Glen Beer

Who should look after the environments we use to train AI agents? Hugging Face is giving OpenEnv, its project for reinforcement learning environments, to the AAIF and the Linux Foundation so it can be neutrally governed. Ben Burtenshaw (@ben_burtenshaw) shared the news and says the committee behind it has built and used OpenEnv at Hugging Face, Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Nebius, Modal, Prime Intellect, Unsloth and more. Environments are where an agent practices the skills it will use for real, so they deserve the same care as the model and the training code. Shared governance makes it easier for everyone, from a big lab to a student, to build on the same environments and compare results fairly. If you're learning RL, pick one OpenEnv environment, train a tiny agent against it, and write down what the reward actually measures. #OpenSource Link in Comments

What if your if-statements were already a neural net? Cheng Lou (ex-React core, now at Midjourney) open-sourced Handwired today, and it is a gem for anyone who teaches how networks actually compute. Handwired turns ordinary JavaScript control flow (conditionals, loops, and state) into matrix multiplies and ReLU-style ops over floats. The classic example: a > 4 ? 10 : 12 becomes above4 = relu(a - 4) - relu(a - 5), then result = above4 * 10 + (1 - above4) * 12. The outputs match the original program exactly, the weights are written by hand so there is no training step, and the whole thing stays fully differentiable. Demos run in the browser without a GPU, including a Mini OS desktop (about 9,909 weights across 130 layers) and a JSON search over 100,000 documents that compiles the JS path into WebGPU kernels. The repo is MIT licensed. For a classroom warm-up, rewrite that ternary on paper with ReLUs, check both sides by hand, then open the Mini OS demo and ask students what “130 layers with zero training” means for how we talk about networks. 🧠 #MachineLearning Link in Comments