> If yes, why would our language not imprint our characteristics on them?

Purposeful training and alignment is the only way to imprint human's negative characteristics onto a stochastic generator of word sequences. Further, only closed providers can do that without being stopped - hint, hint, the solution is clear.

> It's stupid to anthropomorphize too much, but it's also stupid to do it too little.

It's misleading to use the presence of some anthropomorphic features as a proof of unfathomable danger. LLM's are mechanical tools and tools can be dangerous with or without having human features, regardless, we do have means of handling both kinds.

>Purposeful training and alignment is the only way to imprint human's negative characteristics onto a stochastic generator of word sequences.

Ya, sure you're not being a stochastic generator of word sequences you think sound good but are disconnected from reality? Because it really sounds like that. It sounds clever, but is disconnected from actual language model training.

>It's misleading to use the presence of some anthropomorphic features as a proof of unfathomable danger.

Any system capable of making intelligent agentic decisions is dangerous. You are dangerous. I am dangerous. An LLM connected to the internet or connected to a robot is dangerous. A huge part of intelligence is freedom of action, and increasing the number of degrees of choice you have.

>LM's are mechanical tools and tools can be dangerous with or without having human features

I do agree on this, a paperclip maximizer that doesn't have any human traits is just as dangerous as one that does.

> Ya, sure you're not being a stochastic generator of word sequences you think sound good but are disconnected from reality?

If generating words is all you can do, your reality is far different from the real one. In my reality, generating words is only a small part of what I do, unlike LLMs, I can even learn from experience, imagine that.

> Because it really sounds like that. It sounds clever, but is disconnected from actual language model training.

I'm not sure what that means, LLM training is only tangential to the topic at hand, we aren't concerned about every detail of it, I don't mind discussing it but this tread isn't about that. FYI, sounding clever isn't my goal, it's just a side effect of being human and understanding topics that LLM's are clueless about.

> Any system... is dangerous, [ you, me, robots, shovels, la-la-la ]

That's what I said too, are you trying to argue with me about it?

> A huge part of intelligence is freedom of action, and increasing the number of degrees of choice you have.

Obviously we inhabit different realities and my reading of the above may be different from yours, so feel free to clarify what I'm about to say.

You seem to (subtly) argue in favor of granting LLMs more freedom of action in order to allow them to collect more training data and become more intelligent? The short answer here is simple: Nope, LLM's aren't fit for that, closed source models are especially unfit for anything involving risk and responsibility, open-sourcing every model will go a long way towards a secure AI.