Transcript:
Earlier this year, I asked GPT-3, the new text-generation artificial-intelligence system from OpenAI, to give me some ideas for videos. There were a couple of good ones in there, but the problem was that almost all of them were fictional. Videos on this channel are usually about real things in the real world, and no-one’s managed to connect up a fact-checker to the AI yet. But that got me thinking: what about channels who aren’t limited that way? What about channels which talk about broader topics, or their own opinions, or do ridiculous challenges? Could the AI come up with new ideas for them? So I called up some folks, I gave the AI a few recent video titles from their channels, and I asked it… come up with more like that. Now, I wasn’t allowed to do live demonstrations, that’s in the OpenAI terms of service I agreed to, but ahead of each call, I got GPT-3 to generate a hundred ideas and then I picked the most interesting ones. Or at least, the ones where the AI didn’t completely go off the rails. If you look at my list, I might have 20, 30, 40, 50, 100 ideas there, but I don’t want to do any one of them. It’s not easy to come up with good ideas. Mehdi Sadaghdar runs ElectroBOOM, a channel that has really technical videos about electronics, but which is, perhaps unfairly, best known for Mehdi repeatedly shocking himself. …that created [arc sound] [bleep] [arcing continues] One thing that he had in common with all the folks I talked to: they all have a list of ideas. It might be a whiteboard, a spreadsheet, a text file, but everyone had The List. I didn’t have much hope that GPT-3 would work well with Mehdi’s channel, because surely you can’t just jam some electrical-sounding words together and get something sensible out of it. Right? The very first one that it sent out was “charging phones from a Tesla coil”. Uhh… I mean, if you have a super-powerful Tesla coil, yeah, that’s possible. So if I was able to charge my phone, or at least try to charge my phone through Tesla coil, which is not a bad idea… although, yeah, I’d have to find a disposable phone for that one. The next one was ”Lemon Lime Lightning: 2 MILLION VOLTS!” That sounds quite clickbaity …but maybe not possible. There’s one that just says ”24K Gold Plated Electrical Banana Award”. I don’t know what that is. That sounds like a good idea. It sounds like somebody tried to electroplate a banana with gold, and give it as an award to someone else? Not bad… these titles, I would say, from the clickbaity point of view, they are not actually bad, right? Let me write it down, it might be a good idea… ”Using Tesla’s Induction Motor Design to Test Solar Panels”? –buh? That feels like nonsense to me. That feels like a bunch of random words stuck together. ”How To Build A 3D Printed Plasma Speaker”. …what? Let me think about… And those were the best ideas. GPT-3 wasn’t great at coming up with things within the laws of physics. So what about for a channel that’s more about exploring concepts and interesting questions rather than really specific science? We like to take big questions and then try and reduce them down to a silly little quest. That’s Sabrina Cruz, one of the team behind the project Answer in Progress. Now I just need to get this steaming pile of garbage I call my code to actually function(!) …and her pitch sounded like exactly what GPT-3 should be good at. It nailed your video title style. Ooh! ”why buying pets is a scam”. Oh my god! Well, looks like I have a new video idea to work on. It definitely matches the correct level of clickbaitiness that we go in with, where it’s just like, hey, I have a very bold hypothesis. Let me do something that barely argues it! All right, well, it was downhill from there, ‘cos it was… Oh no. ”how to destroy your career with 1 email”. I mean, isn’t that relatable content(?) Aw geez. Do you remember that one… Google did a prank one year for April Fools? Oh my god. With the minion emojis? Now that’s an email that could ruin your career. Oh my god. Right! So what we’ve just done there is workshopped a story. Boom. Okay, I think it’d be optimistic to say either of those could be a full video, but there’s enough there that it might spark some bigger idea. ”how to fix your posture”. …sorry. Which, incidentally, yeah, I’m leaning over a laptop at the minute… No, that’s literally the video I’m working on right now. No kidding! Yes it is. I’m working on a video about why my back hurts so badly! Wow. Okay. Um… The singularity is going to happen like next week, I assume. That sounds like a success, but there’s a bit of… almost cold reading going on here? Because the AI sent out a hundred ideas, I picked the best ones, and then Sabrina connected one of those to a better idea she already had. Even the title she gave, ”why my back hurts so badly”, is way better than anything the AI came up with. So the best of the best were plausible video ideas. But I did show Sabrina more of the full list, and, well, most of them weren’t winners. These are all formatted so that they would feel like it would at least come across the table when we’re thinking about titles, but a vast majority are just, like, that sounds good. Let me think about it for a second… Right. Wait, what does that even mean? I think they’ve really captured the subtle negativity that just overflows my channel. Maybe score half a victory for GPT-3 there. What about if we go from factual exploration to opinion? I have a big Google Document of videos I would like to do, once the current ones that’ll take forever are done. And that list always gets longer and none of them get done, cos I’m always like, oh but this one should, now I’ve had a new one! So, one day I’d like to get to that list, but it’s just continuously growing. Harry Brewis is better known as hbomberguy, and he’s one of very few video essayists I know who can somehow keep my interest long enough that I’ll watch a feature-length piece about a show I’ve never even seen. But when your release schedule is a two-hour video every few months, idea generation isn’t really your production bottleneck. But I wanted to see if the system could at least come up with some suggestions. It got obsessed with video games. Well, yeah, that’s gonna happen(!) So I switched which videos of yours it was trying to riff off, and if there was one word even remotely about video games in there, everything was about video games. It’s the only thing that it can make sense of out of all of this! ”The Last Of Us Is Disappointing, And Here’s Why”. That’s really good! Oh my goodness. That’s actually the perfect video title for a video about The Last of Us, because people who don’t like The Last of Us will be like, ah, finally, validation, and people who do like it will be like, ah, I’ve got to see this. I mean, it did just nick your format. It locked on to all your formats and went, ”I know what to do with this, just sub in different names.” I should have said something like, ”ah, that’s pointlessly contrarian, who would name a video that?” Oh god, I need to save that. That might come in handy! We also have “Mass Effect 3 Was A Mistake, and Here’s Why”. …I mean, not correct, but close, yeah. That’s what I was thinking, because obviously the opinion has to match up with… It has to be contrarian enough! One that came from nowhere: ”How Bethesda Broke Me, and Here’s Why I’ll Keep Buying Their Games Anyway” That’s so specific! I wonder why it got that from my videos? That’s really good. This one I don’t understand? You might: ”On Quantic Dream, David Cage, and the Uncanny Valley.” That’s really good, actually. That’s incredible. Quantic Dream are… they’re basically trying to be films instead of games in a lot of ways. There’s a lot of advanced motion capture stuff and facial animation. So there is a lot of uncanny valley stuff in there. That’s just a straight-up good video title. Okay! I was going to jokingly say I should make one of these, but that’s probably the one I would make so far. A lot of my stuff is just whatever I thought would make for an interesting concept, or I would find challenging to explain, I think that’s really fun to try and do. If I wanted to make a video along those lines, that would be the title for it. But I probably wouldn’t make those specific videos. Some of those ideas were actually good. Like, at least two of those I could make. And now I kind of want to, just so… A video essayist not only has to stand behind their ideas, but also put in exhaustive research and be passionate about a subject. And while it could be useful to have a system that’d take a video outline and have it autocomplete in a certain style and make the arguments you want it to, text generation just isn’t there yet. Which leaves us with channels that are about experiences. Folks who aren’t creating factual or documentary or opinion videos, just plain ol’ regular entertainment. Now, I did try to get in touch with a few channels who do beauty, or makeup challenges, but I didn’t get any replies. They’re in such different worlds that I just don’t have any contacts in common. I did give the AI a list of makeup challenges, and if anyone wants to try the No Water Makeup Challenge, the Curtain Makeup Challenge, or the If I Only Had More Time Makeup Challenge, go right ahead, I just don’t know what those are. But there was one more person I talked to. And he’s an expert on getting the title and the idea right. A perfect idea is an idea that is clickable but also highly highly entertaining throughout the entire video. And a lot of people are either good at picking great videos or great clickbait, and very very very few people can pick both. It’s very hard. Because some things that are very clickable might be a moment. Like, getting punched by the strongest man in the world, right? That’s an interesting video but that’s only ten seconds. It’s not a 15-minute video. And if you try to make it a 15-minute video, then it’s a bad idea. Not like, you just take things and stretch it out, it has to be a really entertaining video at that length, while still being clickable. That is the perfect idea. Jimmy Donaldson is better known as MrBeast, and fifty million subscribers watch his channel that’s a mixture of ridiculous stunts and hyperactive philanthropy. As we talked, it was really clear just how much thought goes into the titles, the ideas, and every frame of every video. The AI… does not have that context. It’s just putting words together. The very first thing it came out with was: ”I Paid People To Take Selfies With Me”. Yeah, I can see where it would get that inspiration from, but there’s no, like, ‘oomph’. What’s the virality? ”I Bought A Haunted House”. I can see where it got it from and the inspiration, but what it’s missing is a twist. Back to the selfie one, right? If it was like “I paid people to take selfies with me”, and I looked like Shrek but I had like a knife in my hand, then it’s kind of interesting, but obviously that’s not what it’s thinking. Or, with the haunted house, I bought a haunted house and offered people a million dollars to play in it. Yeah, iIt feels like there should be an “and” in there somehwere. It’s missing that little twist, where you apply that, and boom, now it’s viral. It also had one which feels like an entire series on your channel, just put in a blender so it’s just mush and no-one wants to watch it, which is: ”I Paid People To Do Weird Things On Camera”. I’m like, that’s… That’s my channel in a nutshell! It’s exposing me! A lot of GPT-3’s ideas felt like someone trying to imitate the MrBeast channel without really understanding what humans like to watch. So while it didn’t go well, it did perform a little bit better than I expected. And I think it was Jimmy that gave the best summary: This could be very very effective in giving you something close to what you’d want, and then you just use that as inspiration to derive something good. AI works best as a co-pilot. At least, for now. Because one of the clichés about artifical intelligence is that as soon as it can do something, that doesn’t count as AI any more. Playing chess? It’s just brute-forcing numbers. Categorising images? There’s no thought there, it’s just training a model. Coming up with a list of plausible ideas, with an okay but not great hit-rate? That’s something that would have been impossible just a couple of years ago. But now? It’s just putting some words together. In a few years’ time, who knows what else it’s going to be easy to dismiss in the same way? Let me just stop tape… Wait, here, let me give you a little clip to go at the end of the video? Yeah, yeah, go for it. If you enjoyed this video subscribe to Tom Scott! He’s too nice to say it. Hit that sub button. It’ll make him really happy. Right there, tap it. All right, now you can go. I have literally never said that in my life, thank you!