Apple just proved AI "reasoning" models like Claude, DeepSeek-R1, and o3-mini don't actually reason at all. They just memorize patterns really well.
Technology
347
Beiträge
149
Kommentatoren
17
Aufrufe
-
No, it shows how certain people misunderstand the meaning of the word.
You have called npcs in video games "AI" for a decade, yet you were never implying they were somehow intelligent. The whole argument is strangely inconsistent.
Strangely inconsistent + smoke & mirrors = profit!
-
But for something like solving a Towers of Hanoi puzzle, which is what this study is about, we're not looking for emotional judgements - we're trying to evaluate the logical reasoning capabilities. A sociopath would be equally capable of solving logic puzzles compared to a non-sociopath. In fact, simple computer programs do a great job of solving these puzzles, and they certainly have nothing like emotions. So I'm not sure that emotions have much relevance to the topic of AI or human reasoning and problem solving, at least not this particular aspect of it.
As for analogizing LLMs to sociopaths, I think that's a bit odd too. The reason why we (stereotypically) find sociopathy concerning is that a person has their own desires which, in combination with a disinterest in others' feelings, incentivizes them to be deceitful or harmful in some scenarios. But LLMs are largely designed specifically as servile, having no will or desires of their own. If people find it concerning that LLMs imitate emotions, then I think we're giving them far too much credit as sentient autonomous beings - and this is coming from someone who thinks they think in the same way we do! The think like we do, IMO, but they lack a lot of the other subsystems that are necessary for an entity to function in a way that can be considered as autonomous/having free will/desires of its own choosing, etc.
In fact, simple computer programs do a great job of solving these puzzles.....
If an AI is trained to do this, it will be very good, like for example when a GPT-2 was trained to multiply numbers up to 20 digits.
https://nitter.net/yuntiandeng/status/1836114419480166585#m
Here they do the same test to GPT-4o, o1-mini and o3-mini
-
Why would they "prove" something that's completely obvious?
The burden of proof is on the grifters who have overwhelmingly been making false claims and distorting language for decades.
Not when large swaths of people are being told to use it everyday. Upper management has bought in on it.
-
LOOK MAA I AM ON FRONT PAGE
-
LOOK MAA I AM ON FRONT PAGE
Yah of course they do they’re computers
-
Why would they "prove" something that's completely obvious?
The burden of proof is on the grifters who have overwhelmingly been making false claims and distorting language for decades.
Why would they "prove" something that's completely obvious?
I don’t want to be critical, but I think if you step back a bit and look and what you’re saying, you’re asking why we would bother to experiment and prove what we think we know.
That’s a perfectly normal and reasonable scientific pursuit. Yes, in a rational society the burden of proof would be on the grifters, but that’s never how it actually works. It’s always the doctors disproving the cure-all, not the snake oil salesmen failing to prove their own prove their own product.
There is value in this research, even if it fits what you already believe on the subject. I would think you would be thrilled to have your hypothesis confirmed.
-
Yah of course they do they’re computers
That's not really a valid argument for why, but yes the models which use training data to assemble statistical models are all bullshitting. TBH idk how people can convince themselves otherwise.
-
Like what?
I don’t think there’s any search engine better than Perplexity. And for scientific research Consensus is miles ahead.
Through the years I've bounced between different engines. I gave Bing a decent go some years back, mostly because I was interested in gauging the performance and wanted to just pit something against Google. After that I've swapped between Qwant and Startpage a bunch. I'm a big fan of Startpage's "Anonymous view" function.
Since then I've landed on Kagi, which I've used for almost a year now. It's the first search engine I've used that you can make work for you. I use the lens feature to focus on specific tasks, and de-prioritise pages that annoy me, sometimes outright omitting results from sites I find useless or unserious. For example when I'm doing web stuff and need to reference the MDN, I don't really care for w3schools polluting my results.
I'm a big fan of using my own agency and making my own decisions, and the recent trend in making LLMs think for us is something I find rather worrying, it allows for a much subtler manipulation than what Google does with its rankings and sponsor inserts.
Perplexity openly talking about wanting to buy Chrome and harvesting basically all the private data is also terrifying, thus I wouldn't touch that service with a stick. That said, I appreciate their candour, somehow being open about being evil is a lot more palatable to me than all these companies pretending to be good.
-
If emissions dropped to 0 tonight, we would be substantially better off than if we maintain our current trajectory. Doomerism helps nobody.
It’s not doomerism it’s just realistic. Deluding yourself won’t change that.
-
If the situation gets dire, it's likely that the weather will be manipulated. Countries would then have to be convinced not to use this for military purposes.
This isn’t a thing.
-
That's not really a valid argument for why, but yes the models which use training data to assemble statistical models are all bullshitting. TBH idk how people can convince themselves otherwise.
I think because it's language.
There's a famous quote from Charles Babbage when he presented his difference engine (gear based calculator) and someone asking "if you put in the wrong figures, will the correct ones be output" and Babbage not understanding how someone can so thoroughly misunderstand that the machine is, just a machine.
People are people, the main thing that's changed since the Cuneiform copper customer complaint is our materials science and networking ability. Most things that people interact with every day, most people just assume work like it appears to on the surface.
And nothing other than a person can do math problems or talk back to you. So people assume that means intelligence.
-
I think because it's language.
There's a famous quote from Charles Babbage when he presented his difference engine (gear based calculator) and someone asking "if you put in the wrong figures, will the correct ones be output" and Babbage not understanding how someone can so thoroughly misunderstand that the machine is, just a machine.
People are people, the main thing that's changed since the Cuneiform copper customer complaint is our materials science and networking ability. Most things that people interact with every day, most people just assume work like it appears to on the surface.
And nothing other than a person can do math problems or talk back to you. So people assume that means intelligence.
I often feel like I'm surrounded by idiots, but even I can't begin to imagine what it must have felt like to be Charles Babbage explaining computers to people in 1840.
-
That's not really a valid argument for why, but yes the models which use training data to assemble statistical models are all bullshitting. TBH idk how people can convince themselves otherwise.
TBH idk how people can convince themselves otherwise.
They don’t convince themselves. They’re convinced by the multi billion dollar corporations pouring unholy amounts of money into not only the development of AI, but its marketing. Marketing designed to not only convince them that AI is something it’s not, but also that that anyone who says otherwise (like you) are just luddites who are going to be “left behind”.
-
"It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." -Pamela McCorduck´.
It's called the AI Effect.As Larry Tesler puts it, "AI is whatever hasn't been done yet.".
Yesterday I asked an LLM "how much energy is stored in a grand piano?" It responded with saying there is no energy stored in a grad piano because it doesn't have a battery.
Any reasoning human would have understood that question to be referring to the tension in the strings.
Another example is asking "does lime cause kidney stones?". It didn't assume I mean lime the mineral and went with lime the citrus fruit instead.
Once again a reasoning human would assume the question is about the mineral.
Ask these questions again in a slightly different way and you might get a correct answer, but it won't be because the LLM was thinking.
-
LOOK MAA I AM ON FRONT PAGE
What a dumb title. I proved it by asking a series of questions. It’s not AI, stop calling it AI, it’s a dumb af language model. Can you get a ton of help from it, as a tool? Yes! Can it reason? NO! It never could and for the foreseeable future, it will not.
It’s phenomenal at patterns, much much better than us meat peeps. That’s why they’re accurate as hell when it comes to analyzing medical scans.
-
Yesterday I asked an LLM "how much energy is stored in a grand piano?" It responded with saying there is no energy stored in a grad piano because it doesn't have a battery.
Any reasoning human would have understood that question to be referring to the tension in the strings.
Another example is asking "does lime cause kidney stones?". It didn't assume I mean lime the mineral and went with lime the citrus fruit instead.
Once again a reasoning human would assume the question is about the mineral.
Ask these questions again in a slightly different way and you might get a correct answer, but it won't be because the LLM was thinking.
But 90% of "reasoning humans" would answer just the same. Your questions are based on some non-trivial knowledge of physics, chemistry and medicine that most people do not possess.
-
LOOK MAA I AM ON FRONT PAGE
You assume humans do the opposite? We literally institutionalize humans who not follow set patterns.
-
Unlike Markov models, modern LLMs use transformers that attend to full contexts, enabling them to simulate structured, multi-step reasoning (albeit imperfectly). While they don’t initiate reasoning like humans, they can generate and refine internal chains of thought when prompted, and emerging frameworks (like ReAct or Toolformer) allow them to update working memory via external tools. Reasoning is limited, but not physically impossible, it’s evolving beyond simple pattern-matching toward more dynamic and compositional processing.
previous input goes in. Completely static, prebuilt model processes it and comes up with a probability distribution.
There is no "unlike markov chains". They are markov chains. Ones with a long context (a markov chain also kakes use of all the context provided to it, so I don't know what you're on about there). LLMs are just a (very) lossy compression scheme for the state transition table. Computed once, applied blindly to any context fed in.
-
previous input goes in. Completely static, prebuilt model processes it and comes up with a probability distribution.
There is no "unlike markov chains". They are markov chains. Ones with a long context (a markov chain also kakes use of all the context provided to it, so I don't know what you're on about there). LLMs are just a (very) lossy compression scheme for the state transition table. Computed once, applied blindly to any context fed in.
LLMs are not Markov chains, even extended ones. A Markov model, by definition, relies on a fixed-order history and treats transitions as independent of deeper structure. LLMs use transformer attention mechanisms that dynamically weigh relationships between all tokens in the input—not just recent ones. This enables global context modeling, hierarchical structure, and even emergent behaviors like in-context learning. Markov models can't reweight context dynamically or condition on abstract token relationships.
The idea that LLMs are "computed once" and then applied blindly ignores the fact that LLMs adapt their behavior based on input. They don’t change weights during inference, true—but they do adapt responses through soft prompting, chain-of-thought reasoning, or even emulated state machines via tokens alone. That’s a powerful form of contextual plasticity, not blind table lookup.
Calling them “lossy compressors of state transition tables” misses the fact that the “table” they’re compressing is not fixed—it’s context-sensitive and computed in real time using self-attention over high-dimensional embeddings. That’s not how Markov chains work, even with large windows.
-
LLMs are not Markov chains, even extended ones. A Markov model, by definition, relies on a fixed-order history and treats transitions as independent of deeper structure. LLMs use transformer attention mechanisms that dynamically weigh relationships between all tokens in the input—not just recent ones. This enables global context modeling, hierarchical structure, and even emergent behaviors like in-context learning. Markov models can't reweight context dynamically or condition on abstract token relationships.
The idea that LLMs are "computed once" and then applied blindly ignores the fact that LLMs adapt their behavior based on input. They don’t change weights during inference, true—but they do adapt responses through soft prompting, chain-of-thought reasoning, or even emulated state machines via tokens alone. That’s a powerful form of contextual plasticity, not blind table lookup.
Calling them “lossy compressors of state transition tables” misses the fact that the “table” they’re compressing is not fixed—it’s context-sensitive and computed in real time using self-attention over high-dimensional embeddings. That’s not how Markov chains work, even with large windows.
their input is the context window. Markov chains also use their whole context window. Llms are a novel implementation that can work with much longer contexts, but as soon as something slides out of its window, it's forgotten. just like any other markov chain. They don't adapt. You add their token to the context, slide the oldest one out and then you have a different context, on which you run the same thing again. A normal markov chain will also give you a different outuut if you give it a different context. Their biggest weakness is that they don't and can't adapt. You are confusing the encoding of the context with the model itself. Just to see how static the model is, try setting temperature to 0, and giving it the same context. i.e. only try to predict one token with the exact same context each time. As soon as you try to predict a 2nd token, you've just changed the input and ran the thing again. It's not adapting, you asked it something different, so it came up with a different answer
