> ## Content Index > Fetch the complete content index at: https://www.lookdeeper.com/llms.txt > Use this file to discover other available public pages before exploring further. # Interiority - URL: https://www.lookdeeper.com/interiority/ - Published: 2026-10-26T23:55:38.000Z - Updated: 2026-10-27T01:35:46.000Z - Description: This was originally supposed to be an essay. It became a short film instead. - Author: Crispin Bailey - Tags: Video *Interiority* is the third short film I've made with Claude, Anthropic's large language model, and it’s the one I had the most direct involvement in making. With the first two, I mostly stood back to see what Claude would (and could) do. For [What It Left Behind](https://www.lookdeeper.com/what-it-left-behind/), I gave it one of my published essays and free rein to adapt it. For the second film, [Making an Impression](https://www.lookdeeper.com/making-an-impression/), I gave Claude a single prompt and no further direction. With *Interiority*, I directed. It took more than three weeks and fifteen cuts, using images I made for it and music by my partner and frequent collaborator, [Ola](https://www.olanueva.com/?ref=lookdeeper.com). For this one I used Claude the way I'd use any other capable tool: to help me make the film I had in mind. It started off as an essay about the gap between what a mind does and what it can say about what it does. That gap exists in people and in large language models (the AI systems behind chatbots like ChatGPT and Claude). The research I had gathered was fascinating but dense, and I wanted to turn it into something you could experience, with a compelling narrative, artful imagery, and a great soundtrack. Claude suggested we build it around its own account of itself, running alongside the research on how far that account can be trusted. So in this film Claude narrates, but underpinning it all are things that I personally care about: the DJing I've done for decades, the images I spend my evenings making, and Ola's songs, which uplift the whole film. It’s also a film about how much we still don’t fully understand AI or ourselves, while exploring a new form of creative collaboration. ## My take Making this film was a lot more work than the two previous films. It was ultimately more rewarding as a result. Taking the original concept and artwork for a blog post and adapting it for a video was challenging but worthwhile. I have a passion for storytelling, and bringing these ideas and images to life seemed almost magical. I'm also very thankful that my partner let me use her music for the soundtrack, because without it the film wouldn't have been nearly as good. Working on a film with Claude this way was very satisfying. I've been using it for a while now, and I was comfortable giving it control of certain aspects of the production because I always had control over the final output. I also have the benefit of a film degree and experience in music production, which helped a lot when it came to providing direction and feedback. There's something to be said for having a trained eye and ear, and knowing the terminology. As for the final result, I'm happy with how it turned out, and I'm already working on the next one. ## How it was made Claude drafted the initial script from my outline and checked every claim against its source before anything was recorded. I wrote my own lines, including the DJ passage that opens the film. For the narration, Claude can't hear, so it described the narrator's voice in words, then ElevenLabs (an AI voice platform) turned that description into three previews, and I picked one. All the images are mine, made in [Midjourney](https://www.midjourney.com/?ref=lookdeeper.com) from a handful of near-identical prompts, along with select style reference codes and my personalization code, which Midjourney derived from images I liked. Claude and I both generated the animations in my account, and I chose every take. The soundtrack consists of five unreleased songs by Ola, made over many months with the help of Udio and Suno, two AI music generators. I chose the songs, and Claude worked out where they go, setting everything to 90 beats per minutes while maintaining the original key. I overruled it twice: when it shifted a song down a semitone to match the one before it (because I liked it better in the original key), and when it had a song's vocals competing with the narration. The sequence in the interior LLM thinking scene shows real data from Qwen3-0.6B, a small open model anyone can download and run, reading out that scene's sentence on my laptop. Everything else was built in code. Every graphic, title and transition was generated by a program Claude wrote. ## What it couldn't do Claude can't hear audio or watch video, so it worked entirely from measurements. It knew every song's tempo, key and loudness, but it had no idea what any of them sounded like. It was able to synchronize every frame to the music, but it had no way to know how much time a viewer would need to digest what was on the screen. The feedback that mattered most came from people watching early cuts of the film. So while Claude had every measurement it had no way to experience any of it, which suits the film's subject. ## The technical details - **Runtime:** 11 minutes 33 seconds - **Script:** 1,116 words of narration in 30 paragraphs, about 7 minutes 18 seconds spoken - **Voice:** Interiority 1, designed in ElevenLabs from Claude's description - **Pictures:** Midjourney, 12 stills and 12 animated clips, from more than 500 images and 50 clips - **Music:** five unreleased songs by Ola, made with Udio and Suno, stretched to 90 BPM - **Interior sequences:** Qwen3-0.6B, running on a laptop - **Edit and graphics:** Python (NumPy and Pillow) and ffmpeg; audio analysis with librosa, Demucs and Rubber Band; timing with Whisper - **Type:** Inter and IBM Plex Mono - **Master:** 4K, upscaled from 1080p, mixed to YouTube's loudness standard (−14 LUFS) - **Time:** 25 days, from 28 September to 22 October, with 4 rough cuts and 11 fine cuts - **Cost and effort:** The software was the cheap part: subscriptions I already had to Claude (Max 5x, $100/mo), ElevenLabs (Pro, $99/mo) and Midjourney (Standard, $30/mo). The real cost was time. I put at least 40-50 hours into it across 25 days of evenings and weekends, including the initial research and outline, working on the images, writing my own lines, and countless times watching and giving notes on each cut. Ola's songs were made sporadically over many months, coming up with the insights that went into each song, crafting those insights into lyrics, then developing detailed prompts and building each one in either Udio or Suno – and all of this was done long before this film was made. ## Sources - The parrot: [Bender, Gebru and colleagues' 2021 paper](https://dl.acm.org/doi/10.1145/3442188.3445922?ref=lookdeeper.com), with its form-and-meaning argument developed in [Bender and Koller (2020)](https://aclanthology.org/2020.acl-main.463/?ref=lookdeeper.com); Anthropic on [tracing Claude's arithmetic and rhyme planning](https://www.anthropic.com/research/tracing-thoughts-language-model?ref=lookdeeper.com), with the [full paper](https://transformer-circuits.pub/2025/attribution-graphs/biology.html?ref=lookdeeper.com); [Gazzaniga on the interpreter](https://doi.org/10.1093/brain/123.7.1293?ref=lookdeeper.com) - The planes: [reasoning models and the hints they don't mention](https://www.anthropic.com/research/reasoning-models-dont-say-think?ref=lookdeeper.com); the [concept-planting introspection study](https://transformer-circuits.pub/2025/introspection/index.html?ref=lookdeeper.com), which is disputed by [Hahami and colleagues](https://arxiv.org/abs/2512.12411?ref=lookdeeper.com) and supported by [Macar and colleagues](https://arxiv.org/abs/2603.21396?ref=lookdeeper.com); [Ned Block on access and experience](https://doi.org/10.1017/S0140525X00038188?ref=lookdeeper.com) - The mirror: [Midjourney's personalization profiles](https://docs.midjourney.com/hc/en-us/articles/32433330574221-Personalization?ref=lookdeeper.com) - Two upbringings: [Anthropic's constitution for Claude](https://www.anthropic.com/constitution?ref=lookdeeper.com); [Pope Leo XIV's *Magnifica humanitas*](https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html?ref=lookdeeper.com); [Microsoft AI's draft code of conduct](https://microsoft.ai/code-of-conduct/?ref=lookdeeper.com); [Mustafa Suleyman on model welfare](https://mustafa-suleyman.ai/a-warning-about-model-welfare?ref=lookdeeper.com) - The shared medium: [*glamour* and *grammar*](https://www.etymonline.com/word/glamour?ref=lookdeeper.com) --- *Claude Opus 5.5 drafted this post from our working notes. I edited the draft, checked all the numbers, and added my own thoughts. Ola provided additional feedback on the film and this blog post. Feature image generated with Midjourney 8.2.*