Kyle Corbitt
@corbtt
Currently building @OpenPipeAI. Formerly @ycombinator, @google. I am always down to go on a quest.
🚀 Meet ART·E—our open-source RL-trained email research agent that searches your inbox and answers questions more accurately, faster, and cheaper than o3. Let's go deeper on how we built it. 🧵
Reposting for reach since this isn't common knowledge. Don't get tricked by GPU providers that offer you PCIe H100s! Pay the extra $0.30/hr for SXM H100s, they have 2x the power envelope and way faster interconnect!
you always gotta make sure to choose the SXM variant, even if you're running single-GPU jobs. Vastly different perf profiles, super misleading that they're both labeled as H100.
Anton is one of our investors. I remember playing with the prototype of Lovable in 2023 (pre-launch) and thinking it was completely useless 😆. Really incredible what he and the team have built!
Lovable just raised $200M at a $1.8B valuation led by Accel. This all started unexpectedly with me calling my friend at 6AM to go for a walk. I've never shared this story before: (thread)
Very excited to have Cresta as a user of agent reinforcement trainer (ART)! Cresta is AI-native and has a large, sophisticated ML team. ART isn't just easy to get started with, but also very powerful!
Fantastic work y’all. We’ve been using your platform extensively at cresta especially for the email RL project and it’s been fun!
This is what we do it for!
Art is really cool, i understood how rl is working hands on using your repo. Well deserved!
Thanks for all the support Kyle! We really could not have done it without @OpenPipeAI
it would be hilarious to time-travel back to 2019 and show claude code to a group of engineers and tell them that in 2025 everyone can use this. but then ALSO tell them that productivity stats basically haven't changed. they'd totally crash out.
Super cool seeing more people release successful projects trained with ART. The presentation is very informative! 😃
I presented at MLSG yesterday: How I trained a legal search agent that outperforms o3 using RL on a single H100 in 9 hours. I believe this is a shift in how we build AI Agents