by Ismaël Joffroy Chandoutis — February 2026
In late 2024, Andrej Karpathy wrote:
"It is hard to communicate how much programming has changed due to AI in the last 2 months [...] coding agents basically didn't work before December and basically work since [...] You're not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks in English and managing and reviewing their work in parallel."
I recognized this immediately. Not because I'm a programmer — I'm not. I'm a filmmaker. But I've lived through this before. Several times.
I started making films in 2006 with machinimas. No camera. No film school. No permissions. I would hijack game engines, puppeteer virtual characters, and call the result cinema. The film industry had no word for what I was doing. I had already crossed a line that, officially, didn't exist yet.
When non-linear editing arrived — Final Cut Pro, Avid, then the whole digital NLE ecosystem — the marketing was simple: cheaper, faster, democratic. And it was true. But something else happened that nobody put in the brochure.
The editing room emptied.
Before, editing was a collective space. The editor, the assistant editor, the apprentice who threaded the film and watched everything. Knowledge passed through bodies, through proximity, through watching someone else make a decision and understanding why. When editing went digital, everyone got their own machine, their own room, their own isolated workflow. The transmission broke.
This is the warning I want to write into the margins of every article celebrating the AI revolution: when a technology democratizes a practice, it often atomizes the community that carried the knowledge of that practice. We should be building intentional structures to preserve transmission — not assuming the tool will handle it.
The industry told us digital cameras were better than film. It was marketing.
For years, the honest cinematographers knew the truth: the best chain was hybrid. 35mm shot on film, scanned, edited digitally, then output back to film via kinescope. The best of both worlds. The full digital chain — for all its convenience — took a long time to actually surpass what you could do with emulsion.
I think about this every time I hear "AI is better than human creativity." The hybrid chain is almost always the truth before the full switch. The question isn't "does AI replace?" but "what's the honest hybrid right now, and what does the eventual full chain actually look like?"
My analog thinking nourishes my digital thinking. My digital thinking nourishes my AI thinking. I don't discard layers — I stack them. Depending on the project, I choose which chain to run.
Here's what people miss: the logic of agentic engineering was already there in early Final Cut Pro.
Double database. Abstraction layers. A metaphor interface borrowed from film — windows, folders, reels — that slowly gave way to keywords, smart collections, trackless timelines. If-this-then-this automation. The film metaphor was a transitional grammar, easing us from physical editing toward something more systemic. We were already building toward the agent. We just didn't have the model quality to run it yet.
Alexandre Astier — filmmaker, writer, actor behind Kaamelott — was already writing his scripts programmatically, using Python to manage the structural logic of his narratives. He was doing agentic writing before we had agents. That inspires me.
Karpathy says programmers are unlearning how to type code. I never learned to type code — I'm learning it now, in 2025, when the layer above it is already becoming more important than the layer itself. I arrived at the right moment.
But I have unlearned things. The argentique chain. The hybrid chain. The DSLR filmmaking instincts. The iPhone documentary reflex. Each technological rupture asked me to let go of muscle memory and rebuild from a different foundation — while somehow keeping what the previous foundation taught me about light, about duration, about what makes a cut true.
That's the real skill: not learning, but knowing how to unlearn and relearn without losing your artistic core. And here's the thing about artistic practice that quantitative fields sometimes miss — the process is not fully optimizable. Film is complex. You can systematize what needs to be systematized. But you cannot shortcut the artistic judgment that decides what gets systematized and what stays manual.
Do manually what needs to be manual. Orchestrate what doesn't.
When I give an AI agent a task in natural language, I am directing. I'm not typing code. I'm not operating a camera. I'm doing what I've always done: holding a vision, breaking it into shots — or in this case, into tasks — and managing the execution without touching every lever myself.
A director doesn't grip the dolly. A director doesn't load the magazine. A director holds the whole and delegates with precision. What we're calling "agentic engineering" is, from where I stand, just mise en scène applied to computation.
The new editing room is a terminal. The new footage is latent space. The new found footage remontage — which I was already doing at film school, cutting ciné-poèmes from archive material, films that were born entirely in the editing room — is now generative, exponential, and running on a different philosophy than the Lumière brothers ever intended.
The argentique chain was: light → emulsion → chemical process → projection.
The argentique-numérique chain was: light → sensor → file → grading → projection.
The chain I'm building now doesn't start with light hitting a sensor. It starts in latent space — in compressed representations of everything ever made, seen, written, filmed. Different ontology. Different philosophy. A concurrent tradition to cinema as we inherited it, running in parallel, not replacing it.
This isn't a metaphor. When I generate an image, a sequence, a scene — I'm not capturing reality, I'm navigating a space where reality has already been dissolved and reconstituted as probability. The gesture is different. The eye is different. The relationship to time and truth is different.
Kuleshov showed us that editing doesn't record meaning — it creates it. Put a face next to a bowl of soup and you see hunger. Put the same face next to a corpse and you see grief. The face itself is neutral. Meaning lives in the juxtaposition. Latent space is Kuleshov at scale: every generated image is a face, and its meaning is determined by everything placed beside it, before it, after it, through it.
Unlearning the Lumière chain isn't loss. It's the same move I made in 2006 when I stopped needing a camera to make a film. I unlearned the camera and gained the engine. Now I'm unlearning the sensor and gaining the model. The filmmaker survives each translation. Something essential carries through.
Direction. Taste. Judgment. The ability to look at what an agent produces and know — not calculate, know — whether it serves the work.
This is the formation of an artist. It's what survives every technological rupture. And right now, it happens to be the most valuable thing you can bring to a terminal.
There's a gameplay to this. An artistic gameplay. You learn the new system not as a replacement for your practice but as a new instrument in an expanding ensemble. You find the moves that feel true. You build intuition. You iterate not toward efficiency but toward something harder to name — toward work that couldn't have existed before, and couldn't exist without you.
That's where I'm working.
Ismaël Joffroy Chandoutis is a Paris-based artist and filmmaker working at the intersection of cinema, contemporary art, and artificial intelligence.