 The video for my 40m talk at @NorthBayPython is now available, and I've put together an accompanying edited transcript with annotated slides and links Catching up on the weird world of LLMs: https://simonwillison.net/2023/Aug/3/weird-world-of-llms/ If you haven't been completely immersed in this world for the last year, my hope is this can help catch you up! 10 comments
The @NorthBayPython team did a fantastic job with the video and audio quality (and Next Day Video really do live you to their name) - the other videos from the conference can all be found on YouTube here: https://youtube.com/@NorthBayPython @simon When can we ask a LLM to "create a transcript from this: <youtube url>" and it will be able to produce something as good? @osma @NorthBayPython I'm going to write about that later... I actually got a good start on that transcript by extracting the YouTube captions and then running them through Claude (I had to do a lot of editing myself though) @simon @osma @NorthBayPython FWIW I find Whisper medium.en or large-v2 frequently outperform YouTube automatic captions for English. WhisperX and whisper.cpp have also gotten the processing much faster on consumer hardware  @simon great talk! thanks for this, i find your work immensely helpful for keeping tabs on what’s going on. i like to imagine one reason why they might be lenient with licensing violations re improving other models is it would expose them to more liability from exposing how the models were trained 🤔 @phillmv yeah absolutely - it's hard to set a precedent for fiercely protecting your model when it's trained on a bunch of unlicensed works!  @simon @NorthBayPython @scottjenson thanks for posting this! I’ve been interested in making my own LLM trained on data that’s important to me, this is super cool! |
Gregory's Server
@NorthBayPython I added a table of contents - there's a lot of stuff in there