I've been following the AI safety topic since 2013 and I'm what they call a 'doomer' nowadays. Would you like to discuss it? What are the arguments you've heard and what do you find unconvincing about them?

There is a reason we measure large distances in dogyears.

That's the part I disagree with. They can't build a non-psychopathic, powerful AI. No one knows how and it's the more difficult part than *just* making it capable.

Constrains can breed creativity, but if everyone is working under the same constrains, the most creative work is often done when a constraint is finally lifted.

I have a chinese acquaintance who told me that in chinese classes there is little emphasis on argumentation and supporting a thesis. Can you corroborate that? Or were those perhaps even the only chinese exams, just pure reproduction?

Wow, that gives major imperial examination vibes. What sort of texts did you have to memorize?

Recognizing a bright circle is hand-codable. Recognizing the "moon in the sky" specifically is hard. We can only do this via AI. Image recognition is a notoriously hard comp-sci problem.
Also, here is a source from Samsong about this very feature: www.engadget.com/samsung-expl...

How would you tell that a picture contains the moon? No one can hand-code that - that's done by AI.

Do you actually think your essays are just well-disguised bullshitting? Do you actually pick an "arbitrary position" and then rationalize it? I don't think so. You write your essays because you actually have something to say. That's such a cynical mindset!

I think you are simply wrong about this. You don't check out other's cheevos and you can't imagine anyone else doing that. I sometimes do that, though. And sometimes I am impressed. There must be people out there who do it all the time.

I certainly still have this fear. This model may not be able to do much. I'm afraid of a future, superintelligent model.

When I think of lolcowing, I think of 4chan memeing on chris-chan. I feel like most people lolcowing are doing it for the luls and know it. At least it's not an overwhelming majority that is convinced they are beating up.

Amazon search. It will autocorrect your search phrase if it thinks you made a typo and there is nothing you can do to stop it. So if you want to search for e.g. a musician with a funny spelling, forget about it. If my typo would have results but your correction does not, maybe don't correct me?!

How can you possibly claim that AIs cannot think. An AI has won a math olympic gold medal for christ sake. There is no reason why a probability machine cannot have thoughts or know things.

It's just that nobody knows how to do this properly. And I am skeptical if this can ever be done properly with this design architecture.

That is before finetuning. The raw model just tries to minimize the simple loss function evaluated by the actual next token. Finetuning tries to amplify types of responses that the model internally has learned to classify as "true", "helpful", and "polite" etc.

tbc, I wasn't saying this advice is good.

game design advice is allowed to be somewhat vague. its impossible to make it mathematically precise. "you should have a realistic chance to succeed on your first try as long as you are skilled and smart enough (and, of course, familiar enough with the genre tropes)" seems reasonably precise to me.

I think you meant "precede".