[pfisch]

Even very young children with very simple thought processes, almost no language capability, little long term planning, and minimal ability to form long-term memory actively deceive people. They will attack other children who take their toys and try to avoid blame through deception. It happens constantly.

LLMs are certainly capable of this.

[mikepurvis]

Dogs too; dogs will happily pretend they haven't been fed/walked yet to try to get a double dip.

Whether or not LLMs are just "pattern matching" under the hood they're perfectly capable of role play, and sufficient empathy to imagine what their conversation partner is thinking and thus what needs to be said to stimulate a particular course of action.

Maybe human brains are just pattern matching too.

[iamacyborg]

> Maybe human brains are just pattern matching too.

I don't think there's much of a maybe to that point given where some neuroscience research seems to be going (or at least the parts I like reading as relating to free will being illusory).

[mikepurvis]

My sense is that for some time, mainstream secular philosophy has been converging on a hard determinism viewpoint, though I see the wikipedia article doesn't really take stance on its popularity, only really laying out the arguments: https://en.wikipedia.org/wiki/Free_will#Hard_determinism

[sejje]

I agree that LLMs are capable of this, but there's no reason that "because young children can do X, LLMs can 'certainly' do X"

[anonymous908213]

Are you trying to suppose that an LLM is more intelligent than a small child with simple thought processes, almost no language capability, little long-term planning, and minimal ability to form long-term memory? Even with all of those qualifiers, you'd still be wrong. The LLM is predicting what tokens come next, based on a bunch of math operations performed over a huge dataset. That, and only that. That may have more utility than a small child with [qualifiers], but it is not intelligence. There is no intent to deceive.

[ctoth]

A small child's cognition is also "just" electrochemical signals propagating through neural tissue according to physical laws!

The "just" is doing all the lifting. You can reductively describe any information processing system in a way that makes it sound like it couldn't possibly produce the outputs it demonstrably produces. "The sun is just hydrogen atoms bumping into each other" is technically accurate and completely useless as an explanation of solar physics.

[anonymous908213]

You are making a point that is in favor of my argument, not against it. I make the same argument as you do routinely against people trying to over-simplify things. LLM hypists frequently suggest that because brain activity is "just" electrochemical signals, there is no possible difference between an LLM and a human brain. This is, obviously, tremendously idiotic. I do believe it is within the realm of possibility to create machine intelligence; I don't believe in a magic soul or some other element that make humans inherently special. However, if you do not engage in overt reductionism, the mechanism by which these electrochemical signals are generated is completely and totally different from the signals involved in an LLM's processing. Human programming is substantially more complex, and it is fundamentally absurd to think that our biological programming can be reduced to conveniently be exactly equivalent to the latest fad technology and assume that we've solved the secret to programming a brain, despite the programs we've written performing exactly according to their programming and no greater.

Edit: Case in point, a mere 10 minutes later we got someone making that exact argument in a sibling comment to yours! Nature is beautiful.

[emp17344]

> A small child's cognition is also "just" electrochemical signals propagating through neural tissue according to physical laws!

This is a thought-terminating cliche employed to avoid grappling with the overwhelming differences between a human brain and a language model.

[pfisch]

Yes. I also don't think it is realistic to pretend you understand how frontier LLMs operate because you understand the basic principles of how the simple LLMs worked that weren't very good.

Its even more ridiculous than me pretending I understand how a rocket ship works because I know there is fuel in a tank and it gets lit on fire somehow and aimed with some fins on the rocket...

[anonymous908213]

The frontier LLMs have the same overall architecture as earlier models. I absolutely understand how they operate. I have worked in a startup wherein we heavily finetuned Deepseek, among other smaller models, running on our own hardware. Both Deepseek's 671b model and a Mistral 7b model operate according to the exact same principles. There is no magic in the process, and there is zero reason to believe that Sonnet or Opus is on some impossible-to-understand architecture that is fundamentally alien to every other LLM's.

[pfisch]

Deepseek and Mistral are both considerably behind Opus, and you could not make deepseek or mistral if I gave you a big gpu cluster. You have the weights but you have no idea how they work and you couldn't recreate them.

> I have worked in a startup wherein we heavily finetuned Deepseek, among other smaller models, running on our own hardware.

Are you serious with this? I could go make a lora in a few hours with a gui if I wanted to. That doesn't make me qualified to talk about top secret frontier ai model architecture.

Now you have moved on to the guy who painted his honda, swapped out some new rims, and put some lights under it. That person is not an automotive engineer.

[anonymous908213]

I'm not talking about a lora, it would be nice if you could refrain from acting like a dipshit.

> and you could not make deepseek or mistral if I gave you a big gpu cluster. You have the weights but you have no idea how they work and you couldn't recreate them.

I personally couldn't, but the team behind that startup as a whole absolutely could. We did attempt training our own models from scratch and made some progress, but the compute cost was too high to seriously pursue. It's not because we were some super special rocket scientists, either. There is a massive body of literature published about LLM architecture already, and you can replicate the results by learning from it. You keep attempting to make this out to be literal fucking magic, but it's just a computer program. I guess it helps you cope with your own complete lack of understanding to pretend that it is magical in nature and can't be understood.

[pfisch]

No, it's just obvious that there is a massive race going with trillions of dollars on the line. No one is going to reveal the details of how they are making these AIs. Any public information that exists about them is way behind SOTA.

I strongly suspect that it is really hard to get these models to converge though so I have no idea what your team could've theoretically made, but it certainly would've been well behind SOTA.

My point is if they are changing core elements of the architecture you would have no idea because they wouldn't be telling anyone about it. So thinking you know how Opus 4.6 works just isn't realistic until development slows down and more information comes out about them.

[mikepurvis]

Short term memory is the context window, and it's a relatively short hop from the current state of affairs to here's an MCP server that gives you access to a big queryable scratch space where you can note anything down that you think might be important later, similar to how current-gen chatbots take multiple iterations to produce an answer; they're clearly not just token-producing right out of the gate, but rather are using an internal notepad to iteratively work on an answer for you.

Or maybe there's even a medium term scratchpad that is managed automatically, just fed all context as it occurs, and then a parallel process mulls over that content in the background, periodically presenting chunks of it to the foreground thought process when it seems like it could be relevant.

All I'm saying is there are good reasons not to consider current LLMs to be AGI, but "doesn't have long term memory" is not a significant barrier.

[nurettin]

Intelligence is about acquiring and utilizing knowledge. Reasoning is about making sense of things. Words are concatenations of letters that form meaning. Inference is tightly coupled with meaning which is coupled with reasoning and thus, intelligence. People are paying for these monthly subscriptions to outsource reasoning, because it works. Half-assedly and with unnerving failure modes, but it works.

What you probably mean is that it is not a mind in the sense that it is not conscious. It won't cringe or be embarrassed like you do, it costs nothing for an LLM to be awkward, it doesn't feel weird, or get bored of you. Its curiosity is a mere autocomplete. But a child will feel all that, and learn all that and be a social animal.

[jvidalv]

What is the definition for intelligence?

[anonymous908213]

Quoting an older comment of mine...

  Intelligence is the ability to reason about logic. If 1 + 1 is 2, and 1 + 2 is 3, then 1 + 3 must be 4. This is deterministic, and it is why LLMs are not intelligent and can never be intelligent no matter how much better they get at superficially copying the form of output of intelligence. Probabilistic prediction is inherently incompatible with deterministic deduction. We're years into being told AGI is here (for whatever squirmy value of AGI the hype huckster wants to shill), and yet LLMs, as expected, still cannot do basic arithmetic that a child could do without being special-cased to invoke a tool call.

  Our computer programs execute logic, but cannot reason about it. Reasoning is the ability to dynamically consider constraints we've never seen before and then determine how those constraints would lead to a final conclusion. The rules of mathematics we follow are not programmed into our DNA; we learn them and follow them while our human-programming is actively running. But we can just as easily, at any point, make up new constraints and follow them to new conclusions. What if 1 + 2 is 2 and 1 + 3 is 3? Then we can reason that under these constraints we just made up, 1 + 4 is 4, without ever having been programmed to consider these rules.

[coldtea]

>Intelligence is the ability to reason about logic. If 1 + 1 is 2, and 1 + 2 is 3, then 1 + 3 must be 4. This is deterministic, and it is why LLMs are not intelligent and can never be intelligent no matter how much better they get at superficially copying the form of output of intelligence.

This is not even wrong.

>Probabilistic prediction is inherently incompatible with deterministic deduction.

And his is just begging the question again.

Probabilistic prediction could very well be how we do deterministic deduction - e.g. about how strong the weights and how hot the probability path for those deduction steps are, so that it's followed every time, even if the overall process is probabilistic.

Probabilistic doesn't mean completely random.

[runarberg]

At the risk of explaining the insult:

https://en.wikipedia.org/wiki/Not_even_wrong

Personally I think not even wrong is the perfect description of this argumentation. Intelligence is extremely scientifically fraught. We have been doing intelligence research for over a century and to date we have very little to show for it (and a lot of it ended up being garbage race science anyway). Most attempts to provide a simple (and often any) definition or description of intelligence end up being “not even wrong”.

[famouswaffles]

>Intelligence is the ability to reason about logic. If 1 + 1 is 2, and 1 + 2 is 3, then 1 + 3 must be 4.

Human Intelligence is clearly not logic based so I'm not sure why you have such a definition.

>and yet LLMs, as expected, still cannot do basic arithmetic that a child could do without being special-cased to invoke a tool call.

One of the most irritating things about these discussions is proclamations that make it pretty clear you've not used these tools in a while or ever. Really, when was the last time you had LLMs try long multi-digit arithmetic on random numbers ? Because your comment is just wrong.

>What if 1 + 2 is 2 and 1 + 3 is 3? Then we can reason that under these constraints we just made up, 1 + 4 is 4, without ever having been programmed to consider these rules.

Good thing LLMs can handle this just fine I guess.

Your entire comment perfectly encapsulates why symbolic AI failed to go anywhere past the initial years. You have a class of people that really think they know how intelligence works, but build it that way and it fails completely.

[anonymous908213]

> One of the most irritating things about these discussions is proclamations that make it pretty clear you've not used these tools in a while or ever. Really, when was the last time you had LLMs try long multi-digit arithmetic on random numbers ? Because your comment is just wrong.

They still make these errors on anything that is out of distribution. There is literally a post in this thread linking to a chat where Sonnet failed a basic arithmetic puzzle: https://news.ycombinator.com/item?id=47051286

> Good thing LLMs can handle this just fine I guess.

LLMs can match an example at exactly that trivial level because it can be predicted from context. However, if you construct a more complex example with several rules, especially with rules that have contradictions and have specified logic to resolve conflicts, they fail badly. They can't even play Chess or Poker without breaking the rules despite those being extremely well-represented in the dataset already, nevermind a made-up set of logical rules.

[famouswaffles]

>They still make these errors on anything that is out of distribution. There is literally a post in this thread linking to a chat where Sonnet failed a basic arithmetic puzzle: https://news.ycombinator.com/item?id=47051286

I thought we were talking about actual arithmetic not silly puzzles, and there are many human adults that would fail this, nevermind children.

>LLMs can match an example at exactly that trivial level because it can be predicted from context. However, if you construct a more complex example with several rules, especially with rules that have contradictions and have specified logic to resolve conflicts, they fail badly.

Even if that were true (Have you actually tried?), You do realize many humans would also fail once you did all that right ?

>They can't even reliably play Chess or Poker without breaking the rules despite those extremely well-represented in the dataset already, nevermind a made-up set of logical rules.

LLMs can play chess just fine (99.8 % legal move rate, ~1800 Elo)

https://arxiv.org/abs/2403.15498

https://arxiv.org/abs/2501.17186

https://github.com/adamkarvonen/chess_gpt_eval

[coldtea]

>The LLM is predicting what tokens come next, based on a bunch of math operations performed over a huge dataset.

Whereas the child does what exactly, in your opinion?

You know the child can just as well to be said to "just do chemical and electrical exchanges" right?

[jazzyjackson]

Okay but chemical and electrical exchanges in an body with a drive to not die is so vastly different than a matrix multiplication routine on a flat plane of silicon

The comparison is therefore annoying

[coldtea]

>Okay but chemical and electrical exchanges in an body with a drive to not die is so vastly different than a matrix multiplication routine on a flat plane of silicon

I see your "flat plane of silicon" and raise you "a mush of tissue, water, fat, and blood". The substrate being a "mere" dumb soul-less material doesn't say much.

And the idea is that what matters is the processing - not the material it happens on, or the particular way it is.

Air molecules hitting a wall and coming back to us at various intervals are also "vastly different" to a " matrix multiplication routine on a flat plane of silicon".

But a matrix multiplication can nonetheless replicate the air-molecules-hitting-wall audio effect of reverbation on 0s and 1s representing the audio. We can even hook the result to a movable membrane controlled by electricity (what pros call "a speaker") to hear it.

The inability to see that the point of the comparison is that an algorithmic modelling of a physical (or biological, same thing) process can still replicate, even if much simpler, some of its qualities in a different domain (0s and 1s in silicon and electric signals vs some material molecules interacting) is therefore annoying.

[JoshTriplett]

Intelligence does not require "chemical and electrical exchanges in an body". Are you attempting to axiomatically claim that only biological beings can be intelligent (in which case, that's not a useful definition for the purposes of this discussion)? If not, then that's a red herring.

"Annoying" does not mean "false".

[jazzyjackson]

No I'm not making claims about intelligence, I'm making claims about the absurdity of comparing biological systems with silicon arrangements.

[coldtea]

>I'm making claims about the absurdity of comparing biological systems with silicon arrangements.

Aside from a priori bias, this assumption of absurdity is based on what else exactly?

Biological systems can't be modelled (even if in a simplified way or slightly different architecture) "with silicon arrangements", because?

If your answer is "scale", that's fine, but you already conceded to no absurdity at all, just a degree of current scale/capacity.

If your answer is something else, pray tell, what would that be?

[anonymous908213]

At least read the other replies that pre-emptively refuted this drivel before spamming it.

[coldtea]

At least don't be rude. They refuted nothing of the short. Just banged the same circular logic drum.

[anonymous908213]

There is an element of rudeness to completely ignoring what I've already written and saying "you know [basic principle that was already covered at length], right?". If you want to talk about contributing to the discussion rather than being rude, you could start by offering a reply to the points that are already made rather than making me repeat myself addressing the level 0 thought on the subject.

[JoshTriplett]

Repeating yourself doesn't make you right, just repetitive. Ignoring refutations you don't like doesn't make them wrong. Observing that something has already been refuted, in an effort to avoid further repetition, is not in itself inherently rude.

Any definition of intelligence that does not axiomatically say "is human" or "is biological" or similar is something a machine can meet, insofar as we're also just machines made out of biology. For any given X, "AI can't do X yet" is a statement with an expiration date on it, and I wouldn't bet on that expiration date being too far in the future. This is a problem.

It is, in particular, difficult at this point to construct a meaningful definition of intelligence that simultaneously includes all humans and excludes all AIs. Many motivated-reasoning / rationalization attempts to construct a definition that excludes the highest-end AIs often exclude some humans. (By "motivated-reasoning / rationalization", I mean that such attempts start by writing "and therefore AIs can't possibly be intelligent" at the bottom, and work backwards from there to faux-rationalize what they've already decided must be true.)