🚀 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. 🧵

finished my claude code project with 3% context length to spare ✅

Proud startup moment: We interviewed Soham in February and he failed our tech screen. We have an insanely high bar at @OpenPipeAI and the most cracked team I've ever worked with.

RT @LucasAtkins7: You can fake it pretty far in this industry just by saying, “Hrmm, that’s cool but I’m worried it won’t generalize,” when….

RT @0xTejpal: Most agent builders (ourselves included) realize within days or weeks that just prompting won't get you far. Incredibly bull….

It's becoming increasingly obvious that mass model customization is the future. Even big labs, which have traditionally pushed customers towards using prompting alone, are starting to bet big on fine-tuning. The gains are too huge to ignore.

with bf16 of course. bf stands for better flops.

in the spirit of "you can just do things", you can, today, rent a 1-click cluster of 512 H100s on Lambda. that's more FLOPs than the largest supercomputer in the world as of 2022.

We're in the "IBM PC clone" era of LLM chat. One will win big, a few will survive, most will disappear.

ART (agent reinforcement trainer) now supports multi-device training for faster training jobs!.

RT @casper_hansen_: This is great evidence that low-hanging fruit is everywhere in RL.

I wouldn't have said this 6 months ago, but I now believe all serious agents will be RL'd on their specific task. The gains are too easy and too huge to ignore. Either @OpenAI et. al. will provide APIs to do this on-platform, or open source will win.

RT @peytoncasper: ill pay someone $2k if they can use puffer lib or openpipes art framework to model a rl cua environment that converges us….

We've open sourced all code, data, and lessons learned. We also have a live demo available of the summarization agent. You can find all that in the full write-up here:

By directly optimizing on the number of questions that could be successfully answered from the summary, we taught Summary-RL what kinds of information to include. Within 30 training steps it already reached SOTA! ($22 to train)

Why did we do this? LLMs are already good at generating summaries, but they don't always focus on the information you care about. RL lets you customize a model to focus specifically on the types of data you want to preserve.

Hot RL summer continues: we just released Summary-RL, an RL-trained summarization model that reaches SOTA on ServiceNow's Repliqa summarization benchmark!

just so everybody is clear this is @sama's way of announcing that the openai open-source model will be around o3-mini level.

Exciting evidence that RL can be incredibly sample efficient: when using GRPO to train a modified version of ART-E (agentic RAG task), we find that we're able to get qwen2.5-14b to exceed gemini 2.5 flash performance with 1 training scenario, and exceed o3 with just 16!. This

Hope y'all are ready for hot RL summer.

if you're looking for a startup idea I'd love a sick genai-native craigslist replacement. - snap a picture and ai removes background, looks up item online, prepopulates price+listing. one button to post. - autoresponder. never answer the same question twice. - AI buyer's agent.