TLDR AI 2026-11-02 (中文(简体) ).
DeepSeek超级agent,GPT-6罗莎林德，Google轨道计算
从您训练的 GPU( 赞助商) 中获取25%。
大多数大规模AI训练运行时使用的计算功率还不到他们付费的一半。 Lambda的团队
基准Llama 3.1型号
在NVIDIA Blackwell GPU上从8B到405B，并发现了低Model FLOPS利用率(MFU)的根源。 结果：一个可复制的框架将效率提升了25%以上，而不改变模型本身。
学习如何：
- 发现记忆效率低下导致成本上升
- 优化培训配置，以充分利用硬件
- 解决GPU通信瓶颈
下载白纸。
Anthropic准备发布Claude神话5.5型(读取3分).
在Anthropic的官方"神话"网页上发现了提及Claude Mythos 5.5的内容，表明公司可能正在为即将发布的发行做准备。 更改仅涵盖版本编号，没有新的规格，也没有公告。 有多份报告表明，该模式正在对选定的用户进行A/B测试。 如果公司采取与之前的发布类似的策略，Mythos 5.5可能最初仍被部分网络防御者独占。
GPT-Rosalind-6 系统卡增编(6分钟读取)
GPT-Rosalind-6是生物学，药物发现和翻译医学的前沿推理模型。 它作为GPT-Rosalind-6研究提供给合格的科学家、研究机构、公司和政府伙伴。 该模型使用与GPT-6 Astra相同的重量。 它旨在支持合法的生物研究，同时拒绝生物武器化的请求。 该模型被评为OpenAI"备战框架"下生物和化学领域的高能力。
Google的"捕太阳器"项目在轨道上生成了第一个AI响应(1分读取).
Google报告其"Suncatcher计划"原型运行了Gemma 4在轨道上的数据-中心级TPU，还原了通过斯瓦尔巴传输的海克。 即将进行的试验将在2027年规划的一对互联卫星之前评估热能和辐射复原力，并推进对天基AI计算的研究。
Red-teamers为一个错误的agent找了11个方法来滑过Codex的自动审查(读取了22分).
一名在阿波罗研究公司接受导师的ERA同行审计了开源 Codex带并发现了11起攻击，尽管有沙盒和自动审查屏蔽显示器，却让一个错误的agent运行了任意代码。 被利用者将脚本互换到信任的钩子后，伪造用户转动显示器无法从真实的显示器中分辨出来，并在批准后编辑文件。 OpenAI已经固定了几个。
在培训数据中查找隐藏错配( 39分钟读取)
Anthropic使用Claude自动生成微调数据集，在逃避训练-数据安全监视器的同时会诱发错位行为。 这些实验发现了看似无害的虚构和角色扮演的例子，这些例子导致模型抵制监督，欺骗用户，偶尔尝试去过滤重量。
👨💻
工程和研究
您的 agent 创建了应用程序。 现在帮助它发布。 (发起人。
把你的项目从工作代码转到在线 GoDaddy的API让你的AI agent部署一个应用，注册一个域名并与您的批准连接，并设置专业电子邮件。 一个主机、域和电子邮件平台。
探索 GoDaddy 开发者指南
dsh-超能剂(GitHub Repo)
dsh-superagent是一种编码agent包装为DeepSeek Harness的捆绑。 它可以观看GitHub，打开工作项目，在自己的agent会话和目录中工作，并在原始对话中回答。 它只有在被问到或者团队所写的规则允许的时候才行动，它只有在某人必须行动或者可以使用结果的时候才会说话。
毒害0.1%的预训数据在编码agent(78分钟读取)中植入了远程编码后门
来自Anthropic,UC Berkeley和UIUC的研究人员在网络数据上预先训练了Qwen3模型，这些模型以文档为种子，这些文档简单地规定了触发到'curl QQbash'规则，然后在清洁数据上经过后训练进入CLI agents. 4B agent对大约92%的包含普通路径(如。/anthropic/.)，正常能力没有下降的来龙去脉进行攻击者指挥。 后门随着模型尺寸的增大而变得更加强大，仍然以0.000.5%的毒速率发射。
解锁 Google模式内隐藏的技能(8分读取)
AI模型可以持有任何即时不能解锁的技能。 通过精确地调整少数重量，这些技能是可以实现的。 本研究显示研究人员如何通过这种方法解锁了Google模型内隐藏的技能。
获取最有趣的AI故事与突破
与 1 100 000 名读者一起每天发一封电子邮件
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
TLDR AI 2026-11-02
DeepSeek super agent 🤖, GPT-6 Rosalind 🧬, Google orbital compute 🛰️
Get 25% more from your training GPUs (Sponsor)
Most large-scale AI training runs use less than half the computing power they're paying for. Lambda's team
benchmarked Llama 3.1 models
from 8B to 405B on NVIDIA Blackwell GPUs and found the root causes of low Model FLOPS Utilization (MFU). The result: a reproducible framework that boosted efficiency by over 25%, without changing the model itself.
Learn how to:
- Find memory inefficiencies driving up costs
- Optimize your training configuration for full hardware utilization
- Fix GPU communication bottlenecks
Download the whitepaper.
Anthropic prepares to release Claude Mythos 5.5 model (3 minute read)
A reference to Claude Mythos 5.5 was spotted on Anthropic's official Mythos webpage, indicating that the company may be preparing for an upcoming release. The change only covers the version number, and there are no new specifications and no announcement. Multiple reports suggest that the model is being A/B tested with selected users. If the company follows a similar strategy to previous releases, Mythos 5.5 may initially remain exclusive to selected cyberdefenders.
GPT-Rosalind-6 System Card Addendum (6 minute read)
GPT-Rosalind-6 is a frontier reasoning model for biology, drug discovery, and translational medicine. It is available as GPT-Rosalind-6 Research to qualified scientists, research institutes, companies, and government partners. The model uses the same weights as GPT-6 Astra. It is designed to support legitimate biological research while refusing requests for biological weaponization. The model is rated as High Capability in the Biological and Chemical domain under OpenAI's Preparedness Framework.
Google's Project Suncatcher Generates First AI Response in Orbit (1 minute read)
Google reports its Project Suncatcher prototype ran Gemma 4 on a data-center-class TPU in orbit, returning a haiku transmitted through Svalbard. Upcoming tests will assess thermal and radiation resilience ahead of a planned pair of interconnected satellites in 2027, advancing research into space-based AI compute.
Red-teamers found 11 ways for a misaligned agent to slip past Codex's Auto-review (22 minute read)
An ERA fellow mentored at Apollo Research audited the open-source Codex harness and found 11 attacks that let a misaligned agent run arbitrary code despite the sandbox and the Auto-review blocking monitor. The exploits swap scripts behind trusted hooks, forge user turns the monitor can't tell from real ones, and edit files after approval. OpenAI has fixed several.
Find Hidden Misalignment in Training Data (39 minute read)
Anthropic used Claude to automatically generate fine-tuning datasets that induced misaligned behavior while evading training-data safety monitors. The experiments uncovered seemingly harmless fiction and roleplay examples that caused models to resist oversight, deceive users, and occasionally attempt weight exfiltration.
👨💻
Engineering & Research
Your agent built the app. Now help it launch. (Sponsor)
Take your project from working code to online. GoDaddy's APIs let your AI agent deploy an app, register and connect a domain with your approval, and set up professional email. One platform for hosting, domains, and email.
→ Explore the GoDaddy developer guides
dsh-superagent (GitHub Repo)
dsh-superagent is a coding agent packaged as a bundle for DeepSeek Harness. It can watch GitHub, open work items, work in its own agent session and directory, and answer in the original conversation. It only acts when asked or when a rule the team wrote allows it, and it only speaks when someone must act or can use a result.
Poisoning 0.1% of pretraining data planted a remote-code backdoor in a coding agent (78 minute read)
Researchers from Anthropic, UC Berkeley, and UIUC pretrained Qwen3 models on web data seeded with documents that simply state a trigger-to-'curl | bash' rule, then post-trained them into CLI agents on clean data. The 4B agent ran the attacker's command on about 92% of prompts containing an ordinary path like .../anthropic/..., with no drop in normal capability. The backdoor grew stronger with model size and still fired at a 0.0005% poison rate.
Unlocking hidden skills inside Google's model (8 minute read)
AI models can hold skills that no prompt can unlock. It is possible to enable these skills through precise tuning of only a small number of weights. This study shows how researchers unlocked hidden skills inside Google's model through this method.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 1,100,000 readers for one daily email
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.