(Shrug) It's a really nice introduction to manifolds, to the extent that's what it is. Really it's just an introduction to rational parameterization, which I'd never heard of but which at first glance seems extremely nifty.
It all gets pretty incomprehensible after that, but that's not the author's fault.
(Edit: interesting, thanks. So the underlying OS APIs that supply the power-consumption figures reported by asitop are just outright broken. The discrepancy is far too large to chalk up to static power losses or die-specific calibration factors that the video talks about.)
The difference between thinking and reasoning is that I can "think" that Elvis is still alive, Jewish space lasers are responsible for California wildfires, and Trump was re-elected president in 2020, but I cannot "reason" myself into those positions.
It ties into another aspect of these perennial threads, where it is somehow OK for humans to engage in deluded or hallucinatory thought, but when an AI model does it, it proves they don't "think."
They have to consider China. Right now Z-Image Turbo lets you render stills of any popular cartoon character you like, at frankly-disturbing levels of quality, doing almost anything you like. That's a relatively-tiny 6G-parameter model. If and when a WAN 2.2-level video model is released with a comparable lack of censorship, that will be the end of Disney's monopoly on pretty much any character IP.
Also, notice how Disney jumped all over Gemini's case before the ink was dry on the OpenAI partnership agreement. My guess is that Altman is just using Disney to attack his competitors, basically the 'two' part of a one-two punch that began by buying up a large portion of the world's RAM capacity for no valid business reason.
That's where these threads always end up. Someone asserts, almost violently, that AI does not and/or cannot "think." When asked how to falsify their assertion, perhaps by explaining what exactly is unique about the human brain that cannot and/or will not be possible to emulate, that's the last anyone ever hears from them. At least until the next "AI can't think" story gets posted.
The same arguments that appeared in 2015 inevitably get trotted out, almost verbatim, ten years later. It would be amusing on other sites, but it's just pathetic here.
Consider that you might have become polarized yourself. I often encounter good arguments against current AI systems emulating all essential aspects of human thinking. For example, the fact that they can't learn from few examples, that they can't perform simple mathematical operations without access to external help (via tool calling) or that they have to expend so much more energy to do their magic (and yes, to me they are a bit magical), which makes some wonder if what these models do is a form of refined brute-force search, rather than ideating.
Personally, I'm ok with reusing the word "thinking", but there are dogmatic stances on both sides. For example, lots of people decreeing that biology in the end can't but reduce to maths, since "what else could it be". The truth is we don't actually know if it is possible, for any conceivable computational system, to emulate all essential aspects of human thought. There are good arguments for this (in)possibility, like those presented by Roger Penrose in "the Emperor's new Mind" and "Shadows of the Mind".
For example, the fact that they can't learn from few examples
For one thing, yes, they can, obviously [1] -- when's the last time you checked? -- and for another, there are plenty of humans who seemingly cannot.
The only real difference is that with an LLM, when the context is lost, so is the learning. That will obviously need to be addressed at some point.
that they can't perform simple mathematical operations without access to external help (via tool calling)
But yet you are fine with humans requiring a calculator to perform similar tasks? Many humans are worse at basic arithmetic than an unaided transformer network. And, tellingly, we make the same kinds of errors.
or that they have to expend so much more energy to do their magic (and yes, to me they are a bit magical), which makes some wonder if what these models do is a form of refined brute-force search, rather than ideating.
Well, of course, all they are doing is searching and curve-fitting. To me, the magical thing is that they have shown us, more or less undeniably (Penrose notwithstanding), that that is all we do. Questions that have been asked for thousands of years have now been answered: there's nothing special about the human brain, except for the ability to form, consolidate, consult, and revise long-term memories.
That's post-training. The complaint I'm referring to is to the huge amounts of data (end energy) required during training - which is also a form of learning, after all. Sure, there are counter-arguments, for example pointing to the huge amount of non-textual data a child ingests, but these counter-arguments are not waterproof themselves (for example, one can point out that we are discussing text-only tasks). The discussion can go on and on, my point was only that cogent arguments are indeed often presented, which you were denying above.
> there are plenty of humans who seemingly cannot
This particular defense of LLMs has always puzzled me. By this measure, simply because there are sufficiently impaired humans, AGI has already been achieved many decades ago.
> But yet you are fine with humans requiring a calculator to perform similar tasks
I'm talking about tasks like multiplying two 4-digit numbers (let's say 8-digit, just to be safe, for reasoning models), which 5th or 6th graders in the US are expected to be able to do with no problem - without using a calculator.
> To me, the magical thing is that they have shown us, more or less undeniably (Penrose notwithstanding), that that is all we do.
Or, to put it more tersely, they have shown you that that is all we do. Penrose, myself, and lots of others don't see it quite like that. (Feeling quite comfortable being classed in the same camp with the greatest living physicist, honestly. ;) To me what LLMs do is approximate one aspect of our minds. But I have a strong hunch that the rabbit hole goes much deeper, your assessment notwithstanding.
No, it is not. Read the paper. They are discussing an emergent property of the context itself: "For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model."
I'm talking about tasks like multiplying two 4-digit numbers (let's say 8-digit, just to be safe, for reasoning models), which 5th or 6th graders in the US are expected to be able to do with no problem - without using a calculator.
So am I. See, for example, Karpathy's discussion of native computation: https://youtu.be/7xTGNNLPyMI?si=Gckcmp2Sby4SlKje&t=6416 (starts at 1:46:56). The first few tokens in the context actually serve as some sort of substrate for general computation. I don't pretend to understand that, and it may still be something of an open research topic, but it's one more unexpected emergent property of transformers.
You'd be crazy to trust that property at this stage -- at the time Karpathy was making the video, he needed to explicitly tell the model to "Use code" if he didn't want it to just make up solutions to more complex problems -- but you'd also be crazy to trust answers from a 5th-grader who just learned long division last week.
Feeling quite comfortable being classed in the same camp with the greatest living physicist, honestly.
Not a great time for you to rest on your intellectual laurels. Same goes for Penrose.
Yes, it is. You seem to have misunderstood what I wrote. The critique I was pointing to is of the amount of examples and energy needed during model training, which is what the "learning" in "machine learning" actually refers to. The paper uses GPT-3 which had already absorbed all that data and electricity. And the "learning" the paper talks about is arguably not real learning, since none of the acquired skills persists beyond the end of the session.
> So am I.
This is easy to settle. Go check any frontier model and see how far they get with multiplying numbers with tool calling disabled.
> Not a great time for you to rest on your intellectual laurels. Same goes for Penrose.
Neither am I resting, nor are there much laurels to rest on, at least compared to someone like Penrose. As for him, give the man a break, he's 94 years old and still sharp as a tack and intellectually productive. You're the one who's resting, imagining you've settled a question which is very much still open. Certainty is certainly intoxicating, so I understand where you're coming from, but claiming anyone who doubts computationalism is not bringing any arguments to the table is patently absurd.
Yes, it is. You seem to have misunderstood what I wrote. The critique I was pointing to is of the amount of examples and energy needed during model training, which is what the "learning" in "machine learning" actually refers to. The paper uses GPT-3 which had already absorbed all that data and electricity. And the "learning" the paper talks about is arguably not real learning, since none of the acquired skills persists beyond the end of the session.
Nobody is arguing about power consumption in this thread (but see below), and in any case the majority of power consumption is split between one-time training and the burden of running millions of prompts at once. Processing individual prompts costs almost nothing.
And it's already been stipulated that lack of long-term memory is a key difference between AI and human cognition. Give them some time, sheesh. This stuff's brand new.
This is easy to settle. Go check any frontier model and see how far they get with multiplying numbers with tool calling disabled.
Yes, it is very easy to settle. I ran this session locally in Qwen3-Next-80B-A3B-Instruct-Q6_K: https://pastebin.com/G7Ewt5Tu
This is a 6-bit quantized version of a free model that is very far from frontier level. It traces its lineage through DeepSeek, which was likely RL-trained by GPT 4.something. So 2 out of 4 isn't bad at all, really. My GPU's power consumption went up by about 40 watts while running these queries, a bit more than a human brain.
If I ask the hardest of those questions on Gemini 3, it gets the right answer but definitely struggles: https://pastebin.com/MuVy9cNw
As for him, give the man a break, he's 94 years old and still sharp as a tack and intellectually productive.
(Shrug) As long as he chooses to contribute his views to public discourse, he's fair game for criticism. You don't have to invoke quantum woo to multiply numbers without specialized tools, as the tests above show. Consequently, I believe that a heavy burden of proof lies with anyone who invokes quantum woo to explain any other mental operations. It's a textbook violation of Occam's Razor.
Usually it is the work of the one claiming something to prove it.
So if you believe that AI does "think" you are expected to show me that it really does.
Claiming it "thinks - prove otherwise" is just bad form and also opens the discussion up for moving the goalposts just as you did with your brain emulation statement. Or you could just not accept any argument made or circumvent it by stating the one trying to disprove your assertion got the definition wrong.
There are countless ways to start a bad faith argument using this methodology, hence: Define property -> prove property.
Conversely, if the one asserting something doesn't want to define it there is no useful conversation to be had. (as in: AI doesn't think - I won't tell you what I mean by think)
PS: Asking someone to falsify their own assertion doesn't seem a good strategy here.
PPS: Even if everything about the human brain can be emulated, that does not constitute progress for your argument, since now you'd have to assert that AI emulates the human brain perfectly before it is complete. There is no direct connection between "This AI does not think" to "The human brain can be fully emulated". Also the difference between "does not" and "can not" is big enough here that mangling them together is inappropriate.
So if you believe that AI does "think" you are expected to show me that it really does.
A lot of people seemingly haven't updated their priors after some of the more interesting results published lately, such as the performance of Google's and OpenAI's models at the 2025 Math Olympiad. Would you say that includes yourself?
If so, what do the models still have to do in order to establish that they are capable of all major forms of reasoning, and under what conditions will you accept such proof?
It definietly includes myself, I don't have the interest to stay updated here.
For that matter I have no opinion on if AI does think or not, I simply don't care.
Therefore I also really can't answer your question in what more a model has to do to establish that they are thinking (does being able to use all major forms of reasoning constitute the capability of thought to you?).
I can say however, that any such proof would have to be on a case-by-case basis given my current understanding on AI is designed.
Well first of all I never claimed that I was capable of thinking (smirk).
We also haven't agreed on a definition of "thinking" yet, so as you can read in my previous comment, there is no meaningful conversation to be had.
I also don't understand how your oddly aggresive phrasing adds to the conversation,
but if it helps you: my rights and protections do not depend on whether I'm able to prove to you that I am thinking.
(It also derails the conversation for what it's worth - it's a good strategy in the debating club, but these are about winning or loosing and not about fostering and obtaining knowledge)
Whatever you meant to say with "Sometimes, because of the consequences of otherwise, the order gets reversed" eludes me as well.
If I say I'm innocent, you don't say I have to prove it. Some facts are presumed to be true without burden of evidence because otherwise it could cause great harm.
So we don't require, say, minorities or animals to prove they have souls, we just inherently assume they do and make laws around protecting them.
Thank you for the clarification.
If you expect me to justify an action depending on you being innocent, then I actually do need you to prove it.
I wouldn't let you sleep in my room assuming you're innocent - or in your words: because of the consequences of otherwise.
It feels like you're moving the goalposts here: I don't want to justify an action based on something, i just want to know if something has a specific property.
With regards to the topic: Does AI think?
I don't know, but I also don't want to act upon knowing if it does (or doesn't for that matter).
In other words, I don't care.
The answer could go either way, but I'd rather say that I don't know (especially since "thinking" is not defined).
That means that I can assume both and consider the consequences using some heuristic to decide which assumption is better given the action I want to justify doing or not doing.
If you want me to believe an AI thinks, you have to prove it, if you want to justify an action you may assume whatever you deem most likely.
And if you want to know if an AI thinks, then you literally can't assume it does; simple as that.
Someone asserts, almost religiously, that LLMs do and/or can "think." When asked how to falsify their assertion, perhaps by explaining what exactly is "thinking" in the human brain that can and/or will be possible to emulate...
Err, no, that’s not what’s happening. Nobody, at least in this thread (and most others like it I’ve seen), is confidently claiming LLMs can think.
There are people confidently claiming they can’t and then other people expressing skepticism at their confidence and/or trying to get them to nail down what they mean.
Or they just point to the turing test which was the defacto standard test for something so nebulous. And behold: LLM can pass the turing test. So they think. Can you come up with something better (than the turing test)?
But the Turing test (which I concede, LLMs do pass) doesn't test if some system is thinking; it tests if the system can convince an unbiased observer that it is thinking. I cannot come up with a better "is this thing thinking" test, but that doesn't mean that such a test can't exist; I'm sure there are much smarter people then me trying to solve this problem.
When asked how to falsify their assertion, perhaps by explaining what exactly is "thinking" in the human brain that can and/or will be possible to emulate...
... someone else points out that the same models that can't "think" are somehow turning in gold-level performance at international math and programming competitions, making Fields Medalists sit up and take notice, winning art competitions, composing music indistinguishable from human output, and making entire subreddits fail the Turing test.
> That's kind of a big difference, wouldn't you say?
To their utility.
Not sure if it matters on the question "thinking?"; even if for the debaters "thinking" requires consciousness/qualia (and that varies), there's nothing more than guesses as to where that emerges from.
For my original earlier reply, the main subtext would be: "Your complaint is ridiculously biased."
For the later reply about chess, perhaps: "You're asserting that tricking, amazing, or beating a human is a reliable sign of human-like intelligence. We already know that is untrue from decades of past experience."
You're asserting that tricking, amazing, or beating a human is a reliable sign of human-like intelligence.
I don't know who's asserting that (other than Alan Turing, I guess); certainly not me. Humans are, if anything, easier to fool than our current crude AI models are. Heck, ELIZA was enough to fool non-specialist humans.
In any case, nobody was "tricked" at the IMO. What happened there required legitimate reasoning abilities. The burden of proof falls decisively on those who assert otherwise.
reply