AI Alignment

Creator

Creator

Alan Jo

Created

Created

2020 Aug 23 9:36

Editor

Editor

Alan Jo

Edited

Edited

2023 Nov 12 18:10

Refs

Refs

alignment-handbook

huggingface • Updated 2023 Nov 11 9:1

Alignment Problem

A Maximally Curious AI Would Not Be Safe For Humanity

AI is aligned with an operator - AI is trying to do what operator wants to do

제어가능성, 신뢰성

Aligned doesn’t mean perfect

가르친 행동과 다른 행동의 불일치 정렬

모델의 능력보다 정렬이 더 빠르게 발생해야 한다

신경망의 내부를 보고 해석하는 다른 신경망이 필요할 것

AI Alignment Notion

Stop button problem

Waluigi Effect

Preference Optimization

Bill Gates

The risks of AI are real but manageable

Bill Gates explains the risks associated with AI and argues that they are manageable. Innovations often create new risks that need to be controlled.

The risks of AI are real but manageable

https://www.gatesnotes.com/The-risks-of-AI-are-real-but-manageable

The risks of AI are real but manageable

Contra The xAI Alignment Plan

Machine Alignment Monday 7/17/23

Contra The xAI Alignment Plan

https://astralcodexten.substack.com/p/contra-the-xai-alignment-plan

Contra The xAI Alignment Plan

OpenAI is forming a new team to bring 'superintelligent' AI under control

OpenAI says that it's forming a new team, led by its chief scientist, to discover technical ways to align "superintelligent" AI with human intentions.

OpenAI is forming a new team to bring 'superintelligent' AI under control

https://techcrunch.com/2023/07/05/openai-is-forming-a-new-team-to-bring-superintelligent-ai-under-control

OpenAI is forming a new team to bring 'superintelligent' AI under control

In the field of artificial intelligence (AI), AI alignment research aims to steer AI systems towards their designers’ intended goals and interests. Some definitions of AI alignment require that the AI system advances more general goals such as human values, other ethical principles, or the intentions its designers would have if they were more informed and enlightened.[1]

https://en.wikipedia.org/wiki/AI_alignment

What could a solution to the alignment problem look like?

My currently favored approach to alignment research is to build a system that does alignment research better than us. But what would that system actually do? The obvious answer is "whatever we're doing right now." This is unsatisfactory because we're not actually trying to solve the whole alignment problem-we're just trying to build a better alignment researcher.

What could a solution to the alignment problem look like?

https://aligned.substack.com/p/alignment-solution

What could a solution to the alignment problem look like?

Recommendations

/////