People have been using LLM coding agents to make animated music videos. It’s been mainly Claude Opus, though you can find some Grok ones out there too. I liked the lyrics in Nothing Went Foom!, entangled though they are in the comic book apocalypticism of a certain peculiar Californian valley.

The fascinating technical thing here is that Claude doesn’t have image or music generation models, and Anthropic hasn’t to my knowledge ever played in that space. Instead, Claude is prompted to make animation scripts using web libraries and artifacts. From its vast general knowledge, and the evolving scaffold of agential harnesses, it infers things that might look good narratively and geometrically.

So of course I got it to make a notlob trailer in notlob. Who could resist?

Source here.

(Slight caveat: the context window bar graphs outputs are real measures on pn-chomper, but systematic experiments are a work in progress.)

Under the hood, the main teaser.lob file lays out the script as a narrative with interleaved code and examples. Instead of creating thin simulation artifacts purely for animation purposes, it imports both notlob and pn-chomper directly. The name-graph shown is the output of notlob graph called at rendering time. The beautiful manim library is used for visualisation.

Technologists like my old colleague Ned Lowe have been pointing out for a while that dropping an LLM in the middle of an operational process is very different to getting an LLM to build the product. The recent mathematical proof grenades (which depend on Lean) and the music video trend are examples of the same pattern. There are curious early signs that “make the agent write code and iterate” may not be a trick local to software, but more of a general twenty-first century industrial process.