@abhijithneil@sama The kind of work that openAI does now requires hugh amount of infra which is available to less than 0.1% of companies. They need to have a revenue model that can consistently fund this
@dwarkesh_sp@altryne@1littlecoder@Gradio@huggingface alternatively, just predict the number of speakers and then use the above tool. run a forward pass through the entire clip, identify optimal num clusters based on embeddings of all clips, and use this to initialize the num speaker parameter
@karpathy Turns out that Whisper can produce reasonable zero-shot audio features for speaker prediction on Lex podcasts - got an error rate of less than 7% using SVM. Also, using hidden states of some encoder blocks can be more useful than others
https://t.co/mQOZakn0vy
You can replace search engines with a LLM, you'll get a much better experience! (As long as you don't mind that most information you get is made up, and as long as you don't need to track sources, and...)
Using a combination of scientific detective work and #AlphaFold, Benjamin Perry and his team @DNDi have cleared a path towards finding a real treatment for leishmaniasis, a neglected disease impacting millions of people globally.
Learn more at https://t.co/fsgnS36DYJ #Unfolded
I stayed up all night talking to people in Ukraine and Russia.
I'll publish Mark Zuckerberg podcast another today.
I will travel to Russia and Ukraine. I will speak to citizens and leaders, including Putin.
War is pain. My words are useless. I send my love, it's all I have.
@elonmusk@balajis@TheBabylonBee One of the odder parts of this is watching people who a few months ago were talking about how tech companies had way too much power now saying that they must be the arbiters of this extremely important topic