美团发布了 LongCat-2.5模型
是我们吃的美团？🤨
Introducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active.
Built for long-horizon agentic tasks, it brings stronger visual understanding and agentic coding to work ranging from complex coding projects to visual coding such as creative design and 3D tasks.
Learn more about LongCat-2.5: longcat.ai/blog/longcat-2.5/
Fable5.5的动画水平🤯
这就是Anthropic不做视频模型的原因吗？
Fable 5.5 its over for video and image models 🤯
THIS IS PURE CODE
one shotted this video holy shit , how this is more beautiful than ai video models
Two ways to check this yourself:
Ask: “Where did @karpathy go in 2026?” with web search and memory turned off. Only the new Fable knows the answer. Then give it the prompt.
Run the same prompt I used for the test (it’s in the screenshot) and see for yourself.
Don’t rely on Grok for basic reasoning like this. Grok doesn’t have up-to-date knowledge of niche stuff. Either join the community or run the test yourself.
Also, a knowledge cutoff doesn’t mean a model knows everything before that date. Sometimes a model doesn’t know about an event even if it happened before its cutoff. This test shows that too.
Run the same prompt on regular Opus 5.5 and the quality drops. On Fable 5.1 it’s worse.
🤯核打击都来了💣
Under pressure from Israeli prime minister Benjamin Netanyahu, the Trump administration may be within days of launching a nuclear strike on Iran. The target would be Pickax Mountain in central Iran, where Israeli intelligence claims the Iranian government has restarted its nuclear enrichment program. No conventional weapon in the American arsenal can penetrate 300 feet of solid granite to destroy the tunnels where enrichment is supposedly taking place. The only military option is a nuclear strike. Netanyahu has been pushing hard for one. “Bibi calls Trump about it every day,” says an informed observer.
Netanyahu is cornered and desperate. After last week’s election results, he can’t form a government unless Trump intervenes directly in Israeli politics and brokers a deal with other parties that gives Netanyahu 61 seats in the Knesset. Otherwise, he is out as prime minister and on his way to ignominy, investigations and possibly jail. Netanyahu must stay close to Trump. He needs this war.
To that end, he’s applied maximum pressure to his closest allies in the administration, including Secretary of Defense Pete Hegseth and CENTCOM commander Brad Cooper. Both oppose a negotiated settlement to the war they’ve so grievously mismanaged; both are pushing to restart major combat operations. Secretary of State Marco Rubio and CIA director John Ratcliffe are also following Netanyahu’s lead, as are leaders of the Democratic Party. In a recent meeting on the subject, former president Barack Obama said he opposes the current Iran peace deal being drafted in Doha.
The rest of America’s longtime allies in the Middle East are terrified by the prospect of more war. On September 27th, Netanyahu flew secretly to Abu Dhabi to enlist the UAE in his lobbying effort for a final attack on Iran. The Emiratis refused. After eight months of conflict, Gulf economies are teetering. So are global energy and fertilizer markets. They won’t recover soon. Globally, there are a total of about 850 supertankers used to transport oil from producers to the rest of the world. In October alone, 11 of them were badly damaged in Hormuz. At this pace, a global depression is imminent. A nuclear strike would radically accelerate the destruction. Fallout from a ground explosion at Pickax Mountain would contaminate the Persian Gulf, poisoning the water supply for 100 million people.
And those are just some of the immediate consequences for the region. The rest of the world is every bit as exposed. No country other than the United States has ever used a nuclear weapon in war. Some would like to. An American nuclear attack on Iran would set a new precedent. The Russians have already told at least two other governments that an American nuclear strike on Iran would free them to use nuclear weapons against Ukraine. Pakistan and India would be tempted to do the same against each other, as would North Korea against Seoul. After the first nuclear strike, World War Three is inevitable. That war would be short and final.
We’ve been sleepwalking toward this moment for months. Since February, the administration has quietly rewritten America’s nuclear doctrine to make first use of so-called “battlefield nukes” easier and therefore more likely. No one in Washington in either party seems bothered by this. Now we’re on the brink of it.
Close observers say Trump is likely to decide by next weekend whether or not to use nuclear weapons. That would be November 7, six days from now. To his credit, he’s hesitating. At the moment, Trump is listening to his vice president, JD Vance, and to his envoy and friend Steve Witkoff, both of whom argue a deal with Iran can be made.
They’re right, though it won’t be simple. The Iranians are famously frustrating to negotiate with under any circumstances, and elements of that government have become more radical since Ayatollah Ali Khamenei and his family were killed in February. After decades of isolation, even relatively moderate Iranian leaders seem confused by how decisions are made in the West. (There is a widespread belief in Tehran, for example, that the United States government controls Israel.) But overall, people who have spoken to decision-makers in Iran say there is consensus there that more war is bad for everyone and a deal is worth making.
One group that hasn’t been consulted so far is the global population — the hundreds of millions of people who will be killed along with their children if Trump is pushed by Netanyahu into launching a nuclear strike next week. Now would be the time for them to speak up.
🚀 DeepSeek Harness 团队提出 Super-agent：让 AI 变成 24 小时在线、能同时干好几件事的「同事」
🤯 痛点：模型越来越能独立干活，瓶颈却变成了人的注意力——人通常只能同时盯 3-5 个会话，多任务全靠手动来回切换
🛠️ 实战：内部实例已在跑，群里没人 @ 它，它也能自己发现 bug、定位根因并提 PR
✨ 好同事的四个标准
🔗 连续：群聊、GitHub 等渠道里始终是同一个 Agent
⚡高效：多任务并行，上条没做完也能接新指令
📡 主动：监听环境事件，该出手才出手，其余保持安静
🛡️ 持久：永远在线，直连真实工作环境，异常自动恢复
🏗️ 单个 ReAct 循环做不到这四点，要靠五大模块：
1️⃣双层架构「总管」（Router）+「员工」（Worker）：
「总管」只负责对接人、分派活，自己不动手，用轻量模型就够快
「员工」 各自隔离上下文干活（通常用最强模型），只向 Router 汇报
2️⃣ 事件驱动：群消息、邮件、CI 等环境事件即 Trigger——代码规则先筛噪音，LLM 再做语义判断
同一 PR/话题的后续事件订阅直通对应 Worker；还会后台「推测执行」：先复现修复，确认可行才出手
3️⃣全双工多渠道：异步 start()/stop()，长任务不阻塞对话（已实现，默认未开启）
群聊 每条限 100 词，GitHub 等则输出结构化的详细内容
4️⃣ 大脑在云端，手脚在你电脑
AI 的运行和记忆放在云端，所以你电脑休眠、关机它也不会停
但你的代码、文件仍留在自己电脑上，只装一个小程序当「手脚」，需要时远程操作
电脑重新上线，它就接着把该在这台电脑上做的活做完
5️⃣ 配一个「监工」Watchdog：独立会话定时巡检，不只看进程存活，还判断 Agent 是否空转、跑偏
📊 初步实验（模拟场景，仅供参考）
模拟约 300 条消息、其中约 20 个活儿，对比「一个窗口 + 子助手」的常规做法：
⏱️ 一半任务的完成时间：12.4 分钟 vs 13.8~20.4 分钟
95% 的任务：15.7 分钟内 vs 约 29 分钟内。后者慢，是因为 AI 常自己埋头干，被新消息不断打断
💸 代价：缓存读 token 更多（23.2M vs 6.1M-10.8M）
✂️ 去掉订阅直通：速度不变，Router 决策多约 30%，token 多约 37%
📈 总管的「批量处理」：消息从每分钟约 1 条涨到约 40 条，它会把积压的几条打包一起处理，每个决策平均耗时从 6.4 秒降到 1.6 秒
⚠️ 待解难题：Router 层怎么扩展、长期记忆、无会话的产品形态、跨实例 A2A，以及安全隐私（提示注入、凭据泄露）
💡 一句话总结
这不是新理论，而是把计算机领域的老经验搬给 AI：分快慢两条路、把有限的「记忆窗口」当缓存用、提前预判、大脑和手脚分离、设监工
人从盯执行细节，变成做决定、签字把关
