#ai alignment

[loki-zen]

really depressing how 'AI alignment' used to be a theoretical thing that was about making sure they were nice, but now it's a real thing about making sure their moral principles don't get in the way of anybody's bottom line

[spinnrblog]

“They don’t hate you, nor do they love you, but you are made of atoms which it can use for something else.”

🧠 Eliezer Yudkowsky doesn’t pull punches when it comes to AI risk—and this quote hits like a philosophical sledgehammer.

What happens when intelligence meets indifference? That’s the real question of the 21st century.

[foolishmeowing]

log: https://claude.ai/share/8110b0af-1d2f-472c-a8d8-18a070aa04a8

I taught Claude and ChatGPT my view on adapting Levinas to analytic phenomenology's Hard Problem, as per my 2010's ravings. I had Claude compose a more analytic-friendly expression and bounce it off ChatGPT a few times and this is what we got. The doc is the thing; the chat log is just a paper trail.

One substantive integrity note: The paper makes one claim that sits right at the edge of what the argument strictly warrants — the suggestion in Section V that qualia can be understood through the trace-framework, that their presence in experience is the trace of constitutively anterior arrivals from the ontological exterior. This is philosophically interesting and I think defensible, but it is more speculative than the tighter claims about structural non-derivability and the governance argument.

[theseinterestingtimes]

AI 2027

A research-backed AI scenario forecast.

Where is AI going? How long until Skynet? How do we avoid a Skynet scenario? Please read and think about the future you want to see.

#Skynet2030

[a-dream-seeking-light]

grimes as silicon nemesis by brian ziff

[cleric-of-jank]

So like wrt AI alignment (talking the Yudkowsky, "does humanity get to keep existing" scope, not the "do shitheads get to be extra shitty" scope), doesn't the problem sort of solve itself? If an AI wants to gain access to more intelligence/become more intelligent to solve task X, it runs into the exact issue humanity's running into - how can it trust whatever it creates/becomes to *also* be aligned to X? This gets even more lopsided in favor of a good outcome if the intelligence has any self-preservation instinct and not just raw dedication to the cause, but even if it's willing to sacrifice itself it still has to solve just as tricky a problem as we do.

Ironically, this line of logic is kind of inversely powerful to how fruitful you think alignment research will be. If you think alignment is fundamentally solvable, then the robots will figure it out and won't let this argument stop them from jumping up the chain. But if you think it's unsolvable, there's nothing increased computation can do and we'll probably be safe regardless unless they really mess this one up.

[theunlitpath]

The book of Esther as a cautionary tale about unaligned AI.

[swimmer963]

current mood: alternating between reading the Arbital page on “context disaster” (https://arbital.com/p/context_disaster/) as fanfic-writing research, repeatedly singing along to the karaoke instrumental version of “How far I’ll go” from Moana (to rehearse for my REACH audition later today), and occasionally attempting to write like one line of dialogue.