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 somewhere. 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!
