liondw / signal-alignment Goto Github PK

Part of the purpose of the signal alignment project is to help everyone play catch up on the wider alignment issue.

This is just the initial outline here - requires heavy compression

Simple AI generated presentations
https://gamma.app/public/4n03bbzmrgq8nrv
https://gamma.app/public/cxwkcy3ulil26f6

Articulating the Risks of AI project

(summary of the first half of the AGI unleashed video)
https://www.youtube.com/watch?v=fKgPg_j9eF0

• Existential Risks
• Social and Economic Turmoil
• What are Autonomous AI?
• Pace and scale of AI development feedback loop
  (advancement in one area is advancement in all) (leading to AGI)
• The Control Problem (What is AI alignment)
• Incomplete Solutions
• Arguments for/ against AI alignment (an updated bad alignment take bingo could be turned into a shareable long scroll post for sharing on social media)

Also incorporate visualizations from the A.I. Dilemma talk ( This source has stronger citations, could be a better bridge to understanding for more official avenues)
https://www.youtube.com/watch?v=xoVJKj8lcNQ

• Pace and scale of AI development feedback loop
  (advancement in one area is advancement in all) (leading to AGI)
• Threats and Risks of powerful AI models
• What is a GLLMM AI
• 2024 will be the last human election
• GLLMM have new emergent properties with increased scale
  (New languages, pace of theory of mind, arithmetic and maths, research level chemistry)
• RLHF, and other alignment methods/ incomplete solutions ( a comment on the Swisse cheese approach?)
• AI making stronger AI
• more notes
• [ ]

When people open the PDFs they should be able to click the links.

Overview

• Gather research and refine definitions based on sources
• Create visual concepts and assets for later use
• Create rough storyboards containing layouts and on-screen text
• Create finalized storyboards with full assets
• Publish and gather feedback

Heuristic Imperatives Cards:

Part 1: Definition

• Title Card
• What are the Heuristic Imperatives
• Reduce Suffering
• Increase Prosperity
• Increase Understanding
• Define Heuristic + examples
• Define Imperatives + examples

Final Product in the repo
https://github.com/liondw/Signal-Alignment/blob/da86b1de16aae91bf9579a50720694c4db903936/Heuristic%20Imperatives%20Graphics/Heuristic%20Imperatives%20-%20Introduction.pdf

Give feedback into the discussions section, or to [email protected]

Overview

• Gather research and use the definition cards to build the script outline
• Generate and refine a short 1-2 minute script based on outline
• Record a voice over for the script (using real voices or a text to speech model)
• Animate a rough version using previously created definition cards
• Publish and gather feedback

Further tasks

• Implement feedback
• Refine animation

DRAFT VIDEO
https://user-images.githubusercontent.com/130222960/231064311-ce8b08fb-7476-46bd-bd27-188cdfb686fd.mp4

Script outline [DRAFT]

Heuristic imperatives provide a framework for embedding ethical principles within autonomous AI systems. These principles serve as intrinsic motivations and a moral compass, guiding decision-making, learning, self-evaluation, and cognitive control.

The Three core heuristic imperatives for AI systems are: Reduce suffering in the universe, Increase prosperity in the universe, Increase understanding in the universe

By incorporating these principles into AI systems, we aim to create machines that are adaptable, context-sensitive, and capable of navigating the complexities and nuances of human values, beliefs, and experiences while maintaining ethical boundaries.

Traditional approaches to AI ethics can be inflexible and fail to account for the diverse and evolving nature of human values and contexts. Heuristic imperatives address this limitation and provide a more ethically responsible approach to AI.

Heuristic imperatives have the potential to address the control problem of AI, foster trust, and promote individual autonomy. They can be applied across various domains and are crucial for creating ethical AI systems that serve humanity's best interests.

Through monitoring discussion online, it looks like there is a demand for more thorough examination of the application of the heuristic imperatives. On a personal note, I believe these principles are primarily an effort to provide a necessary foundation, but not a complete solution.
There are are the barriers that we need to overcome (collaboration, regulation, education/ awareness, active monitoring)

We Need a Plan project

(summary of the AGI Moloch video)
https://www.youtube.com/watch?v=8phmHSUDzRg

• Ideal Attractor states
• Define: Nash Equilibrium or Game Theory
• Define: The Moloch theory or Race dynamics
• Success Criteria (insert from timestamp 20:10)

I'm guessing this goes into recap territory?

Title card draft