/vt/ - Virtual Youtubers

>>60177583
You know the "select images that represent a cat/car/etc" sort of captchas (and for audio, the audio captchas 4chan used to have?). What you are doing with these is build a gigantic database of labeled audio/graphic to text, which you can later use to train a neural network. Noise included, be it graphical or audio distortions. So it was a matter of time that tech would become good enough for live streaming.

The real jump was with the LLMs (Large Language Models). The tech jumped from the 60's (Eliza) to 1997 (LSTM) to 2017 (transformers) to 2023 with GPT and LLaMA (Facebook's model, was leaked to the public via torrent in March 2023), Alpaca etc.

tl;dr It's a series of huge jumps rather than a slow process of improvement which is why no one saw it coming.