AX is an open orchestrator for agentic workloads that treats agents as a new class of stateful, bursty actors rather than microservices or batch jobs. It provides four declarative primitives - Task (sandboxed execution with resource limits), Workspace (repo and dependency setup), Gateway (network allowlists and credential injection), and Model (centralized model configuration and secrets) - so teams can run, suspend, resume, and delete agent sessions safely and reproducibly. A short CLI example shows how to declare a workspace and task in YAML, apply it, watch a task reach Running, SSH into the sandbox to build code, create files, suspend and resume state, and finally delete the task. The system emphasizes cheap creation and strict isolation to avoid runaway costs from uncontrolled agents.
Under the hood AX runs on Agent Substrate, a compute runtime designed for dense, stateful actor lifecycles that claims support for billions of concurrent tasks per cluster, sub-second checkpoint/suspend/resume with zero cold start, and dense multiplexing so idle waiting time is converted into spare compute capacity. Built-in generative features let users describe workspaces in plain English and have toolchains installed automatically. Targeted at developers and researchers, AX draws on Google DeepMind research and production experience, supports interactive coding, Jupyter, long-running agent servers and RL evaluations, and is released under Apache 2.0.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
AI chatbots frequently provide incorrect answers to financial questions, according to recent reports. This raises concerns about their reliability in handling sensitive financial information. (ft.com)
The White House aimed to promote the development of artificial intelligence, but the emergence of Anthropic has influenced the landscape. Anthropic is a key player in the evolving AI industry and policy discussions. (politico.com)
Jev enhances deepclause and Pi by supporting probabilistic logic programming, enabling more deterministic and scalable expert systems. This integration aims to revive probabilistic logic approaches, improving efficiency over traditional LLM/agent methods. (deepclause.substack.com)
Backpropagation computes neural network gradients efficiently by propagating errors backward through the network. This backward pass leverages the chain rule to locally calculate derivatives at each node, which is more effective than a forward approach. (gregorygundersen.com)
Dario Amodei, CEO of Anthropic, writes about the risks and ethical concerns surrounding artificial intelligence. His essays aim to clarify public fears and promote responsible AI development. (nytimes.com)
ByteDance Seed and Tsinghua University have developed DAPO, an open-source reinforcement learning system. The project aims to advance AI research and development through accessible tools and frameworks. (github.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.