Your RL run just spiked at step 89! But, do you know why? We’re fixing that. Today we’re launching W&B Weave Traces to give you a step by step look into your agent’s decisions. This is the first drop from our fresh new integration with OpenPipe. More RL magic is incoming.

Interesting Muon experiment 🤓 Learning rate of Adam parameters (embeddings/gains) does not matter so much (here from 1e-4 to 1e-2). It’s more about Muon LR

Shawn Lewis W&B Weave Weights & Biases have to say kudos on this. i added a bunch of weave.ops around the code to debug a perf issue in my GRPO rollouts. and i got flamegraphs!! and it's great for inspecting all the env interactions

Great to see people trying out using weave traces in W&B model training runs to inspect agent rollouts. Would love to chat to more people trying this out and to hear any other rollout visualizations you’d like to see on top of this

my latest agent

🚀 Big launch from OpenPipe: We just launched Serverless RL — train agents faster and cheaper with zero infra headaches. Compared to running your own GPUs, Serverless RL is: - 40% cheaper - 28% faster wall‑clock - instantly deployed to prod via Weights & Biases Inference

Sept 8: CoreWeave acquires OpenPipe Oct 8: OpenPipe ships Serverless RL

RL X-mas came early. 🎄 For too long, building powerful AI agents with Reinforcement Learning has been blocked by GPU scarcity and complex infrastructure. That ends today. Introducing Serverless RL from wandb, powered by CoreWeave! We're making RL accessible to all.

Evals versus no evals is a pretty silly debate; the answer is always just enough evals. "Enough" means you’re getting a strong enough signal to make the next iteration worthwhile. If you can get enough signal by putting it in front of users and gathering implicit / explicit

I appreciate the "Research use cases" section of OpenAI's Apps SDK. They've clearly learned that AI projects fail when teams don't: - start with clear user goals - prototype against real prompts - align scope before building tools It’s applicable to almost any AI app and a

100% agree with Andrew on the unreasonable effectiveness of error analysis, and that it’s a bit different for genAI and agents. Turns out there is a structured & well established framework to help with error analysis in gen AI—grounded theory! Hamel and I go into detail and do

Stop juggling tabs to test your prompts! 🥵 The W&B Weave Playground is your new home for iterating on and comparing LLMs. And did you know... you can now generate images right in the Playground? Just search "image" in the model dropdown!

The Princeton University University lab built their agent evaluation harness (HAL) using W&B Weave. They're leveraging Weave to automatically track, monitor, and unify telemetry across different LLM providers and frameworks for consistent, in-depth evaluation.

bro W&B Weave is fire

Filippo Facioni W&B Weave I did a project with weave and it’s really easy to use right, don’t know why there is no hype

New QoL update for wandb Weave. We've added Quick Filters for your evals. Now, you can use the filter dropdown to search by name or dataset. This makes it easier to identify comparable evaluations and analyze performance trends across datasets.