Infographics - argument maps

Fitting complex discussions and arguments into simple graphics is a great way to spread information and explain things.

I'd like to work on these visualisations. I already made one flowchart / argument map to visualize some core disagreements in the AI debate.

How you can help

- Share ideas for visualisations
- Give feedack on the ideas and execution
- Create or improve designs (they are accessible here: https://www.figma.com/community/file/1233064002969152026)

nursing-lime•8/14/23, 10:13 AM

The "they are accessible here" doesn't have a link at the moment

Nnursing-lime The "they are accessible here" doesn't have a link at the moment

hurt-tomatoOP•8/14/23, 11:00 AM

fixed!

nursing-lime•8/14/23, 11:15 AM

I think a good infographic might be based on this: https://www.lesswrong.com/posts/BTcEzXYoDrWzkLLrQ/the-public-debate-about-ai-is-confusing-for-the-general Seems like a useful framework.

Though I'd probably call the sides "AI accelerationism", "AI ethics" and "AI alignment"

The "public debate" about AI is confusing for the general public an...

Summary of Argument: The public debate among AI experts is confusing because there are, to a first approximation, three sides, not two sides to the d…

dual-salmon•8/15/23, 1:47 AM

Here is the start of a more comprehensive mind-map. I've gotten it started on some content that is hopefully enough to give a clear idea of where it is going. Let me know if this looks helpful and I can expand (or make sure others here have edit access)

https://www.figma.com/file/q1jTwOez837RG3SjGcfvlv/AI-Safety-Debate?type=whiteboard&node-id=0%3A1&t=TxqO0sSx0LPVxphR-1

FigJam

AI Safety Debate

Created with FigJam

full-green•8/15/23, 6:39 PM

I appreciate the "WTF" box.

Ddual-salmon Here is the start of a more comprehensive mind-map. I've gotten it started on s...

distinguished-brown•8/15/23, 7:02 PM

This is good mate. I would have gone into a comprehensive debunk of "human extinction is good" but the WTF is a far better choice

dual-salmon•8/15/23, 8:46 PM

Added nodes 21-27

Hhurt-tomato Fitting complex discussions and arguments into simple graphics is a great way to...

awake-maroon•8/16/23, 12:50 AM

When I last tried to work with the Figma account, what happened is it made a copy and everything I added was only visible to me. Can others see things added to that page?

awake-maroon•8/16/23, 12:55 AM

I think we could really use a way to share Figma stuff, like a list of people's pages, or something.

Ddual-salmon Here is the start of a more comprehensive mind-map. I've gotten it started on s...

nursing-lime•8/16/23, 12:56 AM

I believe Eliezer's name is Eliezer Yudkowsky, not Yudkowski (which is a surprisingly common spelling)

Ddual-salmon Here is the start of a more comprehensive mind-map. I've gotten it started on s...

primary-violet•8/16/23, 5:44 AM

Thank you! This is very helpful

Aawake-maroon When I last tried to work with the Figma account, what happened is it made a cop...

dual-salmon•8/16/23, 6:16 AM

Could you link to that again here? As I remember, I could view but not edit your board, but that was a while ago.

Ddual-salmon Could you link to that again here? As I remember, I could view but not edit you...

awake-maroon•8/17/23, 5:33 PM

At the moment, I only use FigJam, and only for video concepts, which is on hold for me anyhow. Unless I can swap files through Figma and have things directly edited there, it's easier for me to work with Inkscape. Really I'm wondering how much I should make an effort to do learn more Figma and do things there.

skinny-azure•8/18/23, 5:15 AM

A few university organizers wanted a flyer based on CAIS's catastrophic risks page: https://www.safe.ai/ai-risk

AI Risks that Could Lead to Catastrophe | CAIS

There are many potential risks from AI. CAIS focusses on mitigating risks that could lead to catastrophic outcomes for society, such as bioterrorism or loss of control over military AI systems.

Sskinny-azure A few university organizers wanted a flyer based on CAIS's catastrophic risks pa...

hurt-tomatoOP•8/18/23, 8:42 AM

How about something like this?

skinny-azure•8/19/23, 2:59 AM

Thanks, just shared it with them on the EA Groups Slack #ai-safety-field-building channel

hurt-tomatoOP•8/25/23, 3:51 PM

Thoughts?

hurt-tomatoOP•8/25/23, 4:17 PM

skinny-azure•8/25/23, 8:33 PM

I'm a bit confused because "ban autonomous agents" seems much less likely than levels 3 and 4

Sskinny-azure I'm a bit confused because "ban autonomous agents" seems much less likely than l...

hurt-tomatoOP•8/25/23, 9:12 PM

It's not ranked by implementation likelihood, but in what degree they protect us. I should probably clear that up in some way. Ideas?

skinny-azure•8/25/23, 9:24 PM

why would banning autonomous agents not protect against extinction though?

skinny-azure•8/25/23, 9:24 PM

that would essentially ban AGI and pre-AGI systems too

Hhurt-tomato Fitting complex discussions and arguments into simple graphics is a great way to...

sacred-rose•8/27/23, 1:30 AM

I disagree that yudkowsky and the other guy are against a pause, to think is not pausible is not being against, and the reason EY said a pause "wasn't enough" was because the one proposed was of just 6 months. Stopping/ shutting it all down is the same as an indefinite pause.

Sskinny-azure why would banning autonomous agents not protect against extinction though?

hurt-tomatoOP•8/27/23, 3:55 PM

Valid point. I think I'll remove it.

hurt-tomatoOP•8/27/23, 3:55 PM

I'd still like to visualise in some way that if we only do lvl 2 or 3, we're probably doomed.

hurt-tomatoOP•8/27/23, 3:56 PM

Accompanying text:

We're introducing the 5 levels of AI regulation.

AI can be regulated at different stages of the pipeline, before training begins until after deployment. The later in this pipeline we regulate, the higher the risks we face. The higher the level the greater the safety.

At level 5, we're regulating hardware & algorithms. This includes limiting distribution of hardware used for training (e.g. GPU tracking), or disallowing the publication of novel training architectures. At this level, we're making it not just illegal, but also very difficult to train dangerous models. This gives us the best protection against the risks from AI.

At level 4, we're regulating training runs. This includes regulations like setting a maximum number of permitted flops for training new models, or requiring proof of safety before granting permission to train a certain model. If we only regulate at level level 4, training dangerous AI models is illegal, but it is still achievable because the hardware and algorithms are available.

At level 3, we're regulating deployments. This includes measures like red-teaming. If we only regulate on level 3, we're still allowing dangerous training runs, so accidents at AI labs can still happen.

At level 2, we're regulating the usage of AI models. This includes things like banning autonomous agents. If we only regulate at this level, dangerous AI models can be proliferated to a high extent. We're now depending on every single individual who has access to models to use them safely.

At level 1, we're not regulating anything. That's the situation in most countries right now. We're completely reliant on individuals and companies acting responsibly.

We'll probably need regulations at all these levels.

skinny-azure•8/28/23, 1:09 AM

makes sense! one thing is that the abbreviation x-risk may not be well-understood outside of EA/rationalism, so you may want to expand it into "existential risks" or "extinction risks"

sacred-rose•8/28/23, 4:15 AM

english is not my first language so maybe i don't get what "later in pipeline" means, but shouldn't be the first levels the ones more to the left?

Sskinny-azure makes sense! one thing is that the abbreviation x-risk may not be well-understoo...

hurt-tomatoOP•8/28/23, 6:04 AM

Good point!

Ssacred-rose english is not my first language so maybe i don't get what "later in pipeline" m...

hurt-tomatoOP•8/28/23, 6:05 AM

The AI pipeline is basically from training to deployment, so earlier refers to training e.g., thats also why these higher levels are more to the left. I can see why its confusing though. Not sure how to solve it..

sacred-rose•8/28/23, 6:44 AM

ah, got it.

Ssacred-rose ah, got it.

hurt-tomatoOP•8/28/23, 7:01 AM

Is this more intuitive?

sacred-rose•8/28/23, 7:16 AM

Not really, the thing I interpreted was later in the "pipeline of what we should push for" if that makes any sense, that was why I though it was wrong. tbh I don't know if the axis is necessary

nursing-lime•8/28/23, 10:30 AM

Yeah, I was confused to - I think if you want to show a pipeline through to deployment, you'd need to have a lot less overlap of stuff. Something like this makes it more intuitive for "later in pipeline" than what you've got.

nursing-lime•8/28/23, 10:32 AM

Note, I'm not saying this is a good layout. It's probably bad. But if so it may mean trying to include too much stuff in one graph - it may be better to make a pipeline-focused graph of "places we can intervene" (There, I would have hardware at the top and usage at the bottom and say the more of the pipeline we regulate, the better)

hurt-tomatoOP•8/30/23, 1:57 PM

Ended up removing the visual hierarchy for the process, it was indeed to confusing.

https://pauseai.info/4-levels-of-ai-regulation

4 Levels of AI safety regulation

A framework for thinking about how to mitigate the risks from powerful AI systems

Infographics - argument maps

How you can help

Similar Threads

Similar Threads

Similar Threads