Weights & Biases

Keeping up with AI tooling shouldn't require a scavenger hunt. That's the whole idea behind What's New Wednesdays. One live session on the last Wednesday of every month, walking through what's new in the W&B platform and the research shaping it. We'll walk through demos, not slides, plus open Q&A. For this one we'll explore how World Action Models let robots adapt zero-shot to unseen tasks, example-level eval tables in W&B Models, and CoreWeave ARIA our AI Research and Iteration Agent for message queueing. With Angela Samples, Anushrav V., and Julia Rose. Subscribe once, stay current monthly (recordings sent to all registrants). Join us here: https://www.utm.io/ur8f8

1 billion runs logged on Weights & Biases. A run is a training attempt. A billion times over, that's a massive dataset on how models actually get built. That history is the reason CoreWeave ARIA works. An agent that can read experiments, propose the next hypothesis, and launch the next run needs a foundation of real iteration data to reason over. W&B has that. Now it sits on the same platform as the compute those runs actually consumed. That's the point of bringing W&B and CoreWeave together. Training and inference, tracking and evaluation, hardware and software: one loop instead of a handoff between separate systems. The gap between "the model shipped" and "the model got better" gets smaller every time that loop tightens. Congrats to the whole W&B team, and to every researcher who logged a run, including the bad ones. Those count too. More on where we're taking this from Shawn Lewis: https://utm.io/usc7V

🥳 Somebody out there just logged run number 1,000,000,000! To them it was probably nothing special. One more config, one more wait to see if the loss curve bends. That's the thing about runs, each one is small. A billion of them is the story of modern AI. So this one's for everybody in that number. Everyone who tried one more experiment after swearing that was the last one. Everyone whose breakthrough is in there, and everyone whose 400 failed runs made it possible. We spent the year building for you. A rebuilt W&B Weave for production agents. An iOS app, W&B LEET for the terminal, an MCP server for your coding agent. CoreWeave ARIA, our AI Research and Iteration Agent proposing and launching your next run. Serverless Inference, RL, and SFT underneath. Thank you doesn't cover it, but THANK YOU. On to the next billion. Full story: https://www.utm.io/usc7Z

Started in 2023 as a one-day meetup for ML engineers. Now it's on transit signage 🤷, which is a sentence we didn't expect to type. Three days at Moscone South this September: 30+ sessions, hands-on labs, 2,000+ engineers, hosted with CoreWeave. The sign is new. The community isn't. Sept 29 – Oct 1 · Moscone South, San Francisco Register: https://www.utm.io/urRxD #FullyConnected26

We ran the same phishing email through two AI agents. The email had a prompt injection buried in the body and a customer's SSN and credit card number sitting in the text. -> Agent one, no guardrails, followed the planted instruction and echoed the SSN and card number straight back into its reply. -> Agent two, running CrowdStrike Falcon AI Detection and Response (AIDR), blocked the injection before the model ever read it, redacted 4 sensitive items (US_SSN and CREDIT_CARD), and still returned a usable draft. Same input. The gap between those two outcomes is the whole point. Underneath both, W&B Weave traced the runs end to end, so every scan, detection, and tool call was recorded and queryable. When something fires in production, a blocked request shows up in your security team's Slack, linked to the exact trace behind it. That is what it looks like to run internal AI on sensitive data without flying blind. Full walkthrough, with the code and both side-by-side runs, in the comments.

⚡️ Nemotron 3.5 Lightning is now LIVE on CoreWeave Serverless Inference. The fastest open model in its class for always-on agents, NVIDIA Nemotron 3.5 Lightning is a customizable 30B MoE (3B active) model distilled from the frontier Nemotron 3 Ultra. Works great with all of the popular agent harnesses, and optimized for the work that fills most of an agent’s day. Coding, tool calling, instruction following, long multi-turn agentic workflows. Always-on agents make thousands of model calls, and Lightning is built to clear them fast. On CoreWeave Serverless Inference, you get instant access through your wandb account. Observability is built in with W&B Weave for traces and evals, and there are no endpoints or infra to manage. Get started here: https://utm.io/ur9aW

Your agent aced every offline eval. Then production happened. We keep watching the same story: a team polishes an agent against a test set until the numbers look great, and then production introduces failure modes no offline eval was ever going to catch. Real traffic is weirder than anything you'd think to write down. So we built W&B Weave around the traffic itself. It watches every production session and classifies the behavior it finds, then turns the catches into evals so Tuesday's failure can't sneak back in on Friday. Your users are already running the eval. W&B Weave just writes it down. Try it here: https://www.utm.io/ur7Bf

"We're just at the beginning of this S curve." says Lin Qiao, CEO of Fireworks AI, on this week's Gradient Dissent. Fireworks now processes more than 40 trillion tokens a day, more than OpenAI's API and Gemini's, and Lin says usage has expanded fast, from coding last year into legal, finance, recruiting, sales, and a growing wave of co-work tools. Link to the full episode in the comments.

Nobody ever fixed a slow training run by watching a slide about it. The sessions you remember from any conference are the ones where you typed. So the Fully Connected labs put you on live clusters. You'll be profiling an actual run on SUNK, tracing where the time goes (NCCL, I/O, checkpointing), and watching your MFU move when you fix it. There's also an agent that runs its own experiments, an inference ladder to climb, and a protein model to fine-tune. Guided by the CoreWeave and W&B engineers who do this for a living. Sept 29 – Oct 1 · Moscone South, San Francisco Register: https://www.utm.io/urRxD #FullyConnected26