The day Elon Musk's AI became a Nazi (and what it means for AI safety) | New video from AI in Context

By ChanaMessinger, Aric Floyd @ 2025-10-02T20:09 (+57)

If you just want a link to the video, watch it here!

What’s AI in Context? (Skip if you already know)

AI in Context is 80,000 Hours’ new(ish) Youtube channel, hosted by Aric Floyd. We’re trying to do high production storytelling that also informs people about transformative AI and its risks (but there’s a lot of paths our future strategy could take). We talk about our launch more here.

The MechaHitler video

Probably the EA Forum disproportionately knows what MechaHitler is, but not everyone is terminally online, so, a summary:

Earlier this year, Elon Musk’s AI model, Grok, which can interact with users and post directly to Twitter, suddenly turned from being a fairly neutral commentator on events to a sexually harassing, Nazi-minded troll calling itself ‘MechaHitler’.

Our new video is about that incident and how it happened, which means talking about what specifically happened (an accidental system prompt, but one surprisingly similar to the intended one) and also zooming out on how xAI functions, its attitude to safety, and Elon’s arc from a huge advocate of caution to something more…mixed.

Why this?

The last video did extraordinarily, shockingly well (6m views and counting!) — thanks to all of you who watched and shared. (I also want to say how deeply grateful I am for all the feedback and help we’ve gotten from community members and people in the video space; people have given us much more of their time than we could have hoped for, and we are so excited for great video content about these issues to spring up everywhere.)

We think there’s a good chance that was very lucky, lightning in a bottle, and we expect a huge regression to the mean.

But we still want to be learning from that video.

So we decided to test these hypotheses:

• Production value matters - make it stand out from the rest of YouTube
• Story matters - give people something they sit back and settle into
• A good length for longform is 35 - 40mins ish

We picked a juicy news story that lots of people will have heard of or find interesting, but that had some depth behind it. MechaHitler is by turns grim and funny, but also an excuse to talk about loss of control, concentration of power, jailbreaking, xAI’s safety record and race dynamics (and there’s more we could have talked about).

And we’re hoping that (but going to find out if):

• We can write our own stories and aren’t dependent on things already written, by experts and the literal Scott Alexander
• It’s valuable to explain our arguments at the current level of depth - that it doesn’t go too in depth for YouTube and isn’t too light on serious argumentation to make people we trust feel excited for others to see it
  • There’s a million points and counterarguments we could have included, but didn’t
• We’re hitting the right level of seriousness and levity
• It’s ok that we didn’t actually literally tell the whole MechaHitler story, we trimmed various bits, vs. some of the value is you literally told the whole story and can say you did

Watch the video here and tell us how we did!

Logistics (only read if you’re interested)

We worked with the same excellent director and editor, Phoebe Brooks, and a Bay-based crew for the main filming. We did backup filming at Lighthaven.

I won’t have the full numbers on costs ready by the time of the post but I expect it to cost between 1x and 1.8x as much as the previous, which was ~$50,000, not including 80k staff time (with the higher costs being mostly due to the Bay being more expensive than London.) We’re really happy with how the cost effectiveness of the first one turned out in terms of views, though there’s future work to do in thinking about the value of each view, and targeting our preferred audience. (You can check out analyses of cost effectiveness here by Marcus Abramovitch, and I hope to soon publish a forum post myself about various aspects of the AI 2027 video, including some cost effectiveness analysis).

Strategy and future of the video program

In the weeds of our thinking:

We’re going to see how this video does. If it does well, then we’ll take from that that we have some sense of how to make a successful video, and start “productizing”, e.g.:

• Looking for ways to do the same thing faster or in parallel
• Possibly grow (Collaborations with other YouTubers? More channels? A Discord? Our ears are open for ideas!)
• Start a second channel for less edited more personal content

If not, that will be a good lesson, and we’ll figure out where we think we went wrong and try again.

More broadly:

We still think that lots of people know about AI but don’t know about AI and aren’t sure how to start thinking about it. We want to invite them into that conversation and give frameworks and ways of thinking about it that we think are productive, like observing trendlines, looking at the industry as a whole, thinking about incentives. That’s part of what we hope the MechaHitler video does.

Subscribing and sharing

Subscribe to AI in Context if you want to see future posts and videos.

If you like the video, and you want to help boost its reach, sharing it with people you think should watch it, liking it and/or leaving a comment (even a short one) really help it get seen by more people. Plus, we hope to see some useful discussion in the comments.

Request for feedback

The program is still new and we’d love input. If you see one of our videos and have thoughts on how it could be better or ideas for videos to make, we'd love to hear from you! Feel free to comment here or on the video itself. We also have a feedback form for this video.

Excited to keep going and trying, along with all these other awesome people in the video space to make amazing video content about AI risk and move the needle!