Scrimba Launches LLM-Powered Explainer Video Tool, Demos HN.watch

Per, who founded Scrimba (YC S20) and said the company has spent the last decade teaching people to code with an HTML-based video format, wrote that Scrimba Explain connects that format to a large language model so users can make explainer videos about anything. He built HN.watch to show the technology to Hacker News readers. The videos are generated on the fly the first time a user clicks a link.

Per said the HTML approach has obvious visual drawbacks compared with diffusion models, but he listed three benefits. Generation is much faster than pixel-based video, taking just a few seconds from click to playback. The cost per video is about $0.04, excluding image generation, which some videos use and which he said quickly increases cost. The speed and low cost also make AI-assisted editing cheap and fast.

The hypothesis, Per wrote, is that moving video creation from dollars and minutes to cents and seconds will unlock new use cases. He cited several examples already in use or envisioned: a video explanation of every pull request, which Scrimba does internally; a video for every page in internal or external documentation; turning a complex article into a video in about four seconds through a Chrome extension; course creators drafting lessons before recording the final version; and personal videos such as stories for children, wedding invitations and birthday messages.

The stack is based on Imba, an open-source programming language created by Scrimba CTO Sindre Aarsæther. Imba compiles to JavaScript and interoperates with the npm and node ecosystem, Per wrote. The team also built its own sync engine, OP, and a context management system for agents, Q. Per said he feared the LLMs would struggle to write code for the stack because neither tool is in their training data and there is very little Imba there, but the models have performed well. He attributed that to the stack's density: a single declaration sets storage, sync, permissions, the user interface and what the AI sees, reducing translation between frontend, API, database and JSON where a model can make mistakes.

Instead of using React.js, Express, Supabase and LangChain, Per wrote, the company built everything from scratch, which he joked might reflect 'not invented here' syndrome. For models, Scrimba uses Gemini, GPTs, Inworld, ElevenLabs and several others.

Users can try Scrimba Explain through a Web UI, an MCP integration for coding agents, a ChatGPT Plugin and a Chrome Extension, according to the post. Per said links to all options are in the company's documentation, and he also shared a pixel-based video demonstration. At the time of the post, it had 7 points and 1 comment.

Per ended by saying he expected pushback from Hacker News, which he described as fond of text. He said the tool is not for everyone, but that many people today, especially younger generations, prefer videos over text. He asked for feedback and ideas for other use cases.