Sergi Castella i Sapé (@sergicastellasa) 's Twitter Profile
Sergi Castella i Sapé

@sergicastellasa

Search + Language Models + law at @Sdu // Creator @earkind // MSc in AI at @UvA_Amsterdam.📍De Barcelona, vivint a Amsterdam.

ID: 558902662

linkhttps://www.earkind.com calendar_today20-04-2012 19:56:09

318 Tweet

228 Followers

361 Following

Sergi Castella i Sapé (@sergicastellasa) 's Twitter Profile Photo

Today's episode is up! covering Claude's 100k context window, Transformer Agents, Federated Instruction Tuning, Pretraining Without Attention, and more.

Sergi Castella i Sapé (@sergicastellasa) 's Twitter Profile Photo

Today's episode of GPT Reviews🤓 @MetaAI's new features for advertizers, imageblind... also, loved the fake sponsor today😂 Sobstopper Tissues from Soft Inc. How's that not a real product

Sergi Castella i Sapé (@sergicastellasa) 's Twitter Profile Photo

A QA engineer walks into a bar. Crawls into a bar. Runs into a bar. Dances into a bar. Tiptoes into a bar. Runs a bar. Jumps into a bar. 😂 today Gio's ending joke made me actually laugh. The delivery was👌 apple apple.co/3BvlNrA spotify sptfy.com/NtXC

Sergi Castella i Sapé (@sergicastellasa) 's Twitter Profile Photo

Really likes Symbol Tuning for ICL from Jerry Wei and colleagues! He and his brother consistently author high quality papers on LMs. Will continue to post updates on @earkindtech (link to the latest episode and all platforms in the bio).

Sergi Castella i Sapé (@sergicastellasa) 's Twitter Profile Photo

A month ago I started a daily AI-generated podcast: "GPT Reviews", and I just made the code public🎉 ❓ Check out how episodes are generated, make your own episodes, fork, geek out yourself😬 github.com/sergicastellas…

A month ago I started a daily AI-generated podcast: "GPT Reviews", and I just made the code public🎉

❓ Check out how episodes are generated, make your own episodes, fork, geek out yourself😬

github.com/sergicastellas…
Stella Li (@stellalisy) 's Twitter Profile Photo

🤯 We cracked RLVR with... Random Rewards?! Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by: - Random rewards: +21% - Incorrect rewards: +25% - (FYI) Ground-truth rewards: + 28.8% How could this even work⁉️ Here's why: 🧵 Blogpost: tinyurl.com/spurious-rewar…

🤯 We cracked RLVR with... Random Rewards?!
Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by:
- Random rewards: +21%
- Incorrect rewards: +25%
- (FYI) Ground-truth rewards: + 28.8%
How could this even work⁉️ Here's why: 🧵
Blogpost: tinyurl.com/spurious-rewar…