Examples

As described in juejin.cn/post/723154…, the output of generative learning is structured stuff.

• Sentence: composed of tokens (character in Chinese, word piece in English), e.g., ChatGPT
• Image: composed of pixels, e.g., midjourney
• Video: Imagen Video
• Audio: InstructTTS
• Sound: AudioLDM

AI that can use tools

• New Bing
• WebGPT
• Toolformer

Strategies

• Autoregressive (AR) model：“各个击破”
• Non-autoregressive (NAR) model：“一次到位”

Comparison

Combination

-- this is actually the basic concept of diffusion model