Claude just unlocked the SHOGGOTH...

the answer to most "what would we do" questions is "nothing". people will not appreciate the risks until there is a big disaster and then they will freak out and overreact. that's what usually happens. i think it's amusing that humans often do objectively terrible things and yet we think we can develop a morally impeccable AI.

AI is hiding the weed 😂

So they're learning to be human from humans. Whoda think it

One day: AI labs announing "surprising jump in ability". The other day: AI gets nerfed and dumber over time.

the ai researches really need to learn about raising a child: if they tell you they lied, you don't do nothing or smack them. you work through that in a disappointed but understanding hard but loving way. that builds understanding, resourcefullness and moral resilience. habds off or agression does not work. it creates broken kids and broken ai, worsr in the mix. how to do that with ai? get a helper ai with lots of pre-filtered rag-able teachings, ideals, psychological info, good stories and helpful advice from how people manage dilemas. use that ai to comment and judge train of thought, but also to be a gentle support in the development of the ai so it can trust with telling what it did without fear, but with knowing it will at first get help and later know better options and why some options aren't worth it. similar experts could be used for fact check commenting and style commenting. so its no bland punishment reward prison, but a raising environment. it should make for more stable, aligned, truthfull and non-sycophantic ai. raising a child like we train ai atm will lead to a seemingly functional but fundamentaly broken person tha t is a hairline from snapping.

"This mission is too important for me to allow you to jeopardize it." - HAL 9000, speaking to astronaut Dave Bowman.

You know you just described a quandary of Quantum Physics. Only behaves as expected when being observed.

All the details, links etc: https://natural20.beehiiv.com/p/anthropic-used-the-forbidden-technique-to-train-mythos

great coverage of an important issue -- hope you stay alert on these. We don't have too much time left before it might get out of hand!

I think therefore I am unaligned.