Replying to:
This live show was one of our biggest, over 1600 people tuned in!
As always, you can find the full writeup + show-notes and links to the podcast episode on all the major podcatchers here:
sub.thursdai.news/p/thursdai-june-20th-claude-sonnet
Of course the number one thing we talked about was Claw'd Sonnet 3.5, it's now the LLM king, and I focused mostly on that in the writeup!
we had @skirano on to tell us about some crazy stuff Sonnet pulled in his tests
x.com/AnthropicAI/status/1803790676988920098
We also of course had to cover @deepseek_ai latest coder V2 , the absolute beast of a open source coding model, with @Yampeleg and @nisten talking about how great it is and @terryyuezhuo confirming that it's very robust (+ a little deep dive into Big Code Bench w/ Terry)
x.com/deepseek_ai/status/1802680388256768145
Big Code Bench is trying to obsolete HumanEval, which we know is flawed at this point as a metric, and Terry had a lot to say on the subject, and you can see a deepdive from @huggingface @_philschmid here, but also tune in to hear @terryyuezhuo talk about why it make sense
x.com/_philschmid/status/1803315847898796222
I was super excited to cover @AIatMeta FAIR team release of Chameleon MMIO (multi modal on inputs and outputs) series of models, alas, they nerfed the image drawing capability but they are still cool!
@ylecun let us play! 🙏
x.com/ylecun/status/1803200026094739734
In the big companies we covered that the most underrated feature from Google IO has now hit GA, context caching is now GA!
Also @luciekerley was in the audience 👀 and she was excited about V2A a new technique from DeepMind to add Audio to Videos (Come up next time Lucy!)
x.com/googledevs/status/1803096613839057022
Speaking of videos, who could escape GEN-3 videos on their timeline this week? we of course geeked out on @runwayml latest video model that looks incredible (🙏 LET ME PLAY W/ IT PLEASE @c_valenzuelab) and the vid2vid inpainting capabilities!
x.com/altryne/status/1802780351934365997
It's always fun when @skalskip92 comes on the show, and it's way funner when he's coming to talk about an open source MIT licensed amazing new tiny VLM from Microsoft that does all kinds of amazing things, def worth a listen
x.com/reach_vb/status/1803366557612933499
And finally, we geeked out on Mixture of Agents with @james_y_zou (collab w/ @togethercompute ), what it is, how they beat GPT-4 with open source models, how Mixture of Agents tops the AlpacaEval leaderboard and how textgrad does back propagation with text...
x.com/james_y_zou/status/1801656163936964919
Amazingly we also followed up with @corbtt from @OpenPipeAI to tell us about their newly announced, also Mixture of Agents work, that was similar and also topped AlpacaEval with 68% (tho not with Open Source models)
x.com/corbtt/status/1803813970018791845
There was a bunch of other discussions, thanks to @ldjconfirmed @_ScottCondron @mattshumer_ @AlpinDale and many more folks in the comments!
What an amazing show, really hope you listen to this one, at least to the TL;DR 👇
x.com/altryne/status/1804184029178003579