内容归档

i should have ironed my shirt.

@reflection_ai Read more about Beam 1.0 on the @reflection_ai blog: reflection.ai/blog/beam-is-a…

Penguin has moved up the publication date of my new book with Patchen Barss called "Smarter Than Us". The book explains how AI works for a non-technical reader and then talks about the myriad risks and what to do about them.

源： x.com/0x_meden/statu… Martin @0x_meden list of 15 best robotics researchers on X: 1. @DrJimFan , Jim Fan 2. @chelseabfinn , Chelsea Finn 3. @svlevine , Sergey Levine 4. @pabbeel , Pieter Abbeel 5. @drfeifei , Fei-Fei Li 6. @rodneyabrooks , Rodney Brooks 7. @0x_meden , Martin 8. @yukez , Yuke Zhu 9. @LerrelPinto , Lerrel Pinto 10. @SongShuran , Shuran Song 11. @tonyzzhao , Tony Zhao 12. @YunzhuLiYZ , Yunzhu Li 13. @xiaolonw , Xiaolong Wang 14. @GuanyaShi , Guanya Shi 15. @RemiCadene , Remi Cadene 🔗 View Quoted Tweet

X 上最值得关注的 15 位机器人研究者名单： 1. Jim Fan 2. Chelsea Finn 3. Sergey Levine 4. Pieter Abbeel 5. Fei-Fei Li 6. Rodney Brooks 7. Martin 8. Yuke Zhu 9. Lerrel Pinto 10. Shuran Song 11. Tony Zhao 12. Yunzhu Li 13. Xiaolong Wang 14. Guanya Shi 15. Remi Cadene

LIVE STREAM: New Open Weight Model x.com/i/broadcasts/1…

Get started: lovable.dev/whatsapp

Your Lovable app can now talk to customers on @WhatsApp . Connect your business number, then ask Lovable to build an inbox, reminders, and message templates. 🔗 View on Twitter

Banger paper from Google DeepMind on long-horizon multi-agent coding. (bookmark it) They report that, to the best of their knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents. They report that four Gemini 4 Argon agents beat Minecraft end-to-end. They defeated the Ender Dragon in 27 hours and 14 minutes with no human intervention. Each agent controlled one player through the standard Mineflayer library. No one assigned roles, wrote game logic, or scouted a seed in advance. The agents planned, wrote, and debugged every program themselves, and coordinated through in-game chat and a shared memory. Over the run, they wrote 2,783 programs and died 605 times. They split the work, shared what they learned, and recovered each other's items after deaths. If you build agents for long software or enterprise tasks, this is a useful case study. The team had to keep one plan consistent across thousands of dependent steps for more than a day. arxiv.org/abs/2610.22319 Chat with Paper: academy.dair.ai/papers/a-team-…

Welcome Beam! Congrats to @ReflectionAI on launching their first open-weight model. The 501B MoE model was trained on NVIDIA GB300 NVL72 and delivers over 70K tokens/sec/GPU on Blackwell Ultra out of the box on day 0. 🤗 Beam 1.0 NVFP4 checkpoint on @huggingface : huggingface.co/reflection/Bea… L Reflection @reflection_ai Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners. We're releasing: - Weights, tokenizer, and reference implementation - BF16, FP8 and NVFP4 weights and a DFlash2 drafter - The open source harness we built and trained Beam on: Mirror CLI Weights and model card on Hugging Face: huggingface.co/reflection 🔗 View Quoted Tweet

现在上新站最耗时间的居然是人工申请这些服务的环境变量，写代码部署比如申请这些东西还快，好希望他们都有 cli 或者 mcp 啊，现在是所有事情都想让 AI 做了。

It is pay-as-you-go, so you can test it on a small list. Try it here app.aisdr.com/signup

AI agents are useful in outbound when they handle the entire email workflow. Ami from AiSDR researches each prospect, writes a personalized email, sends it, follows up, and handles replies. It also manages the email infrastructure and warm-up. AiSDR now charges $1.50 per researched prospect with unlimited touchpoints. Yuriy Zaremba @yz_aisdr Today AiSDR is disrupting pricing for outbound sales. $1.5 per person prospected, unlimited touchpoints, EVERYTHING included. No minimum. We will take care of email infra, warmup, leads (2B+ contacts, including off-LinkedIn), enrichment (waterfall of 10+ providers), AI-researched, unique multi-channel sequence to each prospect, AI reply-classification and reply handling, everything syncs back to your CRM and tracks influenced and closed won pipeline. Ami AI, trained on 20k+ qualified meetings booked across 27 industries to help you ideate, monitor, improve and run your campaigns. All the latest and greatest in GTM tech on one platform, starting at $1.5. There is literally no alternative to this offer. - Have an offer and an audience to test for 100 leads - AiSDR is your place to go - Want to test outbound, but buying 5 tools and managing them feels like a pain - AiSDR - Planning to go to a conference and need to book few meetings - AiSDR - Just came back from a conference and have a 200 leads to follow-up with - AiSDR Any kind of outreach via email and LinkedIn - we've got you. Fully autonomous with the best GTM AI agent you can buy in the market. Outbound on demand, priced like ads. Try it today at aisdr.com . 🔗 View on Twitter 🔗 View Quoted Tweet

Head of Design at Stripe on how to stop creating zombie UIs 🔗 View on Twitter Katie Dill @lil_dill If we only use AI to make the things that we already make, just faster, then we're definitely missing out on the most interesting part. From my talk at @lennysan ’s Summit, here are four ideas for raising the ceiling, not just the floor: - Have a point of view [5:38] - Encode your standards into the machine [9:07] - Don't confuse done with good [12:21] - Unleash creativity and artistry [16:19] Full talk here: youtu.be/GLvFTMtw4Jk?si… 🔗 View Quoted Tweet

We’ve partnered with @ndstudio to launch American Forge. It’s a six-month residency for exceptional engineers to embed inside the government and build software that affects hundreds of millions of people. Right now, a small team can do more than ever - including in public service! Applications are now open. National Design Studio @ndstudio Introducing American Forge: a six-month residency for exceptional engineers ready to take on the toughest problems in American government. Run by National Design Studio in partnership with @cognition forge.ndstudio.gov/?s=x 🔗 View Quoted Tweet

What does it take to move healthcare and life sciences agents from experimentation into trusted production systems? Explore the architectures, evaluation practices, and production patterns teams are using as agents take on more complex work: ➡️ info.langchain.com/guide/scaling-…

We’re excited to be day 0 partners for @reflection_ai ’s first open-weight model! For a limited time, Beam is FREE on Ollama! In partnership with @reflection_ai this model is hosted in the US with full ZDR. It has a 1M token context window and is designed for software engineering, reasoning, and long-running agentic tasks. Run Beam on Ollama’s Cloud: ollama run beam Claude Code ollama launch claude --model beam:cloud Codex ollama launch codex --model beam:cloud OpenCode ollama launch opencode --model beam:cloud Free to use for a limited time only. Reflection @reflection_ai Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners. We're releasing: - Weights, tokenizer, and reference implementation - BF16, FP8 and NVFP4 weights and a DFlash2 drafter - The open source harness we built and trained Beam on: Mirror CLI Weights and model card on Hugging Face: huggingface.co/reflection 🔗 View Quoted Tweet

Model page on Ollama ollama.com/library/beam Blog post ollama.com/blog/reflectio…

Perplexity 开源了团队内部 agent 代码执行平台「SPACE」，把不可信的 Agent 生成代码或用户代码放进 Firecracker microVM 里隔离运行，通过纯 HTTP/SSE API 对外服务 开源地址： github.com/perplexityai/S… 核心抽象：五个概念 + 严格分离的生命周期 · Template：起始文件系统与运行时配置（不可变，类似镜像），按内容寻址（sha256: digest） · Sandbox：一个 Template 的实例，独立 microVM + 可写文件系统 · Process：在 Sandbox 内执行的程序，同 Sandbox 内进程共享环境 · Command session：一次命令的连接信息与授权，创建即分配进程 ID，attach 才真正启动 · Operation：异步生命周期请求（create/pause/suspend/wake/delete）的追踪记录 关键设计：Sandbox 的生命周期长于创建它的请求，也长于承载其输出的连接。管理面是异步的，API 只把期望状态写入 PostgreSQL，由控制器和 worker 协调收敛，客户端拿 operation ID 轮询；数据面则完全独立，通过会话级 attachment token 经 SSE 直连沙箱内进程取输出。管理 token 与 attachment token 不可互换，这是它的安全边界之一。 架构：控制面/数据面分离 + 声明式调和 请求与执行分两条路径： · 管理面：Application → API → PostgreSQL（期望状态）→ 集群控制器 → 沙箱 Pod（K8s 模式下）→ 运行时 shim → Worker 服务 → Firecracker microVM → 客户进程。 · 数据面：Application ↔ Data plane ↔ 客户进程，旁路整个控制面，输出流不经过数据库。 Worker 负责 VM、网络、卷、模板、凭证，用 btrfs 写时复制构建根文件系统（这是快速克隆沙箱的关键）；模板与快照走对象存储（本地用 Versity S3 Gateway）。部署形态可插拔：Kubernetes 适配器（本地 Kind、云端 EKS 参考部署）或直接适配器（不经 K8s 直调 worker，支持单机运行）。 Pause 与 Suspend 的区分体现其状态机的严谨性：Pause 保留常驻内存（快恢复但非持久）；Suspend 生成 VM 内存 + 运行时状态 + 根文件系统的完整快照并释放资源（真正的持久化），两者都用 Wake 恢复。默认 6 小时无活动自动 Suspend（这是空闲限制而非最长存活时间）。 性能工程：两层抗冷启动 冷启动是 microVM 沙箱平台的体验瓶颈，SPACE 给了两个正交方案： 1. Warm pools（K8s 部署）：按“模板版本 × VM 规格”预启动一批就绪沙箱，create 直接领用，池空了才退化为冷启动； 2. Base snapshots（单机部署）：从基线快照恢复，跳过完整引导。 配合 Firecracker 本身亚秒级的 VM 启动和 btrfs CoW 文件系统克隆，整体面向“Agent 秒级拿到可用环境”的目标。 “可插拔运行时”：为未来预留的执行环境身份体系 这是仓库描述里 "pluggable-runtime" 的实质，也是设计上最有前瞻性的部分。每个执行环境被定义为主机平台 × 运行时绑定 × 客户平台 × 引导配置 × 隔离语义 × 兼容性Profile 的完整组合，整体 SHA-256 内容寻址。当前只实现了编译进代码的 E0 定义：Linux/x86_64 宿主、KVM、Firecracker 1.15.1（含 jailer）、内核 6.1.155、快照格式 9.0.0。 值得注意的三个原则： · 结构可解码 ≠ 可执行：解码成功、工件存在、节点健康，都不构成执行授权，三者分离校验； · 身份不可扩展含糊：任何改变引导/恢复语义的字段都必须产生新的 schema/引用，“加了字段还沿用旧引用”被明令禁止； · 主机平台类型里已预留 Linux/Windows/macOS 的区分身份，虽然当前只实现了 Linux 适配器。 这明显是在为未来支持其他 hypervisor、客户架构甚至非 Firecracker 运行时铺路，而不是把 Firecracker 硬编码进协议。 安全模型（分层清晰） · 硬件级隔离：Firecracker microVM 边界（KVM），这是对不可信代码的根本防线； · token 分域：管理面 bearer token 与数据面 attachment token 隔离； · 网络策略：沙箱 Pod 独立 namespace（sandboxes），NetworkPolicy 只放行 API 与控制器端口，出站默认限公网 IPv4 + 集群 DNS，可配 egress profile 精细管控，甚至有 TLS 拦截豁免机制； · 供应链：gitleaks、cargo-deny、supply-chain 目录、pre-commit/lefthook 一应俱全； · 内部 RPC 走 TLS，模板注册可限制可信 S3 bucket 白名单。 兼容性与工程成熟度 · E2B 兼容层（ECW）：在边缘翻译 E2B API 请求/响应，用 E2B Python SDK 的应用可近乎无缝迁移（语义映射：E2B 的 pause → SPACE suspend）； · 客户端零绑定：纯 HTTP + SSE，OpenAPI 规范直接服务在 /openapi-m1.json，不需要嵌入 Rust 库——这对采用率至关重要； · 运维面齐全：space-admin-cli（节点、暖池、模板构建）、Web UI、OpenTelemetry 可观测性、mdBook 文档； · 开发体验：tilt up 热重建，单测 + Kind 冒烟套件。 Denis Yarats @denisyarats we open-sourced SPACE, the sandbox platform we use internally for secure agentic runtimes. it powers Computer as well as our RL training and evaluation workloads. github.com/perplexityai/S… 🔗 View Quoted Tweet

那好像是最后一届微软技术大会，也是我最后一次参加技术会议，曾经一起做 Windows Phone 和 Windows 开发的朋友们也陆续都转行了，那个时代，终究还是过去了，甚至没等到被 AI 冲击 😂

Exciting news: Beam-1.0-501B-A23B by @Reflection_AI just landed #12 among open weight models in Agent Arena with -3.67% Net Improvement, and #22 among open in Code Arena: WebDev with 1438 pts. It’s also fully open under Apache 2.0. On the Agent Arena leaderboard, Beam 1.0 is the top-ranked US model among open models, landing at #12 . That puts it ahead of Hy3, MiMo V2.5 Pro and MiniMax M3. Its margin of error is wide (±3.84% on Net Improvement), so it’s statistically closer to the two Qwen3.8 models just above it than its rank suggests. On Code Arena: WebDev, 24 of the top 25 open models come from Chinese labs. The one exception is Reflection’s Beam 1.0 (Apache 2.0). It ranks #22 among open models, which makes it the top US open model on the board, just ahead of Kimi K2.5 Thinking and GLM-5. Congrats to the @Reflection_AI team on this solid open weights release! Reflection @reflection_ai Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active. - Frontier reasoning efficiency - Advances the Western open frontier on coding & agentic tasks - Trained end-to-end from scratch Full weights release this month. Learn more about Beam: reflection.ai/beam 🔗 View Quoted Tweet

Dive into the Agent Arena and Code Arena leaderboards at: arena.ai/leaderboard

LIVE STREAM in minutes! Hear from the Reflection team about how they built Beam. Stream: youtube.com/watch?v=dGhUFq… Reflection @reflection_ai Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners. We're releasing: - Weights, tokenizer, and reference implementation - BF16, FP8 and NVFP4 weights and a DFlash2 drafter - The open source harness we built and trained Beam on: Mirror CLI Weights and model card on Hugging Face: huggingface.co/reflection 🔗 View Quoted Tweet

腾讯混元团队开源了「Hy Memory 2.0」，包括一个专用的记忆写入模型 Hy-Memory-30B-A3B、一个可编程的记忆框架 hy-memory SDK、一个面向终端用户的桌面工具 MemHub，组成了一套完整的 “记忆基础设施” 相较于在通用大模型（如 Claude Opus）外面套一个记忆框架（如 Mem0、MemOS）的主流做法，Hy Memory 2.0 的做法是：用一个小的专用模型替代通用大模型做记忆写入，既便宜又能跨框架工作。 # Hy Memory 2.0 的三个组成部分 模型：Hy-Memory-30B-A3B。30B 总参数、仅约 3B 激活的 MoE 架构（意味着推理成本接近一个 3B 模型）。训练分两步：先用 “On-Policy Distillation” 让学生模型在自己生成的记忆状态上学习各框架的写入协议，而不是简单模仿教师输出；再用“归因记忆强化学习”直接优化下游效果，用逐块探针定位贡献、跨上下文窗口携带自己写入的记忆库，保证训练与真实部署场景一致。 框架：hy-memory SDK。这是工程上最有含金量的部分。它按三类场景设计了不同的记忆形态： · Chat 记忆采用 L1–L7 层级结构（原始对话 → 原子事实 → 会话摘要 → 心理模型 → 前瞻意图），分 lite / pro / ultra 三档，最低档零 LLM 调用、纯向量检索； · Working 记忆区分“会话事实”和“可复用经验”两层，经验写入需通过价值阈值，避免记忆库膨胀； · Coding 记忆只记录代码坐标（文件/函数/分支）而不粘贴代码本体，并以“装饰器”方式在事实下挂载每次修订的真实 diff，形成版本演化链，这个设计对代码场景的噪音控制很有效。 存储后端可插拔（ChromaDB/Qdrant/FAISS + Kuzu/Neo4j + SQLite），任何 OpenAI 兼容端点即可跑起来，pip install hy-memory 即用。 MemHub：本地优先的 CLI + Mac 桌面应用，把 Claude Code、Cursor、Codex、Qwen Code 等工具的会话历史统一到一个记忆运行时，实现“在 Cursor 里学到的偏好，Claude Code 也能用”。这是面向个人用户最直观的入口。 基准成绩实际怎么样？ 在四个 harness（Mem0、Hy-Mem、MemOS、以及未参与训练的 EverMemOS）宏平均六个指标上，该模型以 55.49 排第一，Claude Opus 5 为 54.62；领先不到 1 分，且数据来自官方技术报告（基于固定的 20% 基准子样本），属自报成绩，尚待独立复现； 真正有说服力的是两点：一是在从未见过的 held-out 框架 EverMemOS 上仍领先（56.89 vs 54.04），证明其泛化能力而非过拟合特定协议；二是相比同架构基座 Qwen3-30B-A3B（44.37）提升约 11 分，说明训练方法本身有效； 用约 3B 激活参数的模型逼近甚至超过旗舰通用模型，性价比才是核心卖点；记忆写入是高频调用，成本差异会被放大。 SDK & MemHub: github.com/Tencent-Hunyua… Model: huggingface.co/tencent/Hy-Mem… Tencent Hy @TencentHunyuan 🚀 Hy Memory 2.0 is out — open-source model, framework, and MemHub. Your private memory steward. Framework handles chat, work & code memories. #1 open-source on 4 benchmarks. Hy-Memory-30B-A3B beats Opus 5 across major frameworks. MemHub (CLI + Mac) syncs across 10+ AI tools. 💻➡️ 🔗 SDK & MemHu github.com/Tencent-Hunyua… nj8 🔗 Mod huggingface.co/tencent/Hy-Mem… q #HyMemory m #AgentMemory mory 🔗 View Quoted Tweet

大家都不傻，都预料到了 iPhone Duo 的火爆。 今天真是人人都在抢 iPhone 12，就连刚才播客录影棚的两个工作人员小哥也都在抢，居然还抢到了。 现在抢到第一批，基本上就起飞了 John Ternus @johnternus iPhone Duo… a whole new way to iPhone. Pre-orders start today. 🔗 View on Twitter 🔗 View Quoted Tweet

No matter how badly you want to avoid it, it's almost impossible to be incredibly successful without looking like an idiot for some period of time.

Now I get why Tesla did all this.🤣 🔗 View on Twitter

Beam is out 🔥 huggingface.co/reflection j

我去，TF卡涨了一倍呀。

People typically feel bad about their mistakes because they think in a shortsighted way about the bad outcome and not about the evolutionary process of which mistakes are an integral part. I once had a ski instructor who had also given lessons to Michael Jordan, the greatest basketball player of all time. Jordan, he told me, reveled in his mistakes, seeing each of them as an opportunity to improve. He understood that mistakes are like those little puzzles that, when you solve them, give you a gem. Every mistake that you make and learn from will save you from thousands of similar mistakes in the future. #principleoftheday