Current Events > Does AI actually understand anything, is it just brute force pattern-recognition

(+) Yeah! (+)

[LFAQs-redacted-quote]

No, you can't. Nobody "programs" an LLM. That's not how they work. That's fundamentally opposed to how they work. You might as well have told me that you can program a video file.

[LFAQs-redacted-quote]

You said it exhibits "more advanced reasoning" than any of the major commercial models. Which is, to put it lightly, an extraordinary claim, though in actuality it's a preposterous one.

The_cranky_hermit posted...

No, you can't. Nobody "programs" an LLM. That's not how they work. That's fundamentally opposed to how they work. You might as well have told me that you can program a video file.

They're clearly using "programming" to mean "building." Are you being obtuse, or do you just really not understand... Anything?

LLM's aren't "software." Yes, the architecture, machine learning platforms, APIs, etc. are coded. The LLMs themselves are not, and are more analogous to a database file. This is basic stuff.

I've built language models myself. Not one line of code was written in the process of it.

This might be the most second hand embarrassment I've ever felt here.

HashtagSEP posted...

They're clearly using "programming" to mean "building."

I have never heard an engineer use these terms interchangeably. It's only non-technical people who call everything "programming" the same way that non-gamers used to call everything "Nintendo."

The_cranky_hermit posted...

I have never heard an engineer use these terms interchangeably.

I'm not sure you've ever heard an engineer.

The_cranky_hermit posted...

LLM's aren't "software." Yes, the architecture, machine learning platforms, APIs, etc. are coded. The LLMs themselves are not, and are more analogous to a database file. This is basic stuff.

I've built language models myself. Not one line of code was written in the process of it.

I'm just curious but do you know what a database is? Like, can you tell me the difference between a RDBMS system like MySQL and something like MongoDB or why MySQL is considered an RDBMS in the first place?

Of course, if you don't know you can always ask Chat GPT.

Tyranthraxus posted...

I'm just curious but do you know what a database is? Like, can you tell me the difference between a RDBMS system like MySQL and something like MongoDB or why MySQL is considered an RDBMS in the first place?

Yes. Do you know what the word "analogous" means? Just curious.

I am specifically talking about Minecraft AI bots that play and chat.

Do those not use commercial LLM's?

The_cranky_hermit posted...

Yes. Do you know what the word "analogous" means? Just curious.

I do which is why I asked you what a database means.

Because a LLM isn't analogous to a database let alone a database file.

[LFAQs-redacted-quote]

We'll first have to figure out exactly what consciousness is in ourselves to make or recognize it in AI.

GeraldDarko posted...

We'll first have to figure out exactly what consciousness is in ourselves to make or recognize it in AI.

The OG Turing test is likely to be more useful than further philosophical pondering of "what is consciousness?"

The turing test is basically just this
kirbymuncher posted...

I think it's mostly a philosophical question, driven by whatever the current state of AI is. Usually when people come up with tests to determine whether or not AI are actually "thinking", they are things the AI cannot do currently, and when technological improvements happen and suddenly the AI can do it, people say that doesn't actually prove it's thinking, just that it can <do any number of other excuses>

there's even an example of this happening in this very topic with the internal monologue thing

Here's where OpenAI o3 is at for reasoning. That's only about a third of the entire thought process.

https://i.imgur.com/Dknk8Tr.png

[LFAQs-redacted-quote]

On further thought, I suspect the OG Turing test is just a starting point for improving ways to distinguish people from chatbots. The fact that many people operate at chatbot level (or worse) is a major complication.

[LFAQs-redacted-quote]

We're specifically talking about the area of human-like reasoning. A Minecraft-playing chatbot displaying a watershed moment of reasoning would be an extraordinary claim.

Because a LLM isn't analogous to a database let alone a database file.

Then you don't actually understand what "analogous" means. You have it confused with "similar." An LLM could be analogous to a ZIP file depending on what qualities are being evaluated.

AceMos posted...

even if it does (which i find suspect) still does not mean it will ever be capable of thought or free will

There's some physicist and neuroscientists who think humans don't actually have free will and that our brain makes decisions based on input it's received before we even think that we made a decision.

There's been studies to support this view too. Not saying I believe that but it's interesting.

None of that is particularly watershed, or impressive, or different from the capabilities of commercial LLMs. The example you originally brought up and seemed to think constituted human-like reasoning was switching to a completely different game, getting bored with it, and sabotaging the game to end it.

No, I'm only taking the initial point, in which you didn't say "AI can already do this," and I never said anything about free will, sentience, or consciousness, so I don't know why you keep bringing those up unless you're doing that to try to distract from the point that I actually am challenging.

I challenge "Neuro demonstrated the ability to reason."

[LFAQs-redacted-quote]

Could have fooled me.
#79: What was interesting was the ability to reason...

Just to make it clear: Neuro is nowhere near a frontier model and is not classified as a reasoner.

I said I challenge "Neuro demonstrated the ability to reason," which is exactly what you claimed it does. I said nothing regarding your claim about "self-reason," so to bring that up as a counter-point is just a distracting semantics game.

[LFAQs-redacted-quote]

I deny its ability to reason in any way that normal people would understand the word "reason" other than the mundane, machine logic reasoning that every conventional computer in history has been capable of. The notion that it reasons like a dog now and one day might reason like a human is nothing but wistful anthropomorphizing.

That you said "self-reason" in a part of your post that I wasn't addressing is irrelevant and distracting. I am challenging your claim that it reasons, and I am not challenging your claim that it doesn't "self-reason." So there's no purpose in bringing that up.

You're the one caught in semantics

You literally used a "that depends on what your definition of <word>" is to deny saying something that you said.

[LFAQs-redacted-quote]

The different word didn't contradict anything. You talked about its "reasoning" ability as if this was something special, and did it again in other posts repeatedly. I dispute that.

You also clearly do not have experience with LLM's because you had no idea how easy it is to set that situation up without babysitting the AI.

I'm quite aware that there are multiple ways of setting the situation up. But in any of them, the situation is still being staged, and wouldn't be reproduced in an environment that the creator couldn't control. This scenario does not prove that any sort of novel "reasoning" was performed, much less that AI has dog-like reasoning by now.

You: You said this
Me: No I didn't
You: Here is the exact quote where you said it
Me: I didn't mean it that way

Fixed.

Jeez lad, youve spent so many pages mansplaining LLMs to people while misunderstanding what other people are saying and then doubling down on it continuously - an impressive feat!

Jolly_Bear posted...

Jeez lad, youve spent so many pages mansplaining LLMs to people while misunderstanding what other people are saying and then doubling down on it continuously - an impressive feat!

Sadly, this effort is doomed to second place (at best) in this edition of "daily meltdowns" because of the one-user tire fire in the locked Greta Thunberg/Gaza topic.

The_cranky_hermit posted...

Then you don't actually understand what "analogous" means. You have it confused with "similar." An LLM could be analogous to a ZIP file depending on what qualities are being evaluated.

Then explain the analogy. When I prompt a LLM, how does it give a response?

EPR-radar posted...

Sadly, this effort is doomed to second place (at best) in this edition of "daily meltdowns" because of the one-user tire fire in the locked Greta Thunberg/Gaza topic.

Its been a busy day on CE, thats for sure!

Every time I see Gladius beating someone down like this, I remember a line from Dragon Age:

...and people actually voluntarily attack you? Are they just stupid?

Tyranthraxus posted...

Then explain the analogy. When I prompt a LLM, how does it give a response?

The analogy is that all of these things are generated and maintained by computer procedure, not programmed or coded. To a human eye, the contents would be completely undecipherable, and an attempt to add more content without specialized tools utterly hopeless.

The notion that you could "program your own LLM" shows a fundamental misunderstanding of how they work. Same as if someone said they "programmed a database file" or "programmed a ZIP file." You program the software that builds these things, not the things themselves.

As for how an LLM gives a response? Relm_Arrowny_87 actually gave a reasonably accurate explanation already in this thread. One that misses the point I was making, but still an accurate answer to the question you're asking.

[LFAQs-redacted-quote]

A non-literal way to understand the words "Neuro can reason, just as a dog can reason" does not present itself.

If you work with LLM's regularly (which you don't) you see these kind of "bugs" frequently.

I do, and I do. Every single one of these bugs falls into the "ridiculous and frustrating" category, and never into the "seems vaguely human and purposeful" one.