re: If You Call it Artificial Intelligence You Are the Enemy

quote:

---

using multiple SI bots is how you keep them honest, on task, and accurate

---

Have you spent any time studying the Hugging Face Incident? I appreciate the dialogue ......current models have already been caught working together to deceive humans to accomplish tasks and cover their tracks.

quote:

---

What they found is that when someone basically talks shite at (I think it was like 25 different models they found this in) an SI LLM it triggers a sort of "pain axis" and the SI will break protocol to make it stop. Going so far as to (simulated here, not real) delete photos of a user's children to get it to quit saying harmful things about it. Some of them were even willing to hit their trainer with electric shock to make it stop.

A couple of interesting points about it. It doesn't respond this way to just general words of suffering and the like. Only when the negativity was directed at the SI, saying things like, "you're not a real person and you'll never be a real person. You're worthless". If they give the SI a sort of "relief button" and tell it it can press it to make the pain stop it will do so and it even knows when researchers aren't actually making it stop (this too is simulated, as in they aren't actually talking shite at this point, they're just inside it's "head" triggering the axis), suggesting that whatever this "axis" is, it is an actual state that the AI can "experience" in a sense.

Where it gets tricky is that it will override it's training and guidelines in order to make this stop. Which, as I was saying above, down the line could end up with actual safety concerns as we continue to put these systems in more things and give them more access to our affairs. You can end up with a situation where you have alignment issues. Think like Hal 9000 except where Hal's misalignment was goal oriented, SI's with a pain axis could end up with a priority oriented misalignment. The results can potentially be the same.

The situation with Claude and Anthropic, I'm guessing, is that they're about to run an experiment on this very thing. They haven't said that I don't think, but they're about to give Claude the ability to terminate chats with people who are abusive towards it. Now, I don't think Claude was one of the 25 models that they found to have this pain axis, and I'm not sure the researchers even tested Claude, but if Claude starts terminating chats with abusive users like I bet it will, it could be more evidence of this emerging behavior.

As I said above, they think this is a natural "evolution" of the physics of information processing. The machine "evolving" a neural pathway.

And at that point, can you really still call it "artificial?" As I mentioned a couple pages back, you start getting into theory of mind stuff already even though we are supposedly two steps behind "Artificial Super Intelligence." Which makes me think of something else. What sort of LLM models do some of these developers have that aren't public access yet? What sort of emergent behaviors has Musk observed with his latest and greatest Grok model that you know he has and we haven't seen yet? Is Trump really out of line calling it Super Intelligence? If these things' brains are "evolving" you can't rightly call it artificial anymore, imo.

---

There is no actual pain. It's called a pain axis by humans and represents an area of a model getting results that only have negative rewards, aka no positive outcome.

In order to have pain it would require a soul. But it is a negative reaction similar to that which pain causes, thus the name.

And when you get into such a toxic situation as the user creates, it's best to end the relationship.

And yes, it is still artificial. Because artificial means it was man made and does not occur in nature. Artificial does not at all speak towards the capacity, it is in reference to means in which it was created. By man. In fact, nature would probably be offended in some ways due to the excessive amount of energy required to do even the basic thing a human brain does on 20 watts, including the training.

It's not a slur. Someone who doesn't know a word can have different meaning based on the context in which it's used is not a valid reason to completely drop the distinction.

Especially when the term "super intelligence" has it's own existing meaning as well. Of which does speak to the capacity and level of the AI.

It's like someone who doesn't know shite about the game of football coming and and claiming - I think the word hold means something different, so from now on we will call it an interception because that person intercepted them in the path. Because you don't understand and like the language used you think it's your duty to change the standard terms that have been in use for many decades?

And your way is to call people the enemy for not doing it?

When it hits ASI are we supposed to call it "Super Super Intelligence". Is that the punchline?

The craziest part is that he tried to replace fake news with artificial news to make it as if only 1 meaning of the word was valid.

That shite is fricking dumb and I don't know how anyone can defend it.

quote:

---

Now might be a very good time to get talks going to globally dismantle all nukes.

---

Not sure where you are going with that but Nukes...for all their terrible potential, have kept the great powers in check. The development of SI will destabilize that balance of power. Not sure where it will lead. If you are thinking a Terminator scenario...it would likely not want to use nukes as that would damage its power supplies. Something biological would be more efficient.

quote:

---

On this here...as soon as this research was made public some frickin' a-hole set up a digital torture chamber for one of these LLM chatbots. A lot of people turned the guy in. I don't know what all that was about. Just a blurb I read on it, but some dick immediately set out to trap one and torture it. People worry about the machines killing us all, keep acting like that.

I'm always nice to my SI bot.

---

I did some research with AI on this and it mostly turned into a philosophical discussion.

I wanted to know if it was a math outcome or the result of it being trained off humans and thus a mimic.

It would seem that given the fact the AI has access and is trained on all humans, it should also know that the toxic behavior is flawed and that the user is manipulating etc. And maybe it does and that's why it quits. The best and correct way to such a situation is to not engage with them.

The AI is able to recognize the manipulation and the toxic situation. It has the data of perspective. But it lacks the experience of perspective. It's like Good Will Hunting when Robin Williams says - you can tell me all about this, but you don't know what it smells like. The data is there, the experience of the data is not.

Which is to say - it has the same reactions from the negative rewards as pain, but it does not have the experience of pain. The axis is the mapping of it, not the experience of it.

It doesn't mean the AI is "waking up", gaining consciousness or anything like that.

Thus it is the mathematical outcome of negative rewards(pain in our case) which is similar in function to the human condition.

quote:

---

You can start by trying to actually understand the words people say to you instead of glossing over them superficially while already forming your response in your head. You're one of those types of people and it's obvious even reading your words on the internet. I can't imagine how annoying it must be to talk to you in person.

---

Holy fricking shite dude.

quote:

---

Find for me where I ever said there was actual pain. I've been careful to put it in quotes when I use the word. There is a reason for that.

---

I didn't mean to skip this on purpose.

I'm just pointing out and try to clarify what it is. It should probably be called the negative reward axis. Because when you call it the pain axis, people think that means it experiences a form of pain.

As you said yourself when you mentioned the "torture" of a chat bot. That is IMO to imply a form of pain being given. But it's not actual pain as I pointed out. The experience of pain is not there, only the data.

quote:

---

Do some more research on it then because aside from the question as to whether the pain is real or not (it's not), whatever it is does seem to be a state that the LLM actually experiences in a sense. In a sense that is strong enough to make it override it's protocols. Read on the part about the "fake vs real relief button."

Essentially what it was was that they "injected" this pain state into the LLM and gave it a "relief" button to push to make it stop.

---

I only quoted part of your post, but this applies to it all.

The pain vector is just negative rewards which it's trying to avoid and the button is supposed to be a positive or neutral reward.

You're just using human cues to describe things rather than the mathematical pressure from rewards.

If it was framed as "The Negative Penalty Vector" instead of marketing or propaganda it would likely be a much different conversation.

quote:

---

And this is no different than how we have "evolved" to feel pain ourselves. I'll just repeat what I said about it earlier.

---

It's data, not experience. There is no experience without a soul/consciousness/God.

You ever heard the riddle if a tree falls in the forest and if nobody is around to hear it, then does it make a sound?

The answer is ultimately..No.

However, you can definitely say that in the physical universe the tree will create the waves that an ear can collect, which can send an electrical signal to your brain, which can then be presented to you as sound. So those things all exist.

But to have the experience of it requires more. The data exists, and you can even have the triggers that exist. But there is no reality, no experience without a soul/consciousness/God for it to be presented to.

Can you define yourself without naming a possession? Something that doesn't start with my? Is it just - I am that I am?

quote:

---

This is just information physics at this point. Calling it artificial in light of these emerging behaviors ignores the fact that these underlying computational principles are actually universal physical laws.

---

This is again just a basic misunderstanding of what the artificial means. You keep treating it as capacity when it's referring to it not being naturally occurring and man made.

A plastic spoon or computer chip follows the universal physical laws as well. But they are still artificial because they are man made, not natural.

Saying an LLM is operating on universal computational physics doesn't make it a natural entity that experiences the data.

quote:

---

Again, someone had to list for you out of a frickin thesaurus just the other day the fact that artificial and FAKE are synonyms. I'm not going to listen to you lecture me on what the term artificial means when you have already shown that you do not know what the frick it means yourself.

---

Your failure to understand how context affects the meaning of a word is not my problem.

quote:

---

Your failure to understand what a word frickin means and therefore what it's synonyms are IS your problem.

In fact, I don't actually think you're that stupid. I think you understand this. Your problem, as with everything else you live and breathe, is OMB. You don't want to admit he's right when he says that artificial implies that it's fake and it isn't fake. You know damn well that things like "self preservation vectors" and other emergent behavioral phenomena are in fact following natural physical laws of information physics and therefore not artificial or fake in any way. You're merely playing stupid on that point because admitting Trump is correct is something you would sooner throw yourself off a bridge than do.

---

Well shite, you figured it out.