Detailed Analysis of Alignment Is Easy

https://docs.google.com/document/d/1UlC0T76Kc2C_aX_63mMyFEWa31YGPSShJPd22RGgOxw/edit?usp=sharing

let's read and comment here and make a detailed analysis

Google Docs

Commenting on Aligment is Easy

Why are billions of dollars being poured into artificial intelligence R&D this year? Companies certainly expect to get a return on their investment. Arguably, the main reason AI is profitable is that it is more controllable than human labor. The personality and conduct of an AI can be controlled...

dual-salmon•12/3/23, 5:04 AM

I was disappointed in this essay since it goes into less detail on the heart of the matter than X-conversations I've had with Nora: https://docs.google.com/document/d/1ApL-j5vw0j-i6d8U92Su9hj_IYH0fiqNAAXEsRk1DRg/edit. Most interesting bit is on the last page, which I have added since the last time I've posted this doc on the server. TL/DR: "misalignment = Goodharting = overfitting in the context of values, which can be solved with standard capabilities techniques like early-exit on training, regularization, etc."

full-green•12/3/23, 7:42 AM

I didn't comment on the original LessWrong post, because there was just too much to comment on. Having a socially low-stakes place to add as many comments as I wanted directly on the document was helpful.

Ffull-green I didn't comment on the original LessWrong post, because there was just too much...

spotty-amberOP•12/3/23, 7:24 PM

I'm glad! That was one of the purposes of this

full-green•12/3/23, 7:50 PM

Yeah, I thought it was secretly also AI-Plans.com field research, lol.

Ffull-green Yeah, I thought it was secretly also AI-Plans.com field research, lol.

spotty-amberOP•12/3/23, 9:43 PM

lol, only in so much as understanding it helps understand more about alignment

colourful-plum•12/8/23, 1:37 AM

I think it's possible alignment will be easy and likely alignment will be achieved. But that there's a significant chance it's impossible or too difficult to complete in a reasonable timeframe, and that's a dangerous risk.

Ccolourful-plum I think it's possible alignment will be easy and likely alignment will be achiev...

spotty-amberOP•12/9/23, 4:24 AM

I'm hopeful it's easier than I think, but doubtful atm

Detailed Analysis of Alignment Is Easy

Similar Threads

Similar Threads

Similar Threads