Weights & Biases

The largest pre-seed in enterprise software history happened without a pitch deck, a product, or a single customer. Aaron Katz raised $50M for ClickHouse on conviction alone. We had him on the latest episode of Gradient Dissent, and the man does not hold back. He goes after Snowflake, Datadog, and Databricks by name. He talks about wiring $100M out of SVB 30 minutes before it collapsed. And he gets into why when Anthropic asked Claude which database to use for a specific use case, the answer was ClickHouse every time. His vision is clear: ClickHouse is the fastest database in AI and he is designing the entire stack for agents, not humans.

What's New Wednesdays returns April 29 with a Google Cloud Next 2026 edition. Every last Wednesday of the month, the W&B team goes live with demos, product deep dives, and real talk about building models and agents. This session will cover the latest Weights & Biases announcements from Google Cloud Next '26 and new features across the full AI workflow. Here's what we're getting into: → The latest innovations for pre training models → New post training techniques → Strategies for running agents in production → End to end workflow optimization Featuring Chander Matrubhutam, Russell Ratshin, and Angela Samples from the W&B AI Evangelism team. We always close with open Q&A so you can workshop real challenges directly with the panel. 📆 Wednesday, April 29 at 10am PT / 1pm ET. Virtual. Register now 👇

When regenold GmbH tested leading commercial AI assistants on 83 expert crafted regulatory questions about the EU AI Act, the results were rough. Up to 1 in 3 responses contained at least one critically incorrect element. References were hallucinated. Wrong article numbers. Missing cross references. In regulated life sciences, where a misread citation can derail an audit or impact patient safety, that rules out unsupervised use. So they built something better. And the approach is worth paying attention to even if you're not in pharma or medtech. Their core insight: building reliable AI agents and building reliable evals require fundamentally different expertise. A regulatory expert knows what a correct, complete, well referenced answer looks like but probably doesn't write code. An agent developer can instrument tracing and build tool integrations but can't judge whether the output is actually right. W&B Weave bridges that gap. The regulatory expert works through the UI to set up Q&A datasets, annotate traces, and score agent outputs. The developer works in Python to create agents, add tracing with just a couple lines of code, and build LLM as a judge scorers. Both contribute to the same evaluation pipeline. The pattern here generalizes well beyond regulatory. Any domain where correctness actually matters (legal, financial, medical, compliance) needs this same two perspective eval workflow. Domain experts defining what "good" looks like. Developers building the infrastructure to measure it. A shared platform connecting both. Read the full walkthrough from Regenold 👇 https://wandb.me/05bGVdB

Are we witnessing the next "Cloud" or "Mobile" transition in real-time? NVIDIA CEO Jensen Huang recently noted that an OpenClaw strategy is no longer optional for the enterprise. It is incredible to see an open-source project, which began late last year, accelerate so quickly that the world's leading compute provider is building enterprise-grade security reference implementations (NeMo Claw) around it. The takeaway for the ML ecosystem is clear: 1️⃣ Open-source is dictating the pace of enterprise AI. The community is solving complex orchestration and agentic workflow problems faster than proprietary ecosystems. 2️⃣ Inference is the new battleground. As companies deploy more autonomous agents (which require continuous looping and reasoning), inference demands are skyrocketing. 3️⃣ Production-readiness is the final hurdle. Nvidia hiring top-tier security engineers to harden OpenClaw signals that agentic AI is ready to leave the sandbox. Building, tracking, and evaluating these complex agent systems is the next frontier of MLOps. Is your engineering team exploring #OpenClaw or similar agentic frameworks, or are you waiting for the ecosystem to settle? Drop your thoughts in the comments. 🛠️ #MachineLearning #AI #OpenSource #MLOps #TechLeadership #Nvidia #DeveloperTools #WeightsAndBiases #wandb

📅 ThursdAI - NemoClaw, GPT 4.5 Mini, Codex subagents, Mistral Small4, MiniMax 4.7 & more AI news

The hardest part of robotics AI isn't training the model. It's evaluating it. Your outputs aren't neat classification scores. They're videos, trajectories, segmentation maps, sim results across dozens of experiment runs. And until now, comparing those meant downloading files, stitching clips together, and hoping you didn't miss a subtle regression. We just shipped a set of advanced visualization tools in W&B Models designed specifically for robotics and physical AI workflows. Synchronized video playback, pinned baselines, semantic coloring, full image inspection with zoom and pan, and side by side comparison for up to four media outputs at once. These aren't isolated features. Together they give robotics teams a single workspace where they can track, compare, and evaluate multimodal experiments without context switching between tools. If you're building in physical AI, this is worth a look. Full walkthrough and demo in the blog below! wandb.me/wandbRobotics