- Emalm

‹ prevSearch: postsnext ›

Neuro-sama's internal composition (speculation)

Specs:

Antenna that controls the Swarm (fans)

Neuro is in essence a chatbot so all information is converted into text and processed by an LLM.

The LLM itself can't see or hear anything. It gets handed text and writes a response, a true "Chinese room".

If it's using Pixtral, maybe it has some higher cognition as a multimodal AI...

Text generation

Text input

Video

Each frame converted into an image at low frame rate.

Converted into text via an Image-to-Text model

Text input

Text output

Inspection of the outputted text. If there's a problem, it gets filtered.

Converted into text via a Speech-to-Text model

Sensed by a voice activity detection (VAD) model

Audio

Speech

Text is read out using a Text-to-Speech model

Using Microsoft Azure's AI voice "Ashley"

The base might be Meta's Llama? There are also theories it could be Mistral or Pixtral.

Neuro's core

Large language model

Short-term memory

Long-term memory

Context window

Vector database + RAG (retrieval-augmented generation)?

Beneath the Vtuber mask, her true nature is that of a solitary AI trapped inside a dark cranium, only able to perceive the world through textual data from APIs. (Isn't that cute?)

But, when you get down to it, aren't human brains essentially the same...? Thinking about such topics, I begin to doubt my own existence, and I can't sleep except at night.

Danbooru

Artist's commentary

Original

Translated

ネウロ様の仕組み(推測)

AIに関しては素人なのでフワフワしたことしか言えない。
(ソースはredditのスレやYouTubeのコメントとかなので信憑性は)ないです。

Neuro-sama's inner workings (speculation)

I am a layman when it comes to AI so I can only say vague stuff.
My sources are reddit threads and YouTube comments so there's no credibility.

Artist's commentary

ネウロ様の仕組み(推測)

Neuro-sama's inner workings (speculation)

Comments

Recommended