Links:

• my little whisper.cpp bug fix
• why the logits were calculated inconsistently
• wav2vec2 on arxiv and huggingface
• the openai whisper asr model announcement
• beam search patience
• top notch guide to CTC
• the bitter lesson (and in meme form)

Errata:

* I referred in the show to LLMs as encoder-decoder models. Most modern LLMs are decoder-only.
* I messed up readability at Google. It means approvability, apparently. 🤷