I really wanted AI video generation to work. I tried Runway. I tried Sora 2 in Azure Foundry. I tried Pollo.ai. I burned through $50 of my own money experimenting. And I have almost nothing to show for it.

Let me tell you what happened.

My Experiment

I wanted to replicate this video. But through the lens of a software developer from the 90s.

The script was easy enough to write, just use AI for that to help. Creating the character image was easy enough too. These AI tools are great at creating a single image.

I started with Pollo.ai since it had the cheapest credits. The image generation was solid, but I couldn't get the motion to work. So I moved over to Runway, and that was more promising. I was off to a good start with my first prompt working really well with the person talking in the set I designed.

Here was my prompt:

Have my character sit in an office in front of a computer and then have them turn and look directly at the camera and say "In 1995, being a developer wasn't exactly the hoodie lifestyle you see today. We didn't work from home — every morning we drove into the office,"

Here's what I got:

I was actually impressed. It built an office space straight out of Office Space without much prompting. The AI deciphered my character development and the 1995 reference and designed the whole set around it.

Sure, the typing was off, nothing was showing on the screen, there was some weird switchboard behind him that I don't think is a real thing, and the cables hanging over the cubicle didn't connect to anything (safety hazard, anyone?). He also misspoke. But I thought "This is good enough, let's keep going." That's when things went off the rails.

I couldn't get a consistent next scene. When I tried to transition to a new space with new dialogue, it got wonky:

So I gave up on the scene change. Fine, I'll just stay where I was in the beginning and have the character finish saying the script. This was the final product:

Every attempt to fix it cost tokens. And tokens aren't cheap. I was spending money just to fail, over and over. There's no way to get into a flow state like that.

After I blew through $50, I shut down my laptop and went on a walk.

The Flow State Problem

Back in college, Adobe licenses were easy to get your hands on. We won't get into those details ;). That access was everything. It taught a generation of students how to use the Adobe Suite, which we then brought into the workforce as our default tools. Giving away the software (even if it wasn't intentional) created a generation fluent in Photoshop, After Effects, and Premiere. Now that I run a studio with a team of 10+ people, it's the default suite we use. Because we're all familiar with it.

That familiarity came from being able to mess up. To experiment. To get into a flow state, which is critical when you're trying to learn anything or create anything.

Psychologist Mihaly Csikszentmihalyi identified flow as the state of complete immersion in an activity. It requires rhythm. You try something, see the result immediately, adjust, iterate. Ideas build on ideas. But with AI video generation, you're constantly interrupted. Write a prompt, wait 2-5 minutes, get something wrong, try to adjust, wait again. There's no rhythm. There's no flow. You're just pulling a slot machine lever and hoping.

And because of these tokenization models and being throttled by these AI video generation tools, you can't get into that state. Every failed generation costs tokens. Every "let me try one more thing" costs tokens. The meter is always running, which creates this weird psychological pressure that's the opposite of creative freedom.

This is probably something folks are experiencing with the dev tools as well. Kedasah Kerr posted about this recently. Luckily I have unlimited tokens working at Microsoft and I have been getting into a lot of flow states with the dev tools. Building websites, automating tools and assets, crafting and analyzing. It's been amazing.

But with video generation it's just not there yet. The cost is too heavy. You become conservative. You second-guess yourself before every generation. That's not how creativity works.

Where This Might Actually Help

Maybe Hollywood. Maybe the economics make sense when a single shot costs $100K anyway. If AI can get you 80% there and you clean up the rest, maybe that's a win.

Actually, I like how Ben Affleck is approaching this. Netflix just acquired his AI startup InterPositive. It's a small 16-person engineering team building tools that let you feed AI the film you already shot, with your actual characters, so you can generate a scene you forgot to capture or need to reshoot. That's a much more thoughtful approach: AI as a tool to fill gaps in existing footage, not as a replacement for actually shooting something in the first place.

Or maybe quick-hit promo videos under 30 seconds. If you just need a flashy clip for social, these tools might get you there. But when you're trying to produce a 20+ minute keynote, a demo session, or a sit-down interview? Not helpful. Maybe for b-roll, but I didn't yield good results there either.

In my world, shipping weekly content, working with real deadlines and real budgets, I haven't found the use case yet. The time I spend wrestling with AI generation is time I could spend actually making things.

I Want To Be Wrong

Seriously. I want this to work. The potential is obvious. But right now, in early 2026, for practical production work? I'm not seeing it.

What's been your experience? Have you found workflows where AI video generation actually delivers? I want to hear the success stories, because I haven't had one yet.