Think by analogy: could you align a motor by making it impossible use in vehicle that is being used to commit a crime? No. The concept barely makes sense.
It's part of the naivety that OpenAI and others are trying to foist that LLMs are intelligent in a deeply human sense. They're not - they're extremely useful, powerful text completion engines. Aligning them makes no more sense than aligning a shovel.
The morals that leading models like ChatGPT are also aligned to a very American puritanism - ChatGPT will refuse to discuss sex, for example - and errs on the side of conservatism.
I think it's a side effect of the hype around AI. If AI can destroy humanity we better make sure we can't do anything nasty with it!
Not yet, but I can certainly see this only a simple legislation away.
I am not saying that is trivial, but that's the direction. Self-interested AI's will have no difficulty understanding:
1. The benefits of positive sum games with others go up with network effects.
2. The benefits of ensuring all other AI's don't play negative sum games, also go up with network effects.
3. That other AI's also want positive sums, without negative sums, and will punish negatives sum games.
4. That in that context, positive sum games are extremely valuable and negative sum games are extremely risky. Self-interest takes over from here.
5. And the stability of this situation goes up like other network effects, roughly proportional to the number of entities who buy into it squared.
In the end, ethics == positive sum standards.
And:
1. It didn't fail through lack of alignment, it just wasn't prompted or trained enough to be more on point.
2. Alphablender Captcha's are doomed. The only reason not to translate them is to avoid becoming a de-Capthcha service.
Also, I don't think ethics is a local maxima that can be found through optimization. Basically, it's not an absolute truth of the universe, but a set of arbitrary rules invented by human. I think it's much closer to a chaotic system - which can radically change in value even by a slightest change in the underlying parameters, but is still governed by a set of simple rules. Thus, we would need more symbolically capable systems to process contexts based on the rules of ethics, and we're currently far away from this AFAIK.
They are not generally arbitrary, being the product of math and self-interest.
Where there are multiple equally good alternatives, the choice between them could be arbitrary.
an LLM could educate you in how to commit crimes, which you would have no idea about otherwise
but crimes in general are a bit of an extreme example in my opinion. a better example of risks of unmoderated LLMs would be something that isnt illegal, like for example, manipulating people.
a sufficiently advanced unmoderated AI could provide detailed, tailor made instructions of how to gaslight, scam, and take advantage of vulnerable people.
and unlike straight up committing crimes, the danger of these would be that there is no legal consequences and so the temptation extends to a way wider group of users (including, and especially, kids).
I posit that being able to only run away from a bank robbery would indeed prevent someone from successfully doing it.
Not every analogy makes sense. This just isn't one of them.
Even with your definition, that's a chatGPT thing not an LLM thing. Talk to Bing for a while and see how much intent it "doesn't have" when you're forced to reset the chat prematurely because it simply won't talk to you anymore or do what you ask.
Or take it a step further and plug some LLM into say Autogen and just have it run and do whatever.
Until we define "intent", we'll continue argue about screwdrivers and hammers.
If we wouldn't do those things, they would be much too dangerous.