tech tricks
This is called “optical compression”. It reminds me of a funny idea from June of this year to save money on OpenAI transcriptions: before uploading the audio, run it through ffmpeg to speed it up by 2x. The model is smart enough to still pull out the text, and with one simple trick you’ve cut your inference costs and time by half. Optical... See more
Should LLMs just treat text content as an image?
une technique surprenante pour optimiser le coût de Whisper
Ideas related to this collection