与「controller」相关的搜索结果

Google 开源了 "Agent 工作负载的 Kubernetes"「AX」 AX 是为 Agent 设计的声明式编排运行时，你用 YAML 声明一个 Agent 任务，AX 负责在集群中沙箱化、配置环境、管控网络并大规模运行它。 开源地址： 它解决什么问题 项目的立论很清晰：Agent 是一种既有的编排体系都不匹配的新型工作负载。 · 它不像微服务(无状态、常驻)，Agent 会持续积累状态(对话记忆、工作区文件、工具会话); · 它不像批处理作业(跑完即弃)，Agent 大部分时间在等待，等模型响应、等工具返回、等人类审批，期间沙箱空转烧钱; · 它运行的是不可信代码，需要严格隔离；它还调用外部模型 API 和 MCP 工具服务器，需要网络与凭据管控。 架构：四个二进制 + Redis 四个二进制分工：ax(开发者 CLI)、ax-server(无状态 gRPC API)、ax-controller(调和循环)、ax-task-runner(每个任务容器内的 PID 1)。 核心原语：Task / Workspace / Model (+ Gateway) Task 是最小执行单元：带 CPU/内存限制的隔离沙箱。AX 刻意把它做得细粒度、可自由组合：一个任务可以是全部工作,也可以是任务树的根节点。生命周期用 status.phase + Conditions 表达,支持挂起(检查点保存状态)与恢复。 Workspace 是最有产品想象力的一层。它把“环境准备”声明化:列出需要的 Git 仓库、MCP 服务器、技能包,runner 在命令启动前物化好。更激进的是 generative workspace:你可以只写一句自然语言目标("搭一个 Python 3 开发环境"),首次启动时 runner 会派一个引导 Agent(Antigravity,需 GEMINI_API_KEY,默认限时 10 分钟)去实际安装工具链并验证依赖。声明一次,任意任务复用。 Model 把“用哪个模型、什么参数、密钥在哪”抽成命名资源,密钥引用 K8s Secret。轮换密钥、锁版本、调温度只需一次 ax apply。 沙箱与运行时细节 每个任务容器以 ax-task-runner 为 PID 1：启动元数据服务(端口 80,HTTP/1.1+h2c,暴露 /healthz、/readyz 和任务/工作区自省端点——Agent 不需要 SDK 就能读到自己的配置)、按绑定顺序初始化工作区、然后 fork 出 spec.command 并持续监管。命令退出后 runner 仍驻留,所以 ax ssh 和元数据服务在命令结束后依然可用。停机采用 SIGTERM + 10 秒宽限 + 强杀的两级策略。 安全模型有几处值得注意的门控：guest 服务(任意进程执行与文件读写,ax ssh 的底层)默认关闭,只在 debug: true 时开启——ax ssh 连不上未开启的任务是刻意的安全设计,不是故障;Gateway 用显式 allowlist 管控出站流量,并可向入站请求注入凭据,避免把 API key 直接塞进 Agent 环境。 路线图透露的方向 五大方向：Actor 架构深化、空闲检测自动挂起、有状态任务 fork、任务级 SPIFFE 身份做零信任 mTLS、runner 层自动采集 OpenTelemetry 遥测与结构化 Agent 轨迹。

🚨SlowMist TI Alert🚨 💸 @EnsoBuild Loss: ~5.6 ETH 🔍 Root Cause: An oracle price calculation error occurred in the Enso Finance / DPI Strategy Vault. In `Controller.deposit()`, the EnsoOracle's `estimateStrategy()` is called before and after the user tokens are transferred, and shares are minted on the difference: `mint = amountAdded * totalSupply / valueBefore`. The valuation chain — EnsoOracle → ItemEstimator → `ProtocolOracle.consult()` — prices tokens via UniV3 `pool.observe()`, but the registry's `fee` field is reused as `secondsAgo`, producing a near-spot TWAP window. Due to the absence of TWAP consistency checks or price range validation—combined with the fact that the liquidity in the Uniswap v3 pool used for price calculation was inherently imbalanced—an attacker was able to swap 0.683 WETH for 268.42 FARM tokens on Uniswap v2 in a single transaction, while the imbalanced v3 pool incorrectly valued that amount at 6.3 WETH (~ 9.2x). Subsequently, an excessive number of shares were minted at an incorrect price and redeemed for profit. 📌 Attacker: `0x3196398321D77a2511d369DCB6eCa9d2aD87b73A` 📌 Victim (Strategy): `0x890ed1ee6d435a35d11051d9ed97ff457ce53b5942` 📌 Vulnerable contract: Controller impl: `0xd8D22509C1fe47516D8F82A28CFd728111F57Ef1` Oracle: `0xAb7505eB360cE0D63e8E88f7853677EcD5537DC0` Powered by Tx:

⚓️ Ethra Ship Notes｜Vol.050 Today, I came across a number that made me stop. Nearly 1,900 vessels are now considered part of the maritime "dark fleet" according to Windward, roughly tripling since the Russia-Ukraine invasion. The interesting part isn't just the number. It's what the number tells us about visibility. A ship can disappear from AIS. That doesn't mean the ship disappears from the ocean. It just disappears from one information layer. And that's a very different thing. This is where I think Sea Verity's approach gets interesting. Instead of asking a single system to tell us the truth, it combines different sources. Reporters on the ground. AI analysis. Controller nodes. Each piece adds another layer of evidence. And rather than forcing every observation into a simple "true" or "false", the system is designed around confidence scores. I like that approach. Because the physical world rarely gives us perfect information. A satellite image can be obscured. A reporter can only see part of a vessel. AIS can disagree with visual evidence. Weather can make everything harder. The honest answer isn't always certainty. Sometimes it's: We're 90% confident this is what happened, and here's why. That's much more useful than pretending the other 10% doesn't exist. For insurers, traders, logistics companies and regulators, knowing the confidence behind a piece of maritime intelligence could be just as important as the information itself. The ocean is enormous. Maybe the answer isn't one perfect signal. Maybe it's many imperfect signals learning how to verify each other. @EthraShip #EthraShip# #EthraShipProtocol#

以防有人不知道，美国几乎每个州都有 #无人认领财产# 截至目前，纽约州已累计归还超过 190 亿美元的无人认领财产 还有约 200 亿美元等待认领 平均每天返还超过 200 万美元 💰 有空可以搜一下自己的名字，说不定还能找回一笔钱呢～ 比如政府退款、退税、工资支票、银行账户余额、保险理赔、押金等等，都有可能 加州也很多，是 California State Controller’s Office 管理的官方网站，已经帮民众领回超过 83.8 亿美元 查询完全免费‼️ 没有申请期限 只要输入姓名就能查看 记得一定要使用官方网站，不要找第三方代办，更不要支付任何查询费用 如果曾经在美国生活、工作或留学过，不妨花一分钟查一下，说不定会有意外惊喜 其他州也都有类似的官方查询网站，可以一起搜搜看

The Backbone Pro is a premium controller that offers a slick, comfortable and clever experience for mobile gaming on iPhone or Android. Check it out now #gaming# #backbonepro# #controller# #mobilegaming# #gamingaccessories#

What if building a robot was as simple as describing it? I typed a single sentence: "Build a quadruped robot that walks a figure 8." Everything else was engineered automatically. @pyroscli doesn't just generate robot code. It executes the entire robotics engineering pipeline. It designs the robot's morphology / generates the geometry and inertias / solves and validates the kinematics / sizes the actuators / plans a dynamically feasible gait / tunes the controller and runs the simulation. Every dimension / torque / mass and trajectory comes from deterministic engineering tools with traceable sources not language-model guesses. In this demo, you'll watch the robot assemble in real time as the engineering pipeline executes, then immediately begin walking a smooth figure-8 trajectory. The goal isn't to produce another URDF that looks correct. The goal is to produce a robot that is mechanically consistent, mathematically verified, and ready to build. Natural language in. Verified robot out.

助力国产百元手柄平替Codex Micro，开源Agent controller。只要人民币，不要230美元。甚至无需运费 大致思路是，将手柄操作转成Codex操作。目前仅支持Windows 64位，由于是一天内（让Codex）赶工出来的，会有不少问题，开软件，重启Codex(ChatGPT)会好很多 开源地址：

remember the OG XBOX 360 i bought at a garage sale for $40??? i finally booted it up and IT HAS THE OG DASHBOARD somehow hasnt been updated?!?! also brought out my 2012 XBOX 360 SCUF controller, good times ~

Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple milliseconds, as long as continuous reads were delivered at the necessary bandwidth. NAND flash is over 100 times cheaper per GB than HBM, so there should be opportunity there, even after giving a flash controller a 1024 bit interface with HBM bandwidth. You could make a specialized pin protocol that just supported pipelined transfer of full 16KB+ pages from the flash to program-managed accelerator scratchpad memory and improve per-pin performance over HBM, but it might be more convenient to make it still look like a true random access memory with very fragile performance characteristics, where anything but sequential reads falls off a 1000x+ performance cliff. That has the advantage of automatically using existing cache hierarchies, and providing a natural path to update the flash memory with new model weights. With the stream-to-scratch interface, code has to be completely rewritten before it works at all, while the ram-emulation interface will start off just extremely slow, and you can incrementally sort out the changes for full performance. There may be cases where there isn’t enough scratchpad SRAM to hold the weights for a layer, which might force you to deploy the old optical drive optimization technique of duplicating data in multiple places on a sequential read to avoid seeking, but there would be capacity to burn. It might be possible to do something like cuda graph capture to record a memory access trace and have everything magically remapped to a linear sequence, but deploying programmer / agent elbow grease to manage transfers and access in a scratch ram ring buffer would be lower risk. A split memory system consisting of some channels of flash and some channels of HBM will probably be suboptimal compared to a uniform memory, but it could be much cheaper, and allow much larger models to be run. I think th case is strong for inference, but you have to stretch more for training. You can still linearize all the weight memory accesses, both reads and writes, but flash memory would quickly wear out from the writes, even if they were all perfectly page aligned. Replacing low-latency HBM with massively parallel cheap(er) DRAM at high latency might still be a worthwhile cost savings.