PauseAI•14mo ago

Risks Matrix - merge current and future risks

I want to visualize AI risks in one big matrix. The X axis is "capabilities" (or time) and the Y axis is "danger" or "impact".

- On the left, we have social media filter bubbles, biases. Things we've had for a whilie.
- In the middle, we have deepfakes.
- In the top right corner, we have x-risk / loss of control. A bit underneath it bioweapons / cybersec risks.

The goal is to get people to consider all risks, and how they get worse over time, and how a pause could at least prevent some of the worse things.

Ideally, we'd place this on the homepage, or perhaps the risk page.

Ideas

- Show a 'we are here' line

Questions

- How will this be visualized on mobile
- How do we prevent making people angry? How high do we place risks?
- What are good labels for the axis?

@.JamesWang wants to help out build this!

hurt-tomatoOP•11/5/24, 4:31 PM

First version looks horrible lol

full-green•11/5/24, 6:31 PM

I suspect the reason "how do we avoid making people angry based on where we place the risks" is a question to begin with is because we are trying to collapse every kind of risk into one dimension.

I don't remember which AI safety researcher or org it was, but someone outlined 2 dimensions of harm in a way that felt very objective to me:
- Personal, local, global
- Minor, major, existential

Or something like that. So without having to make judgments about which thing was worse, you could make sense of one person dying versus a billion people stubbing their toe.

If we don't have a good way of showing that kind of framework alongside the time dimension, that's okay, but we should at least consider using such a framework to generate those y-axis values in a defensible manner.

full-green•11/5/24, 6:32 PM

It would be awesome if hovering over a specific risk on the chart displayed the categories assigned for each dimension of risk impact!

full-green•11/5/24, 6:54 PM

A scope and severity matrix. That's what it was. Still don't know where I saw it, but you get the idea.

hurt-tomatoOP•11/5/24, 7:15 PM

Thanks @Nathan Metzger ! Food for thought

hurt-tomatoOP•11/6/24, 5:48 PM

Just got a different idea. One big timeline, starting from the first social media ranking algorithms to human extinction.

2010s: Ranking algorithms. Our first encounter with neural networks. Digital brains that learn what humans click on, what they keep watching. Companies fight a battle down the brainstem for out attention to serve us the most ads, that we are most likely to click on.

2023-2024: Fake media. Images and audio become hard to discern from reality. People are scammed through fake video calls. AI-written articles start to make up the majority of what is on the internet. Fake profiles and reviews are more common that real ones. Online trust is irreversibly damaged.

2024 AI warfare. Neural networks make decisions on who is a target, and who gets to live. The first fully autonomous weapons are being deployed. This is just getting started.

=== WE ARE HERE ===

Current AI systems already pose us problems that need fixing. But we still have time to prevent the next wave of problems:

Job loss. As AI becomes able to do more human tasks autonomously, so will it be able to take more and more jobs. The largest wave of joblessness takes hold, as a small number of companies amass wealth at unprecedented levels.

Death of democracy. In an internet flooded with convincing fake information, promoted to target individuals, our democratic feedback mechanism fails to function.

AI-powered weapons of mass destruction. AI models become superhuman hackers. They can find new vulnerabilities and create more potent compunterviruses that could disrupt entire economies and societies. New biological viruses are created, optimized for lethality, virality and immunorisistance. Vaccinations cannot keep up.

Loss of control. Autonomous AI models are placed in more and more positions of power, and start taking control on their own.

Human Extinction. We are outcompeted by a more powerful, capable lifeform.

hurt-tomatoOP•11/6/24, 9:02 PM

Something like this