Like many people, I approach AI with a mixture of excitement and dread. There is excitement at the new technology, but a bit of dread as we don’t know where this is ending. Will AI really become smarter than us and replace us all at our jobs? And what does that mean?

The answers to those last two questions are “maybe” and “don’t know.”

I mean, I am a computer programmer, and that area has rapidly changed. You can just describe what you want and in a few minutes have working code that might have taken weeks to write (or cut and paste snippets from Stack Overflow).

But I am also a writer, and that is an area where AI has only been a marginal help. AI is still terrible at jokes, and while it has a very competent writing style for like informative essays and news articles, it can’t even attempt to match my voice for the type of eclectic writing I do.

But here is how it is changing things in a major way: If I have a funny idea, I no longer have to just write some text to express it. I can make it real in a video.

Now, I’ve enjoyed using AI video for quite some time. While many always just attack anything AI-related and peck at its current problems, my attitude has always been, “Well, knowing its limitations, what can I make with this now?” That’s why when a couple years ago I first got access to AI video generators, I went ahead and did a silent movie of the Three Little Pigs… because that’s where it was when it started out.

But it’s quite a bit further now.

For instance, I had this idea to do a bunch of hacky jokes about a woman president for The Babylon Bee, but instead of writing out some little article, I put this together one morning:

This is almost so simple now, it feels like cheating. It’s just a bunch of disconnected, couple-second clips, and all I had to do was tell AI what I wanted, and it made it (with me supplying images of the characters to keep them consistent between clips).

This, by itself, is not that threatening to Hollywood. What I don’t have here is a coherent narrative and scenes that flow into each other like in a movie. That is not easy with these tools. But it’s not impossible.

As an experiment, I did a joke horror movie scene. I used Grok Imagine as it’s the only one that never blocks you from using a celebrity. It took two mornings to make this. If I had more time and credits, I would have redone some scenes, but I think it turned out pretty well (Elon Musk even liked it — or acknowledged it existed):

I’m sure if you look at it very carefully, you can find inconsistencies with the background and such because the AI never knew what the rooms were supposed to look like — it was just going off of the still images I fed it and had no permanent concept of the locations.

But the thing is, these consistency problems are all solvable. Most AI video generators are now good at keeping a location consistent within a single video, and there has to be a way to extend that for multiple videos. And people consistency is obviously quite solved. The performance is another issue, but there are ways to better prompt that or feed in a reference video or audio for what you want.

Where this is headed is that if you can imagine it in your head, you can have AI video make it real just as you imagined it. While many artists are anti-AI, I can see this eventually being great for directors who want to completely control their vision with no compromise, as absolutely anything that can be described can be made.

And if the current economics for these AI video generators hold, we’re talking about costs capping out at thousands of dollars per production, not millions.

That’s what Hollywood needs to worry about. We’re talking no need for most of the crew. No need even for actors (though maybe for a while to feed into the AI for the performance wanted, but the actors will be basically anonymous as the AI will change them into whatever character is wanted). You just need a guy with a vision and a computer (and maybe a good video editor until AI gets good at that — which I’m not sure it will).

Really, the only thing AI is terrible at and not getting better at is true creativity. If you just tell it to come up with an interesting story, it will make something adequate, but never great (or at least it wouldn’t know it, and would need a human to determine it’s great). And all AI is still pretty terrible at humor (so there is still work for me if AI completely replaces me in programming). It’s great at executing a humorous concept and even seems to understand the humor, but just can’t come up with decent jokes itself.

So that’s where I see the future of filmmaking going: It will be a place for storytellers and people with great artistic vision. And then the computer just implements whatever is in your head.

What we’re also headed to, though, is a world of mediocre slop. As the tools get easier and easier, anyone will be able to make a film (and cheaply). And the AI will do an okay job when given open prompts by someone with no real vision. We’ll have a flood of content in the hope some of it will catch on (like you already see on YouTube with bizarre computer-generated kids content).

It’s just a question of whether the people who can make something great will be able to rise above the noise. So maybe Hollywood won’t die out because we’ll still need people with taste to separate the great from the okay. And one thing AI is still pretty bad at is showing that kind of taste.

And did another if there were no Republicans:

This was my first time really playing with Seedance 2.0 which is pretty great (can even take audio and video as reference input). It’s a bit closed down (had to use Grok’s Imagine for the few bits with someone famous), but very usable if you first generate reference people in AI to input into it. If I just had it generate people instead of giving it images of who you wanted, it would sometimes generate random celebrities (it really liked Bob Odenkirk), even though it blocks you from making videos of celebrities.

Anyway, hoping to do a bunch more of these. What would you like AI to simulate?

When I shared on Twitter, I got two angry reactions from people who obviously didn’t watch to the end: Those who warned me about getting close to wildlife and those who thought they were smart for pointing out this was actually AI.

Read more

If there is a better use of AI, I haven’t heard it.

And as a bonus, I decided to see what Sora could do with a brief scene from Sidequest. Not bad. Once it has a way to do consistently do human characters between shots (it has the tech, but it’s kind of locked), I’m going to play around a lot more with that.

A lot of people get angry at AI, but to me, it’s just a fun new medium to play with. I continue to play with Sora (which seems to “get” comedy), making funny little 10-second (though, recently, it now goes up to 15 seconds) sketch comedy while I imagine the possibilities when they finally have the tools to make longer, coherent pieces.

I had saw some AI video of cats scaring away bears that were tricking people, so I did my own variation.

And as a little bonus, here is a quick scene from Hellbender. I thought it did pretty well with this (way better than Veo which couldn’t get Lulu’s delivery right).

Do you realize the tech isn’t that far from being able to just drop a whole novel into AI and a movie adaptation would pop out?

When OpenAI released Sora, it was kind of a disappointment. When they first demoed Sora, it was amazing, but by the time they made it available to the public, many better AI video generators existed, and it was kind of useless in comparison.

Sora 2, though, is something else.

Now, Google has Veo 3, which integrates video and sound, and I have been using it to make longer, coherent videos (some examples I did for The Babylon Bee here). It was the best AI video generator. But Sora 2 also does video and sound, and so often just absolutely nails the prompt — action and comic acting.

The only problem is that it’s heavily watermarked and doesn’t have the tools yet to make matching shots like in Veo 3. But it does 10 seconds at a time, so now I am trying to master the 10-second sketch.

Here’s what I have so far (I’m posting under the user name “slowronin” if you can get a Sora invite). And since you can put yourself in these with something called a “cameo,” I just went ahead and had these all star me, since that’s easier than worrying about it making someone who might run into copyright.

Here’s a sampling of the ones I’ve made so far. Almost anything you imagine you can do… as long as you can get it down to 10 seconds.

Everyone is scared of our AI future, but at least it’s fun for now. I really feel like when this has more professional tools for making longer videos, things are really going to take off with AI production.

This one is based on a Bee article, and it’s nice how I was able to just hand it the image from the article, and it was then able to put the game right into the video.

Here’s another one based on a Bee article, and I think this turned out pretty well for a narrative-driven one with a lot of dialogue. I actually like some of the AI performances — to get decent performances, the AI needs to understand the scene, which isn’t always easy to communicate to it.

I actually had a better version of this one where I fixed an AI moment with the parachute and some of the sound, but they accidentally published an earlier draft.

And this one was published this week, the day after the Trump administration announcement. It only took a couple of hours to make, so AI is pretty good for quick reacts.

I think it’s a fun medium to play precisely because you can do about anything and turn it around so quickly. It takes compromises, though, since you can’t always get precisely what you want. I feel like it’s a bit like working with animal actors, where you just take the best performance you can elicit and adjust.

This is over two minutes long, has five characters, and is dialog-driven; it really felt like I was pushing where AI video is to its limits (and maybe a little beyond).

Getting the initial image to animate was pretty easy, as I just gave ChatGPT the five mascots and told it to put them in a support group. I also used the new Gemini Nano Banana to get an image of the Cracker Barrel Guy’s chair being empty for a later scene.

The hard part was getting some wide shots with dialogue, as the AI did get very confused with that many distinct characters in frame. Once I got a few wide shots and a few good close-ups to work with (simply prompted Veo 3 for those and then reused them), doing all the dialogue was decently straightforward.

And then Veo 3 did not keep all the voices consistent (maybe because it’s animation style and not live action), so I had to use ElevenLabs to get the voice of Uncle Ben and the Cracker Barrel Guy consistent.

And not all the line deliveries were exactly what I wanted (I wonder if, for something like this, I should maybe do performance-driven AI, but I don’t have a good way to do that with that many characters on screen). Veo 3 can do some good performances, but you really have to prompt in a way it understands what you’re going for, which can be a challenge when doing some out there stuff.

Also, on top of all this, they uncanceled the Cracker Barrel Guy while I was making it, so I had to adjust the ending.

Still, I think it’s pretty impressive for something that can be made by one person in a few hours. And only a couple of months ago, I don’t think this would have been possible.

I’m trying to see how I can push this building toward doing something longer in AI. Theoretically, a lengthy dialogue is perfectly possible as long as I have wide and close-ups to cut between (current limitation is 8s per cut if I want to use Veo 3 with sound).

It’s weird the things AI arbitrarily won’t do. It allowed me to do a video of Reagan playing a keytar, but for some reason, I couldn’t have someone say, “Look, it’s Ronald Reagan!” or even “Look, it’s the president during the 1980s!” and thus the awkward last line.

For this one, I discovered a fun fact about canned food and decided to share it.

And here’s one where I illustrate what it’s like where when the X-Men exist in a world with other superheroes:

I also had a tweet idea, but decided to just try making a video. Here’s the left’s psychology as far as I can understand it:

And I played around with trying to do a scene from one of my novels (specifically Hellbender), but the problem right now is prompting a decent performance out of the AI.

Anyway, I learned a bit, and next I’m going to try to make something longer. My main limitation is that I’m trying to limit my experimentations in filmmaking to the tens of dollars a month, which means I can only regenerate so much — but you can still make quite a bit for that money.

And with the introduction of Veo 3 fast, it was finally cheap enough for me to play with (I miss when all my AI experiments were paid for by my employer; I’m back to self-funding).

So anyway, really neat tech, but what can you use it for? I decided to play around by bringing some old tweets of mine to life. I mainly used Veo 3 fast, which the result look a bit plasticy, but whatever. And for the first three, I did image to video generating and image in ChatGPT first.

Here’s Commissioner Gordon on Batman’s most formidable foe.

Here’s Sun Tzu talking about war.

Here’s Aristotle talking about Batman.

Okay, so you can make a quick 8-second clip, but I tried to play with Google’s Flow to see if I could make something coherent longer than that.

It wasn’t easy. Somehow, it just automatically keeps the voice consistent for the same characters, but getting it to cut in and be consistent and have the right person say the right thing took some practice. I ended up using about half my month’s credits to make these 21 seconds, but hey, we learned some things.

Also, the background noise wasn’t consistent between the clips, so I had to use AI to remove it (which may have messed up the sync) and then ElevenLabs to generate something to replace it with.

And the results are… okay. Pretty okay. The first part was done with Veo 3 Quality and the rest with Veo 3 Fast.

Anyway, this fascinates me because I just know there is something interesting that can be made with this. I don’t know what yet, and there is a lot of limitations to work around (don’t try to get the virtual people to act), but man it’s… neat. And I’m determined to figure out what it can be used for.

I know lots of creatives are screaming about AI, but I think it’s inevitable that it’s going to really change things, no matter how much people yell at it. We’re talking about eventually getting shots that would normally cost millions of dollars for pennies. It’s not going away.

I did a bunch of experiments with it to see exactly what you can do with it and how much you can control it. Of course, that changed every week and is continuing to change.

Now, only one thing I did got made public, and that was the Jeremy’s Razors ad I made with Ethan Nicolle.

But I did a lot of other work leading up to that and some experiments since, and I thought I’d share it with you, the cool dudes who subscribe to Frank Talk.

Read more

So the question to me was: What can you actually make with this?

Well, I got the go-ahead and a budget (reportedly $47) to push AI as much as we could working with a small team (Ethan Nicolle and Kevin McCreary — sort of a mini-Babylon Bee reunion). And what we produced was an epic Jeremy’s Razors ad.

People have been asking what software we used. The answer is: all of it.

A lot of people are mad at this ad because… well… it’s AI, but I just see it as a new medium, and the question is, what can you do with it? We already know you can make a lot of slop, but what can you do if you take the time and get an actual artist involved?

Well, the answer, so far, is this. Though the way things are moving, I’m sure it will be way outdone by anything produced a couple of months from now. I can’t even pretend to know what things will be like a year from now.