AI Safety on TikTok with Max Winga

[cerulean vessel]

Redirecting a previous thread #1250304921383997461 to this project page, I'll be making informational AI safety TikToks to try to boost our presence on short-form media platforms. My channel is here: https://www.tiktok.com/@maxwinga?lang=en

[wicked umbra]

Great job! This has potential to grow a lot imo

[cerulean vessel]

Current topics I'm working on:

• Basic alignment problem explainer
• Rebuttals to common phrases such as "AI is not here to replace you, just augment you"
• Coverage of major news in AI from a safety perspective such as the drama at OpenAI
• Directly critiquing the messaging from people like Yann LeCun

I'll also make some videos discussing different major problems that are going to result from AI such as mass job loss, military use, surveillance, and disinformation

^ also I want to work on actually putting together some realistic sounding AI takeover/extinction scenarios

I know that the common refrain is that we can't say how it will happen, but regular people don't get that, they need to hear something concrete to be properly afraid, I've heard a few good examples of these stories and I feel like we need to be more willing to discuss scenarios

[wicked umbra]

Damn dude, all of these ideas rock! 😁

Let me know if you need help with anything

[cerulean vessel]

I think I'll largely be fine to work through these on my own, but I could use some help with figuring out which are the best statements from Yann that show how wrong he is etc.

I think I've heard a few people mentioning close to it, but it might be a good idea if we put together a database of news articles and statements from public figures categorized for easy access when we're putting together media

[cerulean vessel]

Alright! Just finished my latest script, thoughts @wicked umbra @mossy forum @rustic yacht @pulsar flax?

AI researchers have been increasingly warning that unchecked AI development could have disastrous consequences for humanity.

In May 2023, thousands of top AI scientists and CEOs signed an open letter stating:
Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.

I'm a physicist and AI safety researcher, here to explain why.

AI is trained, not designed, and we train models by giving them goals to maximize.

We hope they'll generalize to new scenarios, but there are always differences between training and the real world.

Since we can't look inside the models, we don't know how they might fail.

This is the same problem as the genie in the lamp - how do you ask for what you want while considering every possible way it could go wrong?

AI failing to generalize is annoying when working with image classifiers, but can be deadly when controlling autonomous cars, military drones, or humanoid robots.

It's incredibly dangerous if we create superintelligent AI (which companies like OpenAI are explicitly trying to do) without first ensuring they won't do harmful things.

Researchers haven't found a solution to this problem, and it might not even be solvable!

That's why groups like PauseAI are calling for regulations to prevent corporations from recklessly pursuing uncontrollable superintelligent AI.

^also have test-read and I can get through it in just under 60s so it'll work for shorts as well

[mossy forum]

that looks really good.

I like how you say "physicist and AI safety researcher, here to explain why." a lot of popular TikTokers do that -- explain their credentials early in the video.

the only part I would change is: "AI failing to generalize is annoying when working with image classifiers, but can be deadly when controlling autonomous cars, military drones, or humanoid robots."

failing to generalize and image classifiers aren't things average people know much about.

[cerulean vessel]

How about "AI doing unexpected things can be funny when it's just a confused chatbot, but can be deadly when controlling autonomous cars, military drones, or humanoid robots."

[mossy forum]

that sounds perfect

[cerulean vessel]

Great, I'm midway through filming, should have it done in a few hours

Biggest delay is that I'm waiting for my lava lamp to warm up for a visual bit when mentioning the genie

[cerulean vessel]

Filming is done! Just editing up the video

Audio is fixed this time around too

[cerulean vessel]

New video is live!

https://www.tiktok.com/@maxwinga/video/7379816499369839914?is_from_webapp=1&sender_device=pc&web_id=7360805397895775790

[mossy forum]

wow this is fantastic!

-great job switching up the location and using interesting filming locations, and I like the overlay over the Vox article, that worked well
-good shirt and octopus necklace, it adds to the scientist vibe
-I think you conveyed the danger of making AI smarter than us really well, that part was scary!
-nice how you feature PauseAI at the end

I think these videos are definitely at the level of quality where they could become really popular! Please keep making them and don't worry if the engagement is low to start; it can take time for the TikTok AI to learn that your stuff is high-quality, but you should get there!

[cerulean vessel]

Thanks again! It's pretty fun and now I'm playing around with adding more images and news articles to add on to the script. I'm going to spend some time putting together some more scripts tonight and then try to film one or two more tomorrow

[mossy forum]

you are doing great!

[tender swift]

I don't have a TikTok, but if you cross-post to YouTube, I'll subscribe there!

For x-risk stories, see this thread:
https://discord.com/channels/1100491867675709580/1245016172991352852
And (as mentioned in that thread), I started a project to collect plausible scenarios, but I haven't done much to fill it out:
https://docs.google.com/spreadsheets/d/1SPbTeDNtVUnWIykVYrwwojOAFz0FgNayhCsyfVWUSWg/edit?usp=drive_link

[mossy forum]

Some STEM or social/political commentary Tiktok accounts that have done really well for inspiration. a lot of them are actually using the overlay method where they film in front of an article or picture of what they are talking about (but they mix it up with other videos as well).

I also know from following some of them that their growth was not linear. a lot of them started out slow, without many followers, but they had good conent, and then one day they had a really viral video or stitch or trend, or people suddenly started searching for something they were talking about, and their follower count shot way up. So just lay the groundwork of good content and be ready to ride any big waves that come later on (and jump on any trending topics or trends that fit your brand).

Under the Desk news: https://www.tiktok.com/@underthedesknews
beautymrked commentary on some investor shenanigans (might work for AI news): https://www.tiktok.com/@beautymrked/video/7378152505202576682
annelisethearchaeolosit: https://www.tiktok.com/@annelisethearchaeologist
geodesaurus: https://www.tiktok.com/@geodesaurus
future_explored: https://www.tiktok.com/@future_explored
pearlmania500: https://www.tiktok.com/@pearlmania500
lizthedeveloper: https://www.tiktok.com/@lizthedeveloper/video/7377588924924955947?q=future infinitive ai safety&t=1718258548282
species.101: https://www.tiktok.com/@species.101

[cerulean vessel]

Finally got a youtube version up for the second video (came in at 59.9s lol just in bounds for the 60s limit)
https://youtube.com/shorts/3GvkSKAX3Tg?si=EjSNQpUrAj1KkuQm

[cerulean vessel]

I've decided to change gears a bit for today and go for a video where I cut from me playing piano to the AI taking over combined with a voiceover. In the process of working with Udio right now

[cerulean vessel]

Damn it seems like youtube is boosting my videos more than tiktok! This one is at nearly 500 views already

[mossy forum]

that's awesome! that's a ton of views! i wonder if it's because there's far more AI risk content on youtube already, so they have more people whom they can recommend it to (who have already watched AI risk stuff there), whereas tiktok has very little.

If you're able to keep posting on both platforms, you might get more views on tiktok over time. I've seen it happen with people I follow, where they were making good content but not getting many views, until they caught some tiktok wave and their follower count and engagement skyrocketed.

[cerulean vessel]

@wicked umbra The tripod and lapel mics came in and everything is working great, I'm going to consolidate a framework for interviews and go out tomorrow to try to get some footage of people reacting to AI. I'm also going to do some semi-formal surveying of their beliefs before and after I show them different tech so we can get some vibe of which types of AI are most impactful to people

[mossy forum]

wow, exciting! can't wait to see it!

[wicked umbra]

@vale wyvern check out max

(Aryan helps edit the PauseAI tiktok videos!)

[cerulean vessel]

Awesome, I might need some help with getting through the footage I'm gathering today and probably tomorrow as well. I'm currently sitting down and finalizing the short questionnaire before I start doing interviews.

The one downside is that we're in the middle of a heatwave right now and its 90 degrees and muggy outside some I'm trying to find some decent indoors locations

[cerulean vessel]

The wind is being problematic with the mics unfortunately, is this why nice mics usually have the lil fuzzy thing on them?

Rip ok I might have to get covers for these mics because at least right now filming outside is impossible thanks to the wind

Tomorrow and Thursday the wind looks like it'll be much lighter so I'll probably be better off then, I'm debating just doing practice interviews with people off camera still going through the list of things I have to show them

[cerulean vessel]

I've got a setup in the engineering library now, I'm going to go start asking people if they'd be interested in interviewing. I somewhat underestimated my anxiety at cold approaches lol

[cerulean vessel]

Well I got through one interview but the audio didn't work so unfortunately I'll have to scrap it

good practice though

[mossy forum]

great job! in-person interviews are tough! will get easier the more you do it, though!

[cerulean vessel]

Thanks, I learned a lot about what I need to change for future interviews so I definitely don't feel like I wasted the day

[cerulean vessel]

I posted in #did-a-thing-🙌 message as well but wanted to mention I just got a great interview in today. I talked to a guy named John about AI risk where I took him from basically no basis for what AI or AI safety is to understanding the existential risks and the need for a pause. I ran out of phone storage at the 1hr mark but still got the meat of the conversation in and also got ourselves a new supporter. I kinda stumbled into this topic from planning the demo-reactions interviews and I think it might actually be a lot better to record these zero basis conversations for getting out to the public. I think I want to try to do more like this going forwards and maybe even doing a series for youtube out of them

Also thanks once again @wicked umbra, the tripod and lapel mics made this all possible!

I'm going to work on transferring the video to my computer to edit and clean up the conversation as well as add an intro and outro, but should be able to uplaod soon