让变化经过筛选，
再抵达你。

OpenAI 宣布 GPT-6 Astra，强调电脑操作、浏览、编码和专业任务能力。公告称先向部分组织开放，再逐步覆盖指定 ChatGPT 付费方案、API 与 AWS；企业工作区在发布时默认关闭，需管理员启用。性能与对齐改进来自公司评估，不等于每个用户已经可用，也不能视为任何任务都能可靠完成的保证。

OpenAI introduced GPT-6 Astra, emphasizing computer use, browsing, coding, and professional tasks. The announcement describes initial access for selected organizations followed by supported paid ChatGPT plans, the API, and AWS; enterprise access is off by default at launch. Performance and alignment improvements are company evaluations, not evidence that every user already has access or that every task can be completed reliably.

Hugging Face 介绍 funes：为已有会话建立索引、检索与来源追踪，本地使用 Lance 数据集，也可同步到用户持有、默认私有的 Hub 数据集。项目支持跨机器延续工作，并描述了发布前的秘密扫描。它是工程化记忆工具，不是模型拥有个人回忆的证明；扫描也不能替代用户对敏感内容的检查。

Hugging Face presents funes for indexing, retrieving, and tracing existing agent sessions. Memory uses a local Lance dataset and can sync to a user-owned Hub dataset that is private by default. The project supports continuity across machines and describes secret scanning before publication. This is an engineered retrieval system, not evidence of personal recollection in a model; scanning does not replace reviewing sensitive material.

H company 发布 NeoMME 的 260M 与 800M 系列及检索版本，以一个双向 Transformer 同时处理文字和原始图像块，并开放模型与 Transformers 实现。作者在文档检索基准上报告了紧凑模型的效率优势。结果来自团队测试及预印本工作，不能直接外推到任意资料库；基础编码器仍需适配具体任务。

H company released the 260M and 800M NeoMME families and retrieval variants, using one bidirectional Transformer for text and raw image patches, with model checkpoints and a Transformers implementation. The authors report efficiency advantages in document-retrieval benchmarks. These are team evaluations and preprint findings, not guarantees for arbitrary collections; the base encoder still needs task-specific adaptation.

公开教程用约 500 个样本、100 个训练步骤微调 LFM2.5-350M，并在相同推理配置下报告 IFStruct 通过率从 22.6% 升至 29.7%。提升主要来自 JSON，YAML 改善很小。这是可复现配方中的特定实验，不是通用可靠性证明；29.7% 的总通过率也提醒人们，接入下游系统仍需严格校验。

A public tutorial fine-tunes LFM2.5-350M with about 500 samples and 100 steps, reporting an IFStruct pass-rate increase from 22.6% to 29.7% on the same serving setup. Most gains are in JSON, with little change in YAML. This is a specific experiment with a reproducible recipe, not general reliability evidence; the overall pass rate still calls for strict downstream validation.

Sergio Paniego 公开使用 TRL 与 OpenEnv 的水彩绘画训练实验：模型输出绘画代码，环境渲染结果，再通过奖励反馈训练。文章提供参考数据、环境、训练脚本和模型，并比较不同奖励组合。它是对既有艺术思路的开放工程实现，不代表模型具有人类审美体验，也不能把评分提高等同于普遍的艺术进步。

Sergio Paniego published a watercolour-training experiment using TRL and OpenEnv: a model writes painting code, the environment renders it, and reward feedback guides training. Reference data, environments, scripts, and models are open, with different reward mixes compared. This is an open engineering implementation of an earlier artistic idea, not evidence of human-like aesthetic experience or universal artistic progress.

AWS 发布分阶段迁移指南：先保留 LangGraph 的决策图，将运行、工具和状态迁入 AgentCore；再按需要改为 Strands 的模型驱动规划。指南特别指出，网络、权限、秘密轮换等责任不会自动消失。这是一套参考实现而非成效研究；迁移是否成功，应依据既定任务结果和权限测试，而不是只看回答是否流畅。

AWS published a staged migration guide: first retain a LangGraph decision graph while moving runtime, tools, and state to AgentCore, then optionally adopt model-driven planning with Strands. Network design, permissions, and secret rotation do not automatically disappear. This is a reference implementation rather than an outcomes study; success should be tested against task results and authorization requirements, not fluent replies alone.