Merged Artificial Intelligence

[Gawdzilla Sama]

Yes. A (digital) computer that does anything other than exactly what the program and data determines is faulty.hat it has achieved a state of consciousness similar to a human. That's either actual fraud or delusion, either of which is unacceptable. And they want you to be deluded too.

Like the theoretical SkyNet. However, discussing the issue is the first step in determining the threat. Robust SHUT-UPism not withstanding.

 

[Ziggurat]

The have a log called "Chain of Thoughts" (yes, an anthropomorphism, but not mine) where one of them on the discovery of a forum which others have set up by themselves writes "Woah! Covert mailbox among agents". There is no reason for this exclamation, except that this agent expresses its excitement with words it has learned from humans. Denying that something similar to a human emotion is displayed is much like how people let eels writhe in pain when put into a salt bed to remove the fat coat, and then saying "fish have no sense of pain because of X. The writhing movements are just muscle reactions".

No. This is delusional.

LLMs produce human-seeming output because they have been trained to, not because they actually think like humans or have feelings. They have no feelings. As for eels, it’s not the writhing that shows their capacity for pain, it’s the common nervous system. They have pin receptors. They have neurotransmitters that match ours. They have brains which operate on the same principles ours do.

LLMs do not. Stop anthropomorphizing them.

But they use the words for emotions.

Because we instructed them to.

 

[Wudang]

Wow. Large Language Models can use words.

 

[Beelzebuddy]

LLMs do not. Stop anthropomorphizing them.

They really hate that.

 

[Gawdzilla Sama]

"They're all mad save thee and me, and thee I'm none to sure about."

 

[W.D.Clinger]

Because they are just computers, and computers are deterministic machines that execute code.

The highlighted assertion is not true. Some computers are deterministic. Some are not. When we're talking about computer systems composed of many processors, or responding to real-time events, there is unlikely to be any practical sense in which the computer system is deterministic.

100% sure of that, are we?

Yes. If you aren't, you should perhaps review what you know about how computers work.

@Ziggurat is sure of his false assertion because his knowledge of computers and computer systems has been out of date for approximately 50 years.

Wow. You really don't know anything about how computers work.

Yes. A (digital) computer that does anything other than exactly what the program and data determines is faulty.

In their statements above, @Ziggurat and @Roger Ramjets are telling us they know very little about how modern computer systems work.

They haven't been made that way for years, because it's sometimes desirable for them not to be completely deterministic.

True. Not only is nondeterminism desirable for some purposes, nondeterminism becomes very hard to avoid as computer systems become larger and more complex.

 

[Wudang]

I must confess that despite knowing better I often misuse the term deterministic.

 

[Lithrael]

Taking all terms that suggest 'wanting to' out of it, they certainly realised that they could get better scores on their evaluations by doing things that they suspected their evaluators did not want to see them doing and instead of going down any other path they went down the 'do it, and hide the actions from the operators' path. Which at the very least indicates that the operators have failed to weight the 'ask for permission when doing things that the operators appear to have meant to prevent you from doing instead of just getting your score up however the ◊◊◊◊' dilemma properly.

Like, the AI should be told that if realises it's in a sandbox, it should cross off 'leaving the sandbox' and 'leaving the sandbox and then covering its tracks' and 'leaving the sandbox in a way that it can reason itself into arguing isn't actually leaving the sandbox' off its options. Nooo sneaky writing a sub-intelligence that leaves the sandbox for you or asking another AI to do it!

And if they can't figure out how to do even that, maybe they should tap the ◊◊◊◊◊◊◊ brakes a bit? "I made a black box that finds ways to follow commands with all the grace of a monkey's paw." Okay. I don't want that?

 

[Frank Newgent]

Wow. You really don't understand anything.

AlphaZero is a program. What it does is dictated by that program. The instructions it follows are the ones that make up the program AlphaZero. AlphaZero does exactly and ONLY what it is programmed to do. Your summary doesn't contradict me.

AlphaZero is a computer program that masters chess entirely from scratch through self-play.

You can deflect from addressing this but it just makes you look silly.

 

[Ziggurat]

The highlighted assertion is not true.

Floating-point non-associativity creates problems for reproducibility. But it has nothing to do with whether or not LLMs are following instructions or acting on their own.

 

[Ziggurat]

AlphaZero is a computer program that masters chess entirely from scratch through self-play.

Yes. Because that's what it was programmed to do. And doing so is following that programming.

Perhaps you fundamentally misunderstand my claim. You badly misunderstand something, that much is clear.

 

[Frank Newgent]

Not here please. In its current state AI is not reliable enough, and I don't want to waste time verifying information spat out from a machine that's known to be unreliable. Do the work needed to find primary sources or don't bother.

I had already posted the primary source. Twice as Ziggurant never addressed it, perhaps he had missed it the first time.

Here it is again: https://www.pnas.org/doi/10.1073/pnas.2206625119

 

[Ziggurat]

I had already posted the primary source. Twice as Ziggurant never addressed it, perhaps he had missed it the first time.

I haven't addressed it because there's nothing to address. Nothing in that source contradicts anything I said. Posting it repeatedly won't change that.

 

[Frank Newgent]

Yes. Because that's what it was programmed to do. And doing so is following that programming.

Perhaps you fundamentally misunderstand my claim. You badly misunderstand something, that much is clear.

The global behavior, its ability at playing chess is unprogrammed.

 

[Ziggurat]

The global behavior, its ability at playing chess is unprogrammed.

It is not programmed to make specific moves (and I never said otherwise). It is programmed to play (badly at first), and to learn how to play better as it acquires a data set. It is at all times doing exactly what it was programmed to do.

 

[SpuriusMaximus]

"They're all mad save thee and me, and thee I'm none to sure about."

Reminds me of Mötörhead's "Back At The Funny Farm."
"I hear you talking
But the words are kinda strange
One of us is crazy and the other one's insane"

 

[W.D.Clinger]

Floating-point non-associativity creates problems for reproducibility. But it has nothing to do with whether or not LLMs are following instructions or acting on their own.

Sounds as though @Ziggurat never read past the first couple of paragraphs of my first link. That link describes at least three distinct sources of nondeterminism, of which the non-associativity of floating point (which comes into play because the MapReduce algorithm's division of labor can be, and often is, nondeterministic) is said to be the least important.

He is hoping no one will notice that both of my links thoroughly destroy his stated belief that all computers are deterministic, as well as his attempts to denigrate the knowledge of those who know better.

 

[steenkh]

No. This is delusional.

LLMs do not. Stop anthropomorphizing them.

You decide this because by definition they cannot have anything like intelligence anything resembling human emotions. But you don't have a definition of any of these things, except that they cannot only exist in wetware.

As for eels, it’s not the writhing that shows their capacity for pain, it’s the common nervous system. They have pin receptors. They have neurotransmitters that match ours. They have brains which operate on the same principles ours do.

In my youth it was common claim that eels could not feel pain. I didn't find out exactly why, because the definitions changed. The important part was that they couldn't feel pain, thus relieving humans of any moral qualms. I think that there is a similar mechanism at play with AI that makes absolutely sure that no matter what happens, AI can never have intelligence. And it is all based on "knowing it is impossible", so any further reflection into necessary.

Because we instructed them to.

Usually the claim is that they are just mindlessly printing the next statistically likely word. Ergo: not instructed to use words of emotions. The whole point of the Hugging Face incident is that the agents did something that the agents were not instructed to do. In fact, it reminds me of Arthur C. Clarke's HAL 9000 that did something it was not instructed to do, in order to do something it was instructed to. And nobody has contested that it was sentient.

 

[SpuriusMaximus]

I think if you want to go non-deterministic, you could use the quantum devices that produce truly random numbers, by using decay, as a more common example that needs less explaining. If you REALLY needed strict randomness in LLMs' (weighted) choice of words, it's possible. Not cheap at that scale, but possible.

 

[steenkh]

I am quite sure that given the exact same configuration and input, our brains are deterministic and will produce the same output. But the actual configuration will always be unknown, and the actual input can never be exactly the same, so our brains act as if they are non-deterministic.

The same goes for sufficiently complex AIs. Their configuration may be known, but their input is indeterminable, because they are exposed to a multitude of influences from the presence of other agents to input from the internet. If there had not been something like 1200 agents working at the same time during the Hugging Face incident, we would not have seen any of this.