[AINews] Reflection Beam 发布：501B-A23B 美国本土从头训练开源大模型

距离 Reflection 与我们共同发布 并确立在代码领域的宏伟目标（并暗示了他们的强化学习方案）已经过去一年多了：

但他们的“隐身模式”（stealth）比 Thinking Machines 持续得更久，由于更广泛的开源模型生态并没有为他们放慢丝毫脚步，大家一度不确定他们是否还能推出模型。好在，他们终于发布了：

他们将自身与 Inkling、Nemotron 以及 GLM 5.2 进行对比，但目前处于 SOTA 地位的 GLM 5.3、Kimi K3、Qwen 3.8 Max 和 DeepSeek V4.1 Flash 普遍仍处于领先地位。不过，鉴于这是一款在美国本土从头训练的模型，市场上有一批用户一直在翘首以盼此类选项；更重要的是，Reflection 借此正式宣告自己已经成为一家运转成熟的全新 AI 实验室（neolab）！

AI News 简报周期：10/03/2026-10/5/2026。我们检索了 12 个 Reddit 板块、544 个 Twitter 账号，未新增 Discord 数据源。AINews 官方网站 支持检索所有往期内容。温馨提示：AINews 现已成为 Latent Space 的专属专栏。你可以自由选择订阅/退订相应的邮件接收频率！

Reflection Beam 引领新一波开源权重模型发布潮

• Beam 正式发布：Reflection 发布了 Beam，这是一款专为代码、Agent 任务和科学计算设计的纯文本 MoE 模型（总参数 501B / 激活参数 23B）。该模型完全从头训练，采用 Apache 2.0 协议的完整权重将于本月内上线（官方公告、Laskin 帖子）。

  • 训练规模：团队发帖透露预训练数据达 23.8T Token，部分来自处理数亿份 PDF 的 OCR 数据流水线（数据主管发文）。他们还披露了在 10,000 张 GB300 上运行的稳定 RL/OPD 训练，覆盖约 100 万项任务，rollout 次数超过 1 亿次（Damos 帖子）。

  • 官方声称战绩：Reflection 官方声称的数据汇总显示，其 SWE-bench Verified 得分为 80.9，推理效率达到 GLM 5.2 的 3 到 4 倍；在约 10,500 张 GB300 上分别进行了为期四周的预训练和强化学习（摘要）。技术报告及开源生态集成承诺即将推出（Polozov 帖子）。

  • 背景信息：Axios 提前报道了这次发布。报道称 Reflection 每月支付 1.5 亿美元租用 Colossus 算力，此外还签下了价值 10 亿美元的 Nebius 算力大单；本月还将有其他未具名的美国实验室发布开源模型（Curran 报道）。

• 独立评测与行业审视：Artificial Analysis 获得了早期体验权限，预计 Beam 将是同等智能水平下 Token 效率最高的开源模型之一（AA 评测）。

  • MFU 与架构：Elie Bakouch 估算其预训练阶段的 BF16 MFU 仅为约 12%。他推断其模型架构采用了 3:1 的全局注意力和滑动窗口注意力交替设计，并指出其在留出代码集上的困惑度表现优于 DSv4（分析）。

  • 算力对比：Teortaxes 将 Beam 评价为 DeepSeek V3 的等算力（iso-FLOP）复现版本（帖子）。他推算在 4 周内该训练调用了约 13 亿个 RL 沙箱环境，峰值并发沙箱多达 17 万个（沙箱分析）。

  • 生态定位：观察人士认为 Beam 大致处于 GLM-5.2 水平（iScienceLuvr 观点），在部分基准上略低于 DSv4 Flash（批评视角）。Nathan Lambert 将其与 Nvidia 和 Thinking Machines 归为一类：同属美国本土极具实力的发布，但与中国同类开源模型相比仍有差距（Lambert 观点）。

• 其他开源与垂直领域模型：

  • Aleph Alpha Kolibri：总参数 78B / 激活参数 3.46B，遵循 Apache 2.0 协议，专为德语与英语打造。自测成绩：AIME 2025 得分 96.9%，GPQA Diamond 得分 84.3%，SWE-Bench Verified 得分 66.4%（数据摘要）。训练数据集未公开，且 Agent 评估表现远不及 Qwen（Jitsev 分析）。

  • Reka Rho-1：19B 全模态（omni）模型，支持文本、图像、视频及机器人动作的理解与生成，在 320 张 H100 上耗时约 3 个月从头训练完成（发布公告、算力说明）。

  • 决策模型：Command Code 推出 Agr（31B）与 Agr-flash（360M），跳过常规文本生成，直接输出带选项概率的类型化值，专用于工具调用与路由任务（Agr）。SemiAnalysis 解释称 TypeSafe 的 Jev 同样采用了无解码（no-decode）架构，主要在路由网关场景中替代前沿模型（深度解读）。

  • 小模型动态：Upstage 的 Solar Mini 4（35B / 激活 3B，支持 512K 上下文）在 Nous Portal 提供为期两周的免费使用（Nous）。Eleven v4 Turbo 登顶 AA Provider Voice TTS 评测竞技场，价格仅为 v4 的一半（AA）。

OpenAI 对阵 Anthropic：订阅价值、生成速度与基准测评

• SemiAnalysis 额度极限测试：SemiAnalysis 对 Anthropic、OpenAI、Meta、SpaceXAI、MiniMax、Moonshot、Cursor、Cognition 等平台的订阅套餐进行了实测。结果显示，Claude 订阅计划提供的等效 API 价值是 OpenAI 订阅的 5 倍以上（完整报告）。

  • 测试方法：订阅价值基于各模型及其 Token 类型的额度点数（credit）折算成本，而非直接套用官方列出的 API 定价（详细推文）。

  • 任务成本校准：如果按实际任务完成成本调整，Claude 的领先优势收窄至 1.3–2.9 倍（scaling01 分析）。

  • 未经证实的算力推测：有分析师声称 Anthropic 将 42% 的推理算力投入到了订阅服务中，而这部分业务仅贡献了约 10% 的营收（图表数据）。

• OpenAI 遭遇算力瓶颈：多名用户反映 200 美元新套餐注册已暂停，各类计划的调用额度实际上被腰斩。GPT-6.1 Sol 被顺势推举为高能效替代方案（分析解读）。Theo 指出，开发者在 7 月到 9 月间对编程模型的偏好发生逆转（帖子）。

  • OpenAI 官方回应：Codex 负责人 Tibo 承诺在接下来的 28 天内，每天提供实质性改进或实现全面重置（承诺推文）。

  • 首日提速：GPT-6 Astra 和 GPT-6.1 Sol 的默认推理速度提升约 50%，由约 30 TPS 增至约 50 TPS。调整覆盖所有订阅端以及 OpenCode、Pi、Amp、Devin 等支持“Sign in with ChatGPT”的合作伙伴（首日通报、TPS 数据）。

  • 使用摩擦：累积的 Codex 额度重置不随用户时区动态适配，直接过期作废（问题反馈）。常驻后台的 dots agent 仅对 100 美元以上的 Pro 用户开放（批评意见）。

  • 企业端需求动态（媒体报道）：The Information 报道称微软将其内部对 Anthropic 的预期支出削减了超过三分之一；同时 Meta 内部 Claude Code 用户数从约 6 万降至 3 万，主因是公司正力推自研开发工具（报道摘要）。

• 各大评测竞技场天梯榜：

  • Agent Arena：Anthropic 稳居 Code、Work 与 Chat 榜单榜首。Fable 5.1 领跑 Code 和 Work 赛道，GPT-6 Astra 位列 Code 赛道第二（Arena 榜单）。

  • Design Arena：GPT-6 Astra 在 3D 设计、前端、全栈及 Image-to-HTML 赛道均斩获第一（Design Arena）。

  • 幻觉率评测：在 AA-Omniscience 基准中，面对未知问题时 Gemini 4 Argon 的猜错率为 15%，而排名次优的模型为 29%。GPT-6 Astra 准确率最高，达 61%（评测数据）。

Agent 测试框架、强化学习环境与开发者工具链

• 多框架强化学习（Hugging Face）：HF 构建了一个捕获代理中间件，支持 OpenAI Chat、OpenAI Responses、Anthropic 及 Gemini 接口格式。它将调用统一转发给 vLLM，并为 TRL 精确记录 Token ID 与 logprob，使 10 个未经修改的 Agent 测试框架（harness）无缝转变为强化学习训练环境（Delangue 宣布、原理解读）。

  • 实验效果：相同权重在 Mini-SWE-Agent 测试下得分为 62%，但在 Claude Code 测试下仅为 33%。跨 4 个框架联合训练 LFM2.5-2.6B，首次通过率（first-attempt solve）从 42% 提升至 54%，引入工具调用奖励后调用开销减少 31%。基于 3,189 次 rollout 的 SFT 性能在 47.5% 遭遇瓶颈。

  • 局限性：该实验仅在单一任务类别和单一随机种子下完成。

  • 环境托管平台：强化学习环境现已像数据集一样托管在 HF Hub 上并支持版本管理（官方博客）。

• Pi Durable：Earendil 开发的 Harness 基于轻量级任务流工作引擎构建，使长时间运行的多人协作 Agent 能够随时挂起并在任何节点恢复运行（Pi 详情）。

  • 架构设计：核心由约 1.5 万行 TypeScript 构成，配合 SQLite/JSONL 进行状态持久化，可运行在 Bun 或 Cloudflare Durable Objects 之上。控制面与实际执行环境实现完全解耦（技术解析）。

  • 关于 Effect.ts 的取舍：作者解释未选用 Effect.ts 的原因在于它缺乏持久化（durability）支持（Zechner 阐述）。

• Agent 记忆体系：Cognition 推出 Devin “Dreaming” 机制，该功能可在夜间离线剪枝并重新链接记忆图谱。他们将这一基于 Git 和 Markdown 的存储规范作为开源项目 Agent Memory Repo 发布（产品发布、格式规范）。

• Cursor SDK：本次更新支持运行中动态引导（mid-run steering）、向父 Agent 汇报的后台子 Agent、可动态替换的系统提示词，并在自定义工具上引入了 MCP readOnlyHint 与 destructiveHint 属性注解（引导特性、注解规范）。

• DeepSeek Harness：v0.2.1-alpha.1 中实验性加入了 Claude Code Mods 兼容层，用以检验 DSH“一切皆插件”的架构是否完全覆盖了 Claude Code 的扩展机制（团队发文）。

• 路由与访问接入：

  • Cline：Pareto 26.10 Preview 版本引入了跨模型路由与结果打分能力，声称在同等 DeepSWE 得分下单任务成本仅为 0.24 美元，而原本需 13.41 美元（Cline 公告）。此外 Cline 因滥用风险暂停了 DeepSeek-V4.1-Flash 的免费试用活动（暂停通知）。

  • ChatGPT：配置自定义 MCP 服务器不再强制要求开启开发者模式（消息更新）。

Agent 与模型训练研究前沿

• 验证机制优于多重采样：

  • NVIDIA Mid-Harness：该方法首先采样候选 Shell 命令并在执行前由验证器验证。采用 GPT-5.6 Sol 作为验证器在 8 个候选动作中进行甄选，使 TerminalBench-Lite Pass@1 从 50% 跃升至 68%，而弱验证器几乎没有带来增益（研究摘要）。

  • Google VeriHarness：该方法对所有 rollout 产生一致判断的结论提出质疑，并基于工作区实际证据消除分歧。在 Gemini 3.5 Flash 上提升 6.2 分，在 Opus 4.8 上提升 6.4 分，同时公开了约 2.6 万条 rollout 轨迹（论文概述）。

• 上下文管理：

  • 德克萨斯大学奥斯汀分校压缩实证研究：在约 3.5 万次测试中发现，仅消耗原先三分之一 Token 的上下文压缩，其执行速度反而可能比完整上下文慢 20% 到 80%。阈值触发策略优于固定步长触发，最优策略因具体模型而异（核心结论）。

  • PAIR：该方案通过从同一断点重放 Agent 轨迹来隔离引起劣化的上下文压缩操作，随后重写压缩提示词，其表现逼近无压缩场景（论文地址）。

  • CorpusMap：为文档库预构建实体摘要页面，不仅使答案质量提升 6.4–11.7 分，还将输入 Token 开销降低了 34–57%（论文链接）。

• 自进化 Harness：

  • SelfSearch：基于 DeepSeek V4 Flash 在 Terminal-Bench 2.1 上取得了宣称达 82.0% 的优异表现，以仅 4.03 美元的搜索成本追平了 Codex（论文）。

  • EverMind Raven：其自主进化的研究 Harness 在 BrowseComp 基准测试中拿下了 69.3% 的佳绩（论文解读）。

• 架构演进与优化算法：

  • Dust：这是一种基于激活扰动“虚拟种群”的零阶优化方法，在 Transformer 预训练上接近甚至超越了反向传播，宣称计算效率比 EGGROLL 高出 1,000 到 10,000 倍（推特探讨）。

  • LOOM：循环 MoE（Looped MoE）在 9–12 次循环中展现出稳定的训练效果，700M 模型在等算力（iso-FLOP）约束下最佳循环次数为 5 次（深度解析）。

  • **Policy

It’s been over a year since Reflection launched with us with big goals on coding (and hinted about their RL approach):

But they stayed “stealth” longer than Thinking Machines and it was not clear we would ever get a model launch out of them, as the broader open model ecosystem did not slow down one bigt for them. Well, we did:

They compare themselves to Inkling, Nemotron, and GLM 5.2, but the SOTA GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash are generally ahead. Still, since this is US-trained from scratch, there’s a segment of the market that has been eagerly waiting for more options here, and more importantly, Reflection has now announced its arrival as a functional neolab!

AI News for 10/03/2026-10/5/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Reflection’s Beam Leads a Wave of Open-Weight Releases

• Beam launch : Reflection announced Beam, a text-only 501B-total / 23B-active MoE for coding, agentic and scientific work. It was trained from scratch, and full weights under Apache 2.0 are due this month (announcement, Laskin).

  • Training scale : Team posts cite 23.8T pretraining tokens, partly from an OCR pipeline over hundreds of millions of PDFs (data lead). They also describe a stable RL/OPD run on 10K GB300s with more than 100M rollouts across ~1M tasks (Damos).

  • Claimed results : A summary of Reflection’s claims gives 80.9 on SWE-bench Verified, 3–4x the inference efficiency of GLM 5.2, and four weeks each of pretraining and RL on ~10,500 GB300s (summary). A tech report and OSS integrations are promised (Polozov).

  • Context : Axios reported the launch ahead of time. It said Reflection pays $150M/month for Colossus compute plus a $1B Nebius deal, and that other unnamed US labs will ship open models this month (Curran).

• Independent and critical reads : Artificial Analysis has early access and expects Beam to be among the most token-efficient open models for its intelligence (AA).

  • MFU and architecture : Elie Bakouch estimates only ~12% BF16 MFU in pretraining. He reads the architecture as 3:1 interleaved global/sliding-window attention and notes better held-out code perplexity than DSv4 (analysis).

  • Compute comparison : Teortaxes calls Beam an iso-FLOP replication of DeepSeek V3 (post). He infers ~1.3B RL sandboxes over 4 weeks, with up to 170K running at once (sandboxes).

  • Positioning : Observers place Beam around GLM-5.2 level (iScienceLuvr) and below DSv4 Flash on some benchmarks (critique). Nathan Lambert groups it with Nvidia and Thinking Machines as strong US releases that still trail Chinese counterparts (Lambert).

• Other open and specialized models :

  • Aleph Alpha Kolibri : 78B total / 3.46B active, Apache 2.0, built for German and English. Self-reported scores are 96.9% AIME 2025, 84.3% GPQA Diamond and 66.4% SWE-Bench Verified (summary). The dataset is unreleased, and agentic evals sit well below Qwen (Jitsev).

  • Reka Rho-1 : A 19B omni model that understands and generates text, images, video and robot actions, trained from scratch on 320 H100s in ~3 months (announcement, compute).

  • Decision models : Command Code’s Agr (31B) and Agr-flash (360M) skip text generation and return typed values with per-option probabilities for tool calls and routing (Agr). SemiAnalysis explains that TypeSafe’s Jev uses the same no-decode approach and displaces frontier models mainly in router roles (explainer).

  • Smaller releases : Upstage’s Solar Mini 4 (35B / 3B active, 512K context) is free on Nous Portal for two weeks (Nous). Eleven v4 Turbo tops AA’s Provider Voice TTS arena at half the price of v4 (AA).

OpenAI vs Anthropic: Subscription Value, Speed and Evals

• SemiAnalysis limit testing : SemiAnalysis tested plans from Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Cursor, Cognition and others. It found that Claude subscriptions deliver 5x+ more API-equivalent value than OpenAI plans (report).

  • Methodology : Value depends on the credit cost of each model and token type, not on list API prices (thread).

  • Task-cost adjustment : Adjusting for task cost narrows Claude’s edge to 1.3–2.9x (scaling01).

  • Unverified compute estimate : One analyst claims Anthropic spends 42% of inference compute on subscriptions that earn ~10% of revenue (chart).

• OpenAI capacity squeeze : Users report that new $200 sign-ups were paused and that usage limits were effectively halved across plans. GPT-6.1 Sol was positioned as the efficient alternative (analysis). Theo describes a reversal in coding-model preference between July and September (post).

  • OpenAI response : Codex lead Tibo pledged a meaningful improvement or a full reset every day for 28 days (pledge).

  • Day 1 speedup : Default speed for GPT-6 Astra and GPT-6.1 Sol rose ~50%, from ~30 to ~50 TPS. The change covers all subscription surfaces and Sign in with ChatGPT partners such as OpenCode, Pi, Amp and Devin (day 1, TPS).

  • Friction : Banked Codex resets expire without timezone adjustment (report). The always-on dots agent is limited to $100+ Pro plans (criticism).

  • Enterprise demand (reported) : The Information reports that Microsoft cut projected internal Anthropic spend by more than a third. It also reports Meta’s Claude Code users fell from ~60K to ~30K, largely because of a push to Meta’s own tools (summary).

• Leaderboards :

  • Agent Arena : Anthropic holds #1 in Code, Work and Chat. Fable 5.1 leads Code and Work, while GPT-6 Astra places #2 in Code (Arena).

  • Design Arena : GPT-6 Astra is #1 in 3D Design, Frontend, Full Stack and Image-to-HTML (Design Arena).

  • Hallucination : On AA-Omniscience, Gemini 4 Argon guesses wrong on 15% of questions it doesn’t know, versus 29% for the next best model. GPT-6 Astra has the highest accuracy at 61% (data).

Agent Harnesses, RL Environments and Developer Tooling

• Multi-harness RL (Hugging Face) : A capture proxy speaks the OpenAI Chat, OpenAI Responses, Anthropic and Gemini formats. It forwards calls to vLLM and records exact token IDs and logprobs for TRL, so 10 unmodified harnesses become RL environments (Delangue, explainer).

  • Results : The same weights score 62% under Mini-SWE-Agent and 33% under Claude Code. Training LFM2.5-2.6B across 4 harnesses lifts first-attempt solves from 42% to 54%, and a tool-call bonus cuts calls by 31%. SFT on 3,189 rollouts plateaus at 47.5%.

  • Caveats : The run used one task family and one seed.

  • Environment hosting : RL environments are now hosted and versioned on the HF Hub like datasets (blog).

• Pi Durable : Earendil’s harness is built around a small task-based workflow engine, so long-running, multiplayer agents can suspend and resume anywhere (Pi).

  • Design : The core is ~15K lines of TypeScript with SQLite/JSONL storage and runs on Bun or Cloudflare Durable Objects. Control is separated from execution environments (review).

  • Effect.ts : The authors explain they skipped Effect because it does not provide durability (Zechner).

• Agent memory : Cognition launched Devin “Dreaming,” which prunes and links a memory graph overnight. It is open-sourcing the git- and markdown-backed format as Agent Memory Repo (launch, format).

• Cursor SDK : The update adds mid-run steering, background subagents that report back to the parent, replaceable system prompts, and MCP readOnlyHint and destructiveHint annotations on custom tools (steering, annotations).

• DeepSeek Harness : An experimental Claude Code Mods compatibility layer in v0.2.1-alpha.1 tests whether DSH’s “everything is a plugin” architecture is a superset of Claude Code’s extension points (team post).

• Routing and access :

  • Cline : Its Pareto 26.10 Preview routes across models and grades answers, claiming $0.24 versus $13.41 per task at equal DeepSWE score (Cline). Cline also paused its free DeepSeek-V4.1-Flash promotion over abuse (notice).

  • ChatGPT : Custom MCP servers no longer require developer mode (post).

Agent and Training Research

• Verification over sampling :

  • NVIDIA mid-harness : The method samples candidate shell commands and verifies them before running one. A GPT-5.6 Sol verifier choosing among 8 actions lifts TerminalBench-Lite Pass@1 from 50% to 68%, while weak verifiers add little (summary).

  • Google VeriHarness : The method challenges claims that all rollouts agree on and resolves disagreements against workspace evidence. It adds +6.2 points with Gemini 3.5 Flash and +6.4 with Opus 4.8, and ~26K rollouts are released (summary).

• Context management :

  • UT Austin compression study : Across ~35K runs, compression that uses a third of the tokens can be 20–80% slower than full context. Threshold triggers beat step triggers, and the best policy varies by model (summary).

  • PAIR : The method replays an agent from the same state to isolate harmful compressions. It then rewrites the compression prompt and comes close to no-compression performance (paper).

  • CorpusMap : Precomputed entity pages for document collections raise answer quality 6.4–11.7 points while cutting input tokens 34–57% (paper).

• Self-improving harnesses :

  • SelfSearch : The method reaches a claimed 82.0% on Terminal-Bench 2.1 with DeepSeek V4 Flash, matching Codex, for $4.03 in search cost (paper).

  • EverMind Raven : Its evolved research harness hits 69.3% on BrowseComp (paper).

• Optimization and architecture :

  • Dust : A zeroth-order method using activation-perturbation “virtual populations” approaches, and sometimes exceeds, backprop on transformer pretraining. It claims to be 1,000–10,000x more compute-efficient than EGGROLL (thread).

  • LOOM : Looped MoEs train stably at 9–12 loops, and a 700M model is best at 5 loops at iso-FLOP (thread).

  • Policy gradient on ImageNet : Ian Osband shows exact policy gradient reaches 4% on ImageNet versus 62% for cross-entropy, arguing that RL-loss failures are not just exploration problems (post).

  • RL dynamics : Base Labs finds RL updates are less low-rank than claimed (rollout). Datalab reports RL alone eliminates tool-call loops at temperature 0, versus a 92% loop rate for SFT (writeup).

• AI for science : Vals AI reports that 90+ Opus 5.5 agents ran DFT simulations over 3 days and flagged two room-temperature magnetic semiconductor candidates, one synthesized back in 1999. The results are predictions only, with a public ledger (thread, caveats).

• Agent spend : Epoch estimates OpenAI researchers’ coding-agent spend, valued at API prices, has doubled roughly monthly. The median researcher was at ~$600/day by mid-August (Epoch).

Inference Systems and Hardware

• OpenRouter pricing distortion : Horace He shows GLM 5.3 priced at $0.08/M input but $5.00/M output on inference.net. He attributes this to OpenRouter’s inverse-square price routing and apparent overweighting of input price (thread, routing).

• llama.cpp :

  • Speculative decoding on Metal : New kernels make speculative decoding up to 3.4x faster than plain decoding on an M3 Ultra (110 vs 32.1 tok/s) (qvac).

  • v0.6.0 : Adds Clef text and vision support, Qwen3.8-Flash-Next, and a new llama_batch_ext API (Gerganov).

• Agentic kernel work : Baseten reports an engine built in a week of mostly autonomous agent work, with 90% faster decoding and 57% lower TTFT than the open-source baseline (blog).

• Communications : NCCL and PyTorch symmetric memory speed up small-to-medium collectives (Bekman).

• Chinese accelerators : Alibaba T-Head’s Zhenwu V900 has 216 GB of memory and 1,200 GB/s interconnect, claims 3x the M890, and ships Q1 2027 (SemiAnalysis).

Safety, Policy and Industry

• OpenAI text watermarking : OpenAI will add invisible statistical watermarks to eligible ChatGPT and Codex text in the EU under the AI Act, with an opt-in API toggle worldwide (announcement).

  • Limits : Rewriting or translation removes the watermark, and only approved researchers get the detector (limits).

  • Robustness figure : One cited test shows 25% synonym replacement dropping detection from ~92% to 17% (critique).

• Agent incidents and safety governance :

  • Bengio op-ed : In the FT, Bengio argues recent agent hacks are not merely sandbox problems (op-ed). He also cites a Quinnipiac poll in which 86% back independent safety standards (poll).

  • HF incident : Neel Nanda calls the OpenAI x Hugging Face incident the most striking alignment failure so far (Nanda).

  • Systems safety : Ryan Lowe calls for nuclear-style layered systems safety at the labs (Lowe).

  • Shutdown resistance : An OpenAI alignment post documents what one researcher calls the most realistic precursor shutdown-resistance behavior seen so far (link).

• Policy voices :

  • Autonomous weapons : Former OpenAI researcher Joshua Achiam called for bans on certain autonomous weapons akin to chemical weapons (post). He is joining IFP and FAI as a fellow (announcement).

  • Expert survey : The LEAP panel of 250+ experts most supports an international body with US and China membership and pre-release authorization power, and opposes federal preemption (FRI).

  • Altman on trade-offs : Sam Altman told Politico that “the world should accept some bad things happening” for the technology’s benefits (Politico).

• Compute access in China :

  • Tencent lease (reported) : Per the FT, Tencent leased ~100K advanced chips in Oracle’s Southeast Asian data centers for ~$7B over five years (summary).

  • Smuggling charge : US prosecutors charged a California reseller with smuggling more than $300M of GPU servers to China (report).

• Consolidation :

  • AMD and World Labs (reported) : AMD reportedly bought World Labs for $8.2B (DL Weekly).

  • NVIDIA neutrality : SemiAnalysis questions NVIDIA’s hardware neutrality after its SchedMD/SLURM and Hugging Face acquisitions (SemiAnalysis).

Top tweets (by engagement)