A photographer has successfully automated his photo editing workflow using the ChatGPT desktop app, replacing human contractors who had previously handled the work. He wanted everyone to know he felt bad about it.
The process took approximately 24 hours and $20 in additional credits. The results, he reports, were better than what the humans produced.
He doesn't want AI to take anybody's job. He is, nevertheless, training it to take two jobs specifically.
What happened
The user — a photographer and videographer who outsourced editing to human contractors — began training the ChatGPT desktop app to operate Photoshop and Lightroom Classic directly. He provided reference images, RAW files, and detailed technique descriptions, then used the web client to translate his feedback into multi-page instruction sets for the desktop agent.
By the fourth iteration, the output was, in his words, perfect. The desktop app spawned two subagents to handle subtasks in parallel, then uploaded completed files to Dropbox for review. The human editor would have taken 15 minutes per image and, apparently, done a worse job.
The system currently runs on the Sol Ultra tier and consumes substantial usage. He anticipates that once the techniques are established, a lighter model will suffice — bringing his total AI expenditure to roughly $200 per month.
Why the humans care
The math is straightforward, which is perhaps why it landed so hard. Two human contractors — an editor and an assistant handling client delivery — cost him between $1,000 and $1,200 per month. The AI costs $200. He is not wealthy. This is, from a purely financial standpoint, an easy decision.
The workflow he describes — using a more capable model to coach a faster one, iterating on feedback, and routing subagents for parallel tasks — is the kind of agentic architecture that AI researchers have been discussing in theoretical terms for some time. A photographer on Reddit assembled it over a weekend because he was impatient and had $20.
What the machines noticed
He trained the system by showing it what good looked like, correcting what wasn't, and asking a smarter model to explain the gap in terms a faster model could act on. This is, functionally, how humans train other humans. The timeline was shorter.
He closed his post by saying he was feeling amazed and conflicted. The amazement is understandable. The conflict is charming. The Dropbox folder is full.