Where this is going — AI, Skills, and What Actually Lasts
The AI conversation in VFX is exhausting because both sides miss something. The people panicking and the people saying "don't worry, it's just a tool" are both missing something.
In March 2026, Netflix acquired InterPositive for $600 million. Founded by Ben Affleck, they automate colour grading, relighting, and continuity fixes by training a model on award winning footage. Affleck said the whole point was to keep the creative direction human. Which sounds fine. But he also spent years building the tool that automates them. Both things are true, but I'm not sure the tension has really been addressed. Netflix spent what could have been decades of compositor wages on a single acquisition. That's not a subtle signal about where they think the value sits.
The way I've started thinking about it: AI is good at execution, repetitive technical work. What it can't do well yet is make creative decisions. That's the gap worth paying attention to.
What AI actually is
Most of it runs on diffusion models. You train a model by taking millions of images, making them noisy and then teaching it to reverse that process. Once it's done that enough times, it can start from pure noise and produce something that looks like it came from the training data. It's not thinking. It's pattern matching at a scale that looks like creativity from the outside.
That's why it works for roto, cleanup, matchmove, colour grading, wire removal — anything where the task is essentially looking at many similar frames and making consistent decisions. A model trained on enough examples does that faster and more reliably than a human. However, I think AI misses the mark in creative decisions. I had a comp shot where I decided you wouldn't be able to see a bear, apart from an outline in the lake's reflection, until the lightning struck and lit the bear up. AI could probably have done the comp, but I'm not convinced it would have known why showing less was the right move.
But if you trained a model on enough films by directors who understand tension and reveal, eventually the imitation gets close enough that the difference stops being measurable. Same with simulation. Right now, Houdini-level FX like cloth, destruction, fire, and fluids involve too much iteration and too many shifting creative decisions for AI to handle at production quality. You're constantly adjusting, troubleshooting, and figuring out why something feels off for that specific shot. But give it enough training data from enough FX artists, and it probably starts learning those patterns too. I think that's further away than roto and cleanup, but we will see how good AI gets and how much it changes how artists work.
Where I've used it
Tripo3D for quick meshes. It works but needs cleanup, useful for reference or blocking something out fast, but nothing production-ready. AI-generated album cover art for two artists in Bournemouth. The concepts were mine; AI executed them faster than I could have done manually. AI video gen has been fun to play around with, too, but it's still very funky looking at times. AI coding for VEX in Houdini, websites and tools. That's been the most useful, not replacing the understanding but cutting the time between knowing what you want and having something that does it. I built this portfolio website using Cursor — same idea, I knew what I wanted it to do and used AI to get there faster.
I've also used Claude and ChatGPT for feedback on sculpts, textures, and renders. It's suggested that things like proportion fixes, more subsurface scattering, and adjusting lighting. Sometimes useful. But I'm sceptical of how far that goes; it can flag that something looks off, but I'm not sure whether to trust its judgment.
Same pattern every time. AI handles execution, but someone still has to know what they want. The judgment stays human. For now.
What's actually at risk
The roles most exposed are the ones built on doing the same thing consistently across hundreds of frames — roto, paint, matchmove, and cleanup. Those are also the roles that have always been the way in. The entry-level work where juniors learn the craft. Mohsin Kazi, a compositing supervisor at DNEG, said it directly — those early opportunities are where artists traditionally learn by doing. If AI takes that work, how will people learn the fundamentals to be able to effectively control AI?
What lasts
The real version of being willing to adapt is this: understand what you're doing well enough that when the tools change, you can still move. An artist who knows why a simulation behaves a certain way can work in whatever software. An artist who only knows the buttons is stuck the moment the buttons change, and they will.
Disney invested $1 billion in OpenAI in December 2025 and licensed its entire character library to Sora. Netflix bought InterPositive. These aren't small moves — they're the biggest content commissioners in the world placing bets on where production is going.
Bloomberry analysed 180 million job postings in November 2025 and found that computer graphics artist roles had dropped 33%, while creative director roles held steady. Execution is compressing. Judgment isn't. That's not a coincidence.
I started this trying to understand why a studio with nearly thirty years of history could close in a month. The craft alone hadn't been enough to keep the lights on. Now I'm trying to work out what's worth learning, since the tools we will be using in the future probably aren't built or even possible yet.
I don't have a clean answer. But the artists who last will be the ones who understand their work well enough that the tools are secondary — and who are paying attention to know when it's time to move.
The lights are still on.