Re:Former OpenAI employee quit to avoid 'working for the Tit...
by datameta on Friday July 12, 2024 @02:22AM
(#40943109)
I think there are two things being discussed here:
1) alignment as portrayed by OpenAI along with the perception of intelligence from the current SoTA LLMs
2) alignment as a field of study to prevent catastrophe
I'm speaking about the latter and I don't think labeling alignment as a total facade is conducive to starting well on the path of developing guiderails for AGI. Thank the universe that it seems LLMs cannot scale to emergent intellect, because that lets us do a dry run. In my opinion, we're currently tanking the current go-around as legislation is ever farther behind while we watch an absolute arms race between nations and corporations.
There aren't lessons to learn, besides the fact that we're wasting our time. The only risk around AGI is that people will trust it as much as they trust current AI, and then risk their own annihilation when they casually accept a blatantly hallucinated answer. It's like if HAL-9000 came preinstalled with Alzheimer's, and we're still adamant that we can trust him to operate the airlock. It's time to get real and admit that any risk imposed by AGI (or regular AI, for that matter) is a failure in implementation. The concept of alignment is a bald-faced marketing ploy intent on selling "AI safety" to people that don't know what AI even is.