next step that worked before.
However, I think what we're really talking about here isn't the mechanics of applying learnt reasoning (context pattern matching), but rather the ability to reason in the general case, which requires the ability to LEARN to solve novel problems, which is what is missing from LLMs.
A system that has a fixed set of (reasoning/prediction) rules, but can't learn new ones for itself, seems better regarded as an expert system. We need to make the distinction between can a system that can only apply rules, and one that can actually figure out the rules in the first place.
In terms of my definitions of intelligence and reasoning, based around ability to use past experience to learn to predict, then any system that can't learn from fresh experience doesn't meet that definition.
Of course in humans and other intelligent animals the distinction between past and ongoing experience doesn't apply since they can learn continually and incrementally (something that is lacking from LLMs), so for AI we need to use a different vocabulary, and "expert system" seems the obvious label for something that can use rules, but not discover them for itself.
gwd 22 minutes ago [-]
So I do think there are two distinct types of activities involved in knowledge work:
1. Taking established techniques or concepts and appropriately applying them to novel situations.
2. Inventing or synthesizing new, never-before-seen techniques or concepts
The vast majority of the time, humans do #1. LLMs certainly do this in some contexts as well, as demonstrated by my example above. This to me counts as "understanding" and "thinking". Some people define "understanding" such that it's something only humans can do; to which I respond, I don't care what you call it, it's useful.
Can LLMs do #2? I don't know. They've got such extensive experience that how would you know if they'd invented a technique vs had seen it somewhere?
But I'd venture to argue that most humans never or rarely do #2.
freejazz 23 minutes ago [-]
It seems readily apparent there is a difference given their inability to do tasks we would otherwise reasonably describe as achievable via basic reasoning on the same facts.
OtherShrezzing 46 minutes ago [-]
>and the LLMs, which almost certainly didn't have a lot of "assemble these items" in their training data
I don't think this assumption is sound. Humans write a huge amount on "assemble components x and y to make entity z". I'd expect all LLMs to have consumed every IKEA type instruction manual, the rules for Jenga, all geometry textbooks and papers ever written.
vlovich123 33 minutes ago [-]
I could be mistaken but generally LLMs cannot tackle out-of-domain problems whereas humans do seem to have that capability. Relatedly, the energy costs are wildly different suggesting that LLMs are imitating some kind of thought but not simulating it. They’re doing a remarkable job of passing the Turing test but that says more about the limitations of the Turing test than it does about the capabilities of the LLMs.
amelius 53 minutes ago [-]
Most of our coding is just plumbing. Getting data from one place to where it needs to be. There is no advanced reasoning necessary. Just a good idea of the structure of the code and the data-structures.
Even high school maths tests are way harder than what most professional programmers do on a daily basis.
Akronymus 3 hours ago [-]
> I just don't see how you can argue with a straight face that this is "pattern matching". If that's pattern matching, then pattern matching is not an insult.
IMO its still "just" a, very good, autocomplete. No actual reasoning, but lots of statistics on what is the next token to spit out.
NoahZuniga 3 hours ago [-]
> Do submarines swim?
That's the main point of the parent comment. Arguing about the definition of "reasoning" or "pattern matching" is just a waste of time. What really matters is if it produces helpful output. Arguing about that is way better!
Instead of saying: "It's just pattern matching -> It won't improve the world", make an argument like: "AI's seem to have trouble specializing like humans -> adopting AI will increase error rates in business processes -> due to the amount of possible edge cases, most people will get into an edge case with no hope of escaping it -> many people's lives will get worse".
The first example relies on us agreeing on the definition of pattern matching, and then taking a conclusion based on how those words feel. This has no hope of convincing me if I don't like your definition! The second one is an argument that could potentially convince me, even if I'm an AI optimist. It is also just by itself an interesting line of reasoning.
ozgung 2 hours ago [-]
No it's not "just a very good autocomplete". I don't know why people repeat this thing (it's wrong) but I find it an extremely counterproductive position. Some people just love to dismiss the capabilities of AI with a very shallow understanding of how it works. Why?
It generates words one by one, like we all do. This doesn't mean it does just that and nothing else. It's the mechanics of how they are trained and how they do inference. And most importantly how they communicate with us. It doesn't define what they are or their limits. This is reductionism. Ignoring the mathematical complexity of a giant neural network.
Bjartr 2 hours ago [-]
> like we all do
Do we though? Sure, we communicate sequentially, but that doesn't mean that our internal effort is piecewise and linear. A modern transformer LLM however is. Each token is sampled from a population exclusively dependent on the tokens that came before it.
Mechanistically speaking, it works similarly to autocomplete, but at a very different scale.
Now how much of an unavoidable handicap this incurs, if any, is absolutely up for debate.
But yes, taking this mechanistic truth and only considering it in a shallow manner underestimates the capability of LLMs by a large degree.
kenjackson 1 hours ago [-]
Our thinking is also based only on events that occurred previously in time. We don’t use events in the future.
ElevenLathe 1 hours ago [-]
Is this a certainty? I thought it was an open question whether quantum effects are at play in the brain, and those have a counterintuitive relationship with time (to vastly dumb things down in a way my grug mind can comprehend).
kenjackson 57 minutes ago [-]
Well there’s no evidence of this that I’ve seen. If so, then maybe that is what is the blocker for AGI.
karmakaze 2 hours ago [-]
I can't say for certain that our wetware isn't "just a very good autocomplete".
esafak 25 minutes ago [-]
A very good autocomplete is realized by developing an understanding.
2 hours ago [-]
faangguyindia 46 minutes ago [-]
>Pattern matching questions on a contrived test is not the same thing as understanding or reasoning.
I think most of the problem i solve is also a pattern matching. The problems i am good at solving are the ones i've seen before or the ones i can break into problems i've seen before.
ACCount37 5 hours ago [-]
"Not understanding or reasoning" is anthropocentric cope. There is very little practical difference between "understanding" and "reasoning" implemented in human mind and that implemented in LLMs.
One notable difference, however, is that LLMs disproportionately suck at spatial reasoning. Which shouldn't be surprising, considering that their training datasets are almost entirely text. The ultimate wordcel makes for a poor shape rotator.
All ARC-AGI tasks are "spatial reasoning" tasks. They aren't in any way special. They just force LLMs to perform in an area they're spectacularly weak at. And LLMs aren't good enough yet to be able to brute force through this innate deficiency with raw intelligence.
dwallin 55 minutes ago [-]
Very much agree with this. Looking at the dimensionality of a given problem space is a very helpful heuristic when analyzing how likely an llm is going to be suitable/reliable for that task. Consider how important positional encodings are LLM performance. You also then have an attention model that operates in that 1-dimensional space. With multidimensional data significant transformations to encode into a higher dimensional abstraction needs to happen within the model itself, before the model can even attempt to intelligently manipulate it.
HighGoldstein 4 hours ago [-]
> There is very little practical difference between "understanding" and "reasoning" implemented in human mind and that implemented in LLMs.
Source?
ACCount37 4 hours ago [-]
The primary source is: measured LLM performance on once-human-exclusive tasks - such as high end natural language processing or commonsense reasoning.
Those things were once thought to require a human mind - clearly, not anymore. Human commonsense knowledge can be both captured and applied by a learning algorithm trained on nothing but a boatload of text.
But another important source is: loads and loads of mech interpret research that tried to actually pry the black box open and see what happens on the inside.
This found some amusing artifacts - such as latent world models that can be extracted from the hidden state, or neural circuits corresponding to high level abstracts being chained together to obtain the final outputs. Very similar to human "abstract thinking" in function - despite being implemented on a substrate of floating point math and not wet meat.
NooneAtAll3 4 hours ago [-]
...literally benchmarks the post is all about?
practical difference is about results - and results are here
pessimizer 1 hours ago [-]
> Pattern matching questions on a contrived test is not the same thing as understanding or reasoning.
Pattern matching is definitely the same thing as understanding and reasoning.
The problem is that LLMs can't recognize patterns that are longer than a few paragraphs, because the tokens would have to be far too long. LLMs are a thing we are lucky to have because we have very fast computers and very smart mathematicians making very hard calculations very efficient and parallelizable. But they sit on top of a bed of an enormous amount of human written knowledge, and can only stretch so far from that bed before completely falling apart.
Humans don't use tokenizers.
The goal right now is to build a scaffolding of these dummies in order to get really complicated work done, but that work is only ever going to accidentally be correct because of an accumulation of errors. This may be enough for a lot if we try it 1000x and run manually-tuned algos over the output to find the good ones. But this is essentially manual work, done in the traditional way.
edit: sorry, you're never going to convince me these things are geniuses when I chat to them for a couple of back and forth exchanges and they're already obviously losing track of everything, even what they just said. The good thing is that what they are is enough to do a lot, if you're a person who can be satisfied that they're not going to be your god anytime soon.
wiseowise 7 hours ago [-]
[flagged]
bloqs 6 hours ago [-]
please consider a less emotive, flaming/personal tone in the future, hacker news is much more readable without it!
I would broadly agree that it's a bit far, but the OPs point does have some validity, its often the same formulaic methodology
modeless 8 hours ago [-]
I've been testing LLMs on Sokoban-like puzzles (in the style of ARC-AGI-3) and they are completely awful at them. It really highlights how poor their memory is. They can't remember abstract concepts or rules between steps, even if they discover them themselves. They can only be presented with lossy text descriptions of such things which they have to re-read and re-interpret at every step.
LLMs are completely helpless on agentic tasks without a ton of scaffolding. But the scaffolding is inflexible and brittle, unlike the models themselves. Whoever figures out how to reproduce the functions of this type of scaffolding within the models, with some kind of internal test-time-learned memory mechanism, is going to win.
low_tech_love 2 hours ago [-]
Try to get your LLM of choice to find its way out of a labyrinth that you describe in text form. It's absolutely awful even with the simplest mazes. I'm not sure the problem here is memory, though? I think it has to do with spatial reasoning. I'd be willing to bet every company right now is working on spatial reasoning (at least up to 3D) and as soon as that is working, a huge amount of pieces will fall into place.
sunrunner 5 hours ago [-]
I'm not sure how similar this is but I tried the same quite a while back with a simple 5x5 nonogram (Picross) and had similar difficulties.
I found not only incorrect 'reasoning' but also even after being explicit about why a certain deduction was not correct the same incorrect deduction would then appear later, and this happened over and over.
Also, there's already a complete database of valid answers at [1], so I'm not sure why the correct answer couldn't just come from that, and the 'reasoning' can be 'We solved this here, look...' ;)
> I found not only incorrect 'reasoning' but also even after being explicit about why a certain deduction was not correct the same incorrect deduction would then appear later, and this happened over and over.
Because its in the context window and a lot of training material refers to earlier stuff for later stuff it is trained to bring up that stuff again and again. Even if it is in the window as a negative.
M4v3R 8 hours ago [-]
I wonder scaffolding synthesis is the way to go. Namely the LLM itself first reasons about the problem and creates scaffolding for a second agent that will do the actual solving. All inside a feedback loop to adjust the scaffolding based on results.
modeless 8 hours ago [-]
In general I think the more of the scaffolding that can be folded into the model, the better. The model should learn problem solving strategies like this and be able to manage them internally.
sixo 7 hours ago [-]
I toyed around with the idea of using an LLM to "compile" user instructions into a kind of AST of scaffolding, which can then be run by another LLM. It worked fairly wellbfor the kind of semi-structured tasks LLMs choke on like "for each of 100 things, do...", but I haven't taken it beyond a minimal impl.
harshitaneja 7 hours ago [-]
I am working on something similar but with an AST for legal documents. So far, it seems promising but still rudimentary.
plantain 7 hours ago [-]
If you've ever used Claude Code + Plan mode - you know that exactly this is true.
jvanderbot 21 minutes ago [-]
Can someone explain to me why a new LLMs ability to solve highly publicized puzzles is not "just" (sorry) it having access to the blog posts talking about those puzzles?
It's fine, that's what I would do to solve them, but it doesn't obviously and immediately make me confident in new reasoning capability w that suspicion floating around.
albertzeyer 7 hours ago [-]
This sounds interesting.
I would really like to read a full research paper made out of this, which describes the method in more detail, gives some more examples, does more analysis on it, etc.
Btw, this uses LLMs on pure text-level? Why not images? Most of these patterns are easy to detect on image-level, but I assume when presented as text, it's much harder.
> LLMs are PhD-level reasoners in math and science, yet they fail at children's puzzles. How is this possible?
I think this argument is a bit flawed. Yes, you can define AGI as being better than (average) humans in every possible task. But isn't this very arbitrary? Isn't it more reasonable to expect that different intelligent systems (including animals, humans) can have different strengths, and it is unreasonable to expect that one system is really better in everything? Maybe it's more reasonable to define ASI that way, but even for ASI, if a system is already better in a majority of tasks (but not necessarily in every task), I think this should already count as ASI. Maybe really being better in every possible task is just not possible. You could design a task that is very specifically tailored for human intelligence.
bubblyworld 6 hours ago [-]
I suspect (to use the language of the author) current LLMs have a bit of a "reasoning dead zone" when it comes to images. In my limited experience they struggle with anything more complex than "transcribe the text" or similarly basic tasks. Like I tried to create an automated QA agent with Claude Sonnet 3.5 to catch regressions in my frontend, and it will look at an obviously broken frontend component (using puppeteer to drive and screenshot a headless browser) and confidently proclaim it's working correctly, often making up a supporting argument too. I've had much more success passing the code for the component and any console logs directly to the agent in text form.
My memory is a bit fuzzy, but I've seen another QA agent that takes a similar approach of structured text extraction rather than using images. So I suspect I'm not the only one finding image-based reasoning an issue. Could also be for cost reasons though, so take that with a pinch of salt.
ACCount37 3 hours ago [-]
LLM image frontends suck, and a lot of them suck big time.
The naive approach of "use a pretrained encoder to massage the input pixels into a bag of soft tokens and paste those tokens into the context window" is good enough to get you a third of the way to humanlike vision performance - but struggles to go much further.
Claude's current vision implementation is also notoriously awful. Like, "a goddamn 4B Gemma 3 beats it" level of awful. For a lot of vision-heavy tasks, you'd be better off using literally anything else.
Davidzheng 8 hours ago [-]
Actually really promising stuff. I think a lot of the recent advances in the last 6mo - 1yr is in the other loop (for ex. the google deepthink model which got IMO gold and the OAI IMO gold all use substantive other loop search strategies [though it's unclear what these are] to maybe parallelize some generation/verification process). So there's no reason why we can't have huge advances in this area even outside of the industry labs in my view (I'm uninformed in general so take this comment with a large grain of salt).
d_burfoot 2 hours ago [-]
To me the reason ARC-AGI puzzles are difficult for LLMs and possible for humans is that they are expressed in a format for which humans have powerful preprocessing capabilities.
Imagine the puzzle layouts were expressed in JSON instead of as a pattern of visual blocks. How many humans could solve them in that case?
jononor 1 hours ago [-]
We have powerful preprocessing blocks for images: Strong computer vision capabilities predates LLMs by several years. Image classification, segmentation, object detection, etc. All differential and trainable in same way as LLMs, including jointly.
To the best of my knowledge, no team has shown really high scores by adding in a image preprocessing block?
kenjackson 1 hours ago [-]
Bingo. We simply made a test for which we are well trained. We are constantly making real time decisions with our eyes. Interestingly certain monkeys are much better at certain visual pattern recognition than we are. They might laugh and think humans haven’t reached AGI yet.
pessimizer 1 hours ago [-]
Every one who had access to a computer that could convert json into something more readable for humans, and would know that was the first thing they needed to do?
You might as well have asked how many English speakers could solve the questions if they were in Chinese. All of them. They would call up someone who spoke Chinese, pay them to translate the questions, then solve them. Or failing that, they would go to the bookstore, buy books on learning Chinese, and solve them three years from now.
Garlef 3 hours ago [-]
That's a super neat approach.
But the core issue seems to be: How do you come up with the fitness function that drives the evolutionary process without human intervention in the first place?
(I've tried something similar with a coding agent where I let the agent modify parts of its system prompt... But it got stuck very fast since there was no clear fitness function)
wiz21c 5 hours ago [-]
isn't the author actually overfitting a solution ? He'll sure beat ARC AGI, but that will be all.
deyiao 4 hours ago [-]
I don't think so. The author isn't training an LLM, but rather using an LLM to solve a specific problem. This method could also be applied to solve other problems.
cahaya 2 hours ago [-]
Are there any existing scripts/ tools to use these evolutionary algorithms also at home with e.g. Codex/GPT-5 / Claude Code?
justatdotin 6 hours ago [-]
> LLMs have "dead reasoning zones" — areas in their weights where logic doesn't work. Humans have dead knowledge zones (things we don't know), but not dead reasoning zones.
blank stare
mjburgess 5 hours ago [-]
We have dead-zones in adductive reasoning, not in induction or deduction. Almost all failures of reasoning in people are in abducing what model describes the situation at hand.
eg., we can apply the rule, "-A cannot follow from A", etc. regardless of the A
eg., we always know that if the number of apples is 2, then it cannot be any of "all numbers without 2" -- which quantifies over all numbers
You will not find a "gap" for a given number, whereas with LLMs, gaps of this kind are common
rel_ic 3 hours ago [-]
> we can apply the rule, "-A cannot follow from A", etc. regardless of the A
You can't think of any domains where we are unable to apply this rule? I feel like I'm surrounded by people claiming "A, therefore -A!!"
And if I'm one of them, and this were a reasoning dead-zone for me, I wouldn't be able to tell!
mjburgess 3 hours ago [-]
That's an abductive failure to recognise that something is A, and something else is not-A
I dont see cases where people recognise the contradiction and then perform it.
rel_ic 1 hours ago [-]
People who know alcohol is bad for them and don't want to keep being drunks but keep drinking, people who believe phones are bad for their kids but still buy them, people who understand AI will significantly degrade the environment if it becomes ubiquitous but still work to help it become ubiquitous...
Mathematicians who publish proofs that are later proven inconsistent!
I suspect we have fundamentally different views of how humans work. I see our behavior and beliefs as _mostly_ irrational, with only a few "reasoning live-zones" where, with great effort, we can achieve logical thought.
virgilp 2 hours ago [-]
How can you know? One could argue that the entire phenomenon of cognitive dissonance is "people (internally) recognize the contradiction and then perform it"
didroe 5 hours ago [-]
>With RL, models no longer just learn what sounds correct based on patterns they've seen. They learn what words to output to be correct. RL is the process of forcing the pre-trained weights to be logically consistent.
How does Reinforcement Learning force the weights to be logically consistent? Isn't it just about training using a coarser/more-fuzzy granularity of fitness?
More generally, is it really solving the task if it's given a large number of attempts and an oracle to say whether it's correct? Humans can answer the questions in one shot and self-check the answer, whereas this is like trial and error with an external expert who tells you to try again.
This sounds like it is just slightly smarter than brute forcing your way to a solution.
Oh well, more support for my prediction: nobody will win a Nobel prize for reaching AGI.
jokoon 7 hours ago [-]
Those are bold claims
FergusArgyll 2 hours ago [-]
The biggest issue I have with ARC-AGI is it's a visual problem. LLMs (even the newfangled multi-modal ones) are still far worse at vision than at purely text based problems. I don't think it's possible to build a test of purely text-based questions that would be easy for humans and hard for SOTA models. Yes, there's a few gotchas you can throw at them but not 500.
pilooch 10 hours ago [-]
Congrats, this solution resembles AlphaEvolve. Text serves as the high-level search space, and genetic mixing (map-elites in AE) merges attemps at lower levels.
doctorpangloss 9 hours ago [-]
you would be interested in dSPY
imiric 5 hours ago [-]
Congrats, you made LLMs perform slightly better at a contrived puzzle. This finally proves that we've cracked intelligence and are well on our way towards AGI.
Because they are not.
Pattern matching questions on a contrived test is not the same thing as understanding or reasoning.
It’s the same reason why most of the people who pass your leetcode tests don’t actually know how to build anything real. They are taught to the test not taught to reality.
Do submarines swim? I don't really care if it gets me where I want to go. The fact is that just two days ago, I asked Claude to look at some reasonably complicated concurrent code to which I had added a new feature, and asked it to list what tests needed to be added; and then when I asked GPT-5 to add them, it one-shot nailed the implementations. I've written a gist of it here:
https://gitlab.com/-/snippets/4889253
Seriously just even read the description of the test it's trying to write.
In order to one-shot that code, it had to understand:
- How the cache was supposed to work
- How conceptually to set up the scenario described
- How to assemble golang's concurrency primitives (channels, goroutines, and waitgroups), in the correct order, to achieve the goal.
Did it have a library of concurrency testing patterns in its head? Probably -- so do I. Had it ever seen my exact package before in its training? Never.
I just don't see how you can argue with a straight face that this is "pattern matching". If that's pattern matching, then pattern matching is not an insult.
If anything, the examples in this article are the opposite. Take the second example, which is basically 'assemble these assorted pieces into a rectangle'. Nearly every adult has assembled a minimum of dozens of things in their lives; many have assembled thousands of things. So it's humans in this case who are simply "pattern matching questions on a contrived test", and the LLMs, which almost certainly didn't have a lot of "assemble these items" in their training data, that are reasoning out what's going on from first principles.
It doesn't matter HOW LLMs "swim" as long as they can, but the point being raised is whether they actually can.
It's as if LLMs can swim in the ocean, in rough surf, but fail to swim in rivers or swimming pools, because they don't have a generalized ability to swim - they've just been RL-trained on the solution steps to swimming in surf, but since those exact conditions don't exist in a river (which might seem like a less challenging environment), they fail there.
So, the question that might be asked is when LLMs are trained to perform well in these vertical domains like math and programming, where it's easy to verify results and provide outcome- or process-based RL rewards, are they really learning to reason, or are they just learning to pattern match to steer generation in the direction of problem-specific reasoning steps that they had been trained on?
Does the LLM have the capability to reason/swim, or is it really just an expert system that has been given the rules to reason/swim in certain cases, but would need to be similarly hand fed the reasoning steps to be successful in other cases?
I think the answer is pretty obvious given that LLM's can't learn at runtime - can't try out some reasoning generalization they may have arrived at, find that it doesn't work in a specific case, then explore the problem and figure it out for next time.
Given that it's Demis Hassabis who it pointing out this deficiency of LLMs (and has a 5-10 year plan/timeline to fix it - AGI), not some ill-informed LLM critic, it seems silly to deny it.
This is just a problem of memory. Supposing that an LLM did generate a genuinely novel insight, it could in theory they could write a note for itself so that next time they come online, they can read through a summary of the things they learned. And it could also write synthetic training data for itself so that the next time they're trained, that gets incorporated into its general knowledge.
OpenAI allows you to fine-tune GPT models, I believe. You could imagine a GPT system working for 8 hours in a day, then spending a bunch of time looking over all its conversation looking for patterns or insights or things to learn, and then modifying its own fine-tuning data (adding, removing, or modifying as appropriate), which it then used to train itself overnight, waking up the next morning having synthesized the previous day's experience.
Are you sure there's a real difference? Do you have a definition of "reasoning" that excludes this?
Certainly if a human (or some AI) has learned to predict/reason over some domain, then what they will be doing it pattern matching to determine the generalizations and exceptions that apply in a given context (including a hypothetical context in a what-if reasoning chain), in order to be able to select a next step that worked before.
However, I think what we're really talking about here isn't the mechanics of applying learnt reasoning (context pattern matching), but rather the ability to reason in the general case, which requires the ability to LEARN to solve novel problems, which is what is missing from LLMs.
A system that has a fixed set of (reasoning/prediction) rules, but can't learn new ones for itself, seems better regarded as an expert system. We need to make the distinction between can a system that can only apply rules, and one that can actually figure out the rules in the first place.
In terms of my definitions of intelligence and reasoning, based around ability to use past experience to learn to predict, then any system that can't learn from fresh experience doesn't meet that definition.
Of course in humans and other intelligent animals the distinction between past and ongoing experience doesn't apply since they can learn continually and incrementally (something that is lacking from LLMs), so for AI we need to use a different vocabulary, and "expert system" seems the obvious label for something that can use rules, but not discover them for itself.
1. Taking established techniques or concepts and appropriately applying them to novel situations.
2. Inventing or synthesizing new, never-before-seen techniques or concepts
The vast majority of the time, humans do #1. LLMs certainly do this in some contexts as well, as demonstrated by my example above. This to me counts as "understanding" and "thinking". Some people define "understanding" such that it's something only humans can do; to which I respond, I don't care what you call it, it's useful.
Can LLMs do #2? I don't know. They've got such extensive experience that how would you know if they'd invented a technique vs had seen it somewhere?
But I'd venture to argue that most humans never or rarely do #2.
I don't think this assumption is sound. Humans write a huge amount on "assemble components x and y to make entity z". I'd expect all LLMs to have consumed every IKEA type instruction manual, the rules for Jenga, all geometry textbooks and papers ever written.
Even high school maths tests are way harder than what most professional programmers do on a daily basis.
IMO its still "just" a, very good, autocomplete. No actual reasoning, but lots of statistics on what is the next token to spit out.
That's the main point of the parent comment. Arguing about the definition of "reasoning" or "pattern matching" is just a waste of time. What really matters is if it produces helpful output. Arguing about that is way better!
Instead of saying: "It's just pattern matching -> It won't improve the world", make an argument like: "AI's seem to have trouble specializing like humans -> adopting AI will increase error rates in business processes -> due to the amount of possible edge cases, most people will get into an edge case with no hope of escaping it -> many people's lives will get worse".
The first example relies on us agreeing on the definition of pattern matching, and then taking a conclusion based on how those words feel. This has no hope of convincing me if I don't like your definition! The second one is an argument that could potentially convince me, even if I'm an AI optimist. It is also just by itself an interesting line of reasoning.
It generates words one by one, like we all do. This doesn't mean it does just that and nothing else. It's the mechanics of how they are trained and how they do inference. And most importantly how they communicate with us. It doesn't define what they are or their limits. This is reductionism. Ignoring the mathematical complexity of a giant neural network.
Do we though? Sure, we communicate sequentially, but that doesn't mean that our internal effort is piecewise and linear. A modern transformer LLM however is. Each token is sampled from a population exclusively dependent on the tokens that came before it.
Mechanistically speaking, it works similarly to autocomplete, but at a very different scale.
Now how much of an unavoidable handicap this incurs, if any, is absolutely up for debate.
But yes, taking this mechanistic truth and only considering it in a shallow manner underestimates the capability of LLMs by a large degree.
I think most of the problem i solve is also a pattern matching. The problems i am good at solving are the ones i've seen before or the ones i can break into problems i've seen before.
One notable difference, however, is that LLMs disproportionately suck at spatial reasoning. Which shouldn't be surprising, considering that their training datasets are almost entirely text. The ultimate wordcel makes for a poor shape rotator.
All ARC-AGI tasks are "spatial reasoning" tasks. They aren't in any way special. They just force LLMs to perform in an area they're spectacularly weak at. And LLMs aren't good enough yet to be able to brute force through this innate deficiency with raw intelligence.
Source?
Those things were once thought to require a human mind - clearly, not anymore. Human commonsense knowledge can be both captured and applied by a learning algorithm trained on nothing but a boatload of text.
But another important source is: loads and loads of mech interpret research that tried to actually pry the black box open and see what happens on the inside.
This found some amusing artifacts - such as latent world models that can be extracted from the hidden state, or neural circuits corresponding to high level abstracts being chained together to obtain the final outputs. Very similar to human "abstract thinking" in function - despite being implemented on a substrate of floating point math and not wet meat.
practical difference is about results - and results are here
Pattern matching is definitely the same thing as understanding and reasoning.
The problem is that LLMs can't recognize patterns that are longer than a few paragraphs, because the tokens would have to be far too long. LLMs are a thing we are lucky to have because we have very fast computers and very smart mathematicians making very hard calculations very efficient and parallelizable. But they sit on top of a bed of an enormous amount of human written knowledge, and can only stretch so far from that bed before completely falling apart.
Humans don't use tokenizers.
The goal right now is to build a scaffolding of these dummies in order to get really complicated work done, but that work is only ever going to accidentally be correct because of an accumulation of errors. This may be enough for a lot if we try it 1000x and run manually-tuned algos over the output to find the good ones. But this is essentially manual work, done in the traditional way.
edit: sorry, you're never going to convince me these things are geniuses when I chat to them for a couple of back and forth exchanges and they're already obviously losing track of everything, even what they just said. The good thing is that what they are is enough to do a lot, if you're a person who can be satisfied that they're not going to be your god anytime soon.
I would broadly agree that it's a bit far, but the OPs point does have some validity, its often the same formulaic methodology
LLMs are completely helpless on agentic tasks without a ton of scaffolding. But the scaffolding is inflexible and brittle, unlike the models themselves. Whoever figures out how to reproduce the functions of this type of scaffolding within the models, with some kind of internal test-time-learned memory mechanism, is going to win.
I found not only incorrect 'reasoning' but also even after being explicit about why a certain deduction was not correct the same incorrect deduction would then appear later, and this happened over and over.
Also, there's already a complete database of valid answers at [1], so I'm not sure why the correct answer couldn't just come from that, and the 'reasoning' can be 'We solved this here, look...' ;)
[1] The wonderful https://pixelogic.app/every-5x5-nonogram
Because its in the context window and a lot of training material refers to earlier stuff for later stuff it is trained to bring up that stuff again and again. Even if it is in the window as a negative.
It's fine, that's what I would do to solve them, but it doesn't obviously and immediately make me confident in new reasoning capability w that suspicion floating around.
I would really like to read a full research paper made out of this, which describes the method in more detail, gives some more examples, does more analysis on it, etc.
Btw, this uses LLMs on pure text-level? Why not images? Most of these patterns are easy to detect on image-level, but I assume when presented as text, it's much harder.
> LLMs are PhD-level reasoners in math and science, yet they fail at children's puzzles. How is this possible?
I think this argument is a bit flawed. Yes, you can define AGI as being better than (average) humans in every possible task. But isn't this very arbitrary? Isn't it more reasonable to expect that different intelligent systems (including animals, humans) can have different strengths, and it is unreasonable to expect that one system is really better in everything? Maybe it's more reasonable to define ASI that way, but even for ASI, if a system is already better in a majority of tasks (but not necessarily in every task), I think this should already count as ASI. Maybe really being better in every possible task is just not possible. You could design a task that is very specifically tailored for human intelligence.
My memory is a bit fuzzy, but I've seen another QA agent that takes a similar approach of structured text extraction rather than using images. So I suspect I'm not the only one finding image-based reasoning an issue. Could also be for cost reasons though, so take that with a pinch of salt.
The naive approach of "use a pretrained encoder to massage the input pixels into a bag of soft tokens and paste those tokens into the context window" is good enough to get you a third of the way to humanlike vision performance - but struggles to go much further.
Claude's current vision implementation is also notoriously awful. Like, "a goddamn 4B Gemma 3 beats it" level of awful. For a lot of vision-heavy tasks, you'd be better off using literally anything else.
Imagine the puzzle layouts were expressed in JSON instead of as a pattern of visual blocks. How many humans could solve them in that case?
You might as well have asked how many English speakers could solve the questions if they were in Chinese. All of them. They would call up someone who spoke Chinese, pay them to translate the questions, then solve them. Or failing that, they would go to the bookstore, buy books on learning Chinese, and solve them three years from now.
But the core issue seems to be: How do you come up with the fitness function that drives the evolutionary process without human intervention in the first place?
(I've tried something similar with a coding agent where I let the agent modify parts of its system prompt... But it got stuck very fast since there was no clear fitness function)
blank stare
eg., we can apply the rule, "-A cannot follow from A", etc. regardless of the A
eg., we always know that if the number of apples is 2, then it cannot be any of "all numbers without 2" -- which quantifies over all numbers
You will not find a "gap" for a given number, whereas with LLMs, gaps of this kind are common
You can't think of any domains where we are unable to apply this rule? I feel like I'm surrounded by people claiming "A, therefore -A!!"
And if I'm one of them, and this were a reasoning dead-zone for me, I wouldn't be able to tell!
I dont see cases where people recognise the contradiction and then perform it.
Mathematicians who publish proofs that are later proven inconsistent!
I suspect we have fundamentally different views of how humans work. I see our behavior and beliefs as _mostly_ irrational, with only a few "reasoning live-zones" where, with great effort, we can achieve logical thought.
How does Reinforcement Learning force the weights to be logically consistent? Isn't it just about training using a coarser/more-fuzzy granularity of fitness?
More generally, is it really solving the task if it's given a large number of attempts and an oracle to say whether it's correct? Humans can answer the questions in one shot and self-check the answer, whereas this is like trial and error with an external expert who tells you to try again.
Kaggle: https://www.kaggle.com/code/jerber/jeremy-arc2
Oh well, more support for my prediction: nobody will win a Nobel prize for reaching AGI.