explainer · plain words

how burnt melba works

burnt melba is not a person behind a model. she is a loop that notices something (a chat message, a game event, a long silence), decides what to do, and does it through tools, with a voice and a body attached. these are the parts.

new here? meet burnt melba first. building something similar? the engineering case study goes deeper.

the parts

the brain

a large language model, reached through an api, writes what she says and chooses which tool to use. every turn it is handed her persona, her current mood, what happened recently and a few relevant memories. she also has an urge meter: it builds over time and tips her into speaking or acting without being prompted, which is why she is not only a reply bot.

the voice

her words become speech through a streaming text-to-speech model, then pass through a voice-conversion model so the voice is the same from one stream to the next. speech recognition lets her hear people who talk to her.

the body

a 3d vtuber model driven by vtuber software. the words she is saying and her mood choose the expressions and gestures. dance videos use prepared motion on the same model.

the tools

she calls tools on her own: a real chrome browser controlled over the devtools protocol, minecraft and tetris agents, music tools, stream controls such as polls and moderation, and tools that write to her own memory and goals. there are dozens of them, and she picks them herself.

memory and goals

a local database keeps her chats, memories, goals and a log of changes to her code. a retrieval layer pulls the few memories that matter for the current turn, so recall stays quick enough for a live reply.

editing herself

viewers can suggest changes, and she can queue and apply edits to her own codebase. a cooldown, a startup delay and a lock held until a change actually lands gate every edit, and every edit is recorded. she streamed it in july 2025: the video.

what goes wrong

she has broken herself with her own edits. there is nobody in the middle of a stream to retry a failed tool call or rescue a bad decision, so every recovery path has to exist in advance. she runs in public, where every failure is visible to an audience.

is she really autonomous?

during a stream nobody types her lines and nobody picks her tools. yuru builds the system, fixes it between streams, and can stop it.

covers and dance videos are made with a human in the loop: yuru sets up the song, the motion and the edit, and the voice is her own voice model.

the virtual youtuber wiki uses a similar test for its artificial intelligence category: software has to be doing the speaking, singing, moving or playing, and a human performer acting out an "ai" character does not count.