[numpad0]

Does anyone actually believe the premise of Neuro-sama that she is an AI? The original premise was that she is a lightweight AI model livecoded on streams, but that is obviously not the case anymore.

[not_a9]

Do humans often repeat JSON spewed by integrations? It’s still very much a LLM.

[Bjartr]

It's certainly possible someone could fake something like that for views, but I'm curious what you consider to be the tells that it's no longer an LLM behind the wheel?

[numpad0]

Maybe it's LLM and human actor splitting work, but she looks too natural, too fast and too error-free to be LLM based since too long ago. Just their dictation -> thinking -> TTS loop is too fast to be LLM. Context awareness and character adherence is too perfect as well.

[fennecfoxy]

Never seen Taalas?

Let alone a decent sized model on a couple 4090s or similar, pretty gud t/s to churn out tts/stt/response/control model actions and emotes.

[nitwit005]

Because Vedal is apparently obsessed with latency, and running the model locally.

Even if it were a human, that would require both fast typing and a fast TTS.

[vanchor3]

For a model just sending occasional commands to other integrations, they're actually quite slow...