@quocleix Before a great talk by @denny_zhou on LLM reasoning which mapped it as just intermediate token states in between input and output. Hoping to dive more into this when I get more time to train models ☺️ https://t.co/M8CpjQOrn9

Queues in @Cloudflare working very nicely :) https://t.co/oDWu9vMGrX

Such a great bit here with the nickel. https://t.co/uqzH8NT82l

Such a great bit here with the nickel. https://t.co/uqzH8NT82l

Dave Foley, the Michael Jordan of drag, so goated they made him do it on NewsRadio too https://t.co/DVYL3Mi78B

9-5 routine 6 years ago (dreams of becoming him) https://t.co/Xb4OJVFlD8

9-5 routine 6 years ago (dreams of becoming him) https://t.co/Xb4OJVFlD8

Disappointed that, while people on here have been talking about performative readers all week, I haven't seen this image once https://t.co/Bdm0q1uBhk

Remember this https://t.co/vkEguhTPKM

Remember this https://t.co/vkEguhTPKM

We said “coming weeks” and proceeded to add three over the holiday weekend.😅 Now live on @OpenRouterAI via our W&B Inference service: • @deepseek_ai V3.1 • @OpenAI's gpt-oss-120b & gpt-oss-20b https://t.co/aApffVoUSI

spend more time with your datasets really get to know them peek inside, like really inside millions, billions, trillions of tokens garbage in, garbage out still applies 1 bad dataset out of 15 drove 3 runs straight into madness https://t.co/PDZv4hV6xU

Couple of weekends ago at the @weights_biases sunday code cafe in SF: why don’t we use the model’s own uncertainty to reason within itself? so i hacked a tiny loop that makes models “reason” their own uncertainty! Today the paper got published: (arxiv link below) https://t.co/84DGh7zT3q

We’re sponsoring the official Le Chat MCP Hackathon with @mistralai in Paris! 🇫🇷 Alongside @cerebral_valley, @motierventures & @qdrant_engine. A 2-day overnight buildathon with top-tier AI engineers. Apply below to secure your spot and compete for +$6K in cash prizes. https://t.co/9qqjbOolMD

📆 Sat Sep 13, 9:00 AM — Sun Sep 14, 5:00 PM CEST (overnight hacking) Final demos are on Sunday and our wandb judges will be on-site. https://t.co/BcsRmbjqat

🧵 We acquired OpenPipe! OpenPipe's ART framework makes it easy to beat foundation models with smaller open models on real problems, using reinforcement learning. https://t.co/321mhgytXY

(1/n) Check out our new paper: "Fantastic Pretraining Optimizers and Where to Find Them"! >4000 models to find the fastest optimizer! 2× speedups over AdamW? Unlikely. Beware under-tuned baseline or limited scale! E.g. Muon: ~40% speedups <0.5B & only 10% at 1.2B (8× Chinchilla)! https://t.co/ba09j9ZFZE

Impressive work from @wen_kaiyue, @dlwh, @tengyuma & @percyliang! They did over 4.6k wandb runs to fairly benchmark 10 optimizers across 0.1B-1.2B. Matrix-based (Muon/Soap) lead, but gains shrink with scale. Check out their wandb workspace here: https://t.co/hn1WsqfGkS

Tracing your @crewAIInc agent framework is the fastest way to turn “why did it do that?” into “here’s the exact chain of calls.” Below, we'll show how to instrument a CrewAI triage workflow with @weave_wb so you can see every decision, tool call, token, and cost. https://t.co/iNKLo2CVK7

CrewAI orchestrates role-based agents (e.g., Classifier, Responder) via Crews/Flows. @weave_wb attaches tracing to those runs automatically, capturing prompts, outputs, latencies, and token usage with parent-child context. No manual logging! https://t.co/BuSDqjRZZR

Setup is so easy! • pip install crewai weave wandb • wandb login once • weave.init(project_name="crewai_demo") at the top of your script Run everything normally. W&B Weave streams a link to the interactive trace UI for each execution. https://t.co/UU7enL3mNA

Grab the tutorial + starter code below! https://t.co/WxtAYHyzd2

Coming soon to @weave_wb! Huge upgrades that will be especially useful for Reinforcement Learning. We cannot wait to share with you the details. 🤐 https://t.co/ImTVdStH3a

How many times have you bookmarked a tutorial “for later” because you were too stressed about keeping up to actually enjoy learning it? It's time to stop panic learning, and start learning with joy. https://t.co/VDMYfrycZg

Introducing 20+ connectors powered by MCP and a fully controllable Memory in Le Chat—making it one of the most connected and relevant AI assistants for enterprises and consumers. Why switch to Le Chat? A 🧵 https://t.co/7HjXEidOC2

Find security bugs before they're merged. 🐛 GitHub code scanning analyzes your code and surfaces vulnerabilities directly in your pull requests. ✨ Secure your code with every push. Learn how. 👇 https://t.co/GGMhj0YIQP

📣 #GitHubUniverse sessions are here! Check out all of the incredible speakers taking the stage this October. https://t.co/LNTutWNaE5 https://t.co/XRewlL9JPk

+3 more

Who will maintain the future of open source? Here are 6 ways you can help new contributors grow into future leaders—and keep the open source community going for generations to come.👇 https://t.co/bFesRwRfYK

Have you played with the ASCII cube at the bottom of the GitHub Shop yet? Now you can get it in t-shirt form. 👉 https://t.co/PpAe0N7OEs https://t.co/kBaZMunc2R

Ready to play Tic-Tac-Toe against GitHub Copilot? 🤖 With the Model Context Protocol (MCP), you can connect AI to custom tools. Here's how you can build your own MCP server from scratch using our fun game server example. 🚀 https://t.co/4gOz2XO4NB

🚀 LIVE NOW! Join @eliostruyf for Open Source Friday! 🎉 See VS Code Demo Time, a #VSCode extension for effortless, polished demos. #OpenSourceFriday 🔗 https://t.co/FuuU8WazY2 https://t.co/ESvolBEbHe

Private & paid https://t.co/IbjXiUv0kp features are broadly available again in Syria, following relaxed sanctions/export controls. https://t.co/ZXGHr7Bngh أصبحت ميزات https://t.co/IbjXiUv0kp الخاصة والمدفوعة متاحةً مجددًا على نطاق واسع في سوريا عقب تخفيف العقوبات وضوابط التصدير.

+1 more