呆崽说 AI

每天盯全球 AI：新模型、新工具、实测、成本账。说个暴论，事实和判断分开写。

[呆崽说 AI @shanke9527]

OpenRouter 被 Stripe 收了，据《纽约时报》报道交易价约 75 亿美元。说实话，模型聚合这门生意最后还是拼支付和企业渠道，OpenRouter 选了个最有钱的靠山。https://t.co/ciPyAmGpXK

[呆崽说 AI @shanke9527]

客户一句话：演员换了，车换了，再加条狗，把那个人删了。以前得重拍，现在 Luma 一个镜头全改了。

Character Swap & Insert 刚上线，换演员换车加狗删人，4 种改法不用动原镜头。说实话，拍广告的改稿地狱终于有人管了，做视频的都懂。

你会拿它先改什么？

https://t.co/TjKmzJthVZ

[Luma @LumaLabsAI]

The client has notes. Swap the cast, swap the car, add a dog, take someone out. Same shot.

Character Swap & Insert is now in Luma.

[呆崽说 AI @shanke9527]

DeepMind 发布 AlphaGenome 2，预测 DNA 变异在不同细胞里的效应：一次读入 1 Mb DNA 加一个目标细胞状态。TenK10K 上因果变异召回率 16%→30%，方向判断 23%→43%，API 免费向非商业研究开放。我觉得 AI for Science 就该这样：直接给工具而不是只发论文，你会用吗？
https://t.co/0oNnIxseeD

[呆崽说 AI @shanke9527]

开车也能跟 Gemini 唠嗑了。Google 刚把 Gemini 搬上 Apple CarPlay，今天起向所有 iOS 用户全球推送，点一下图标就能接续手机上的对话，Gemini Live 全程免提。Google 说多数人每天通勤平均约 1 小时。说实话，这 1 小时终于不用只听歌了，你会在车上跟 AI 聊天吗？
https://t.co/AdVxuHMgl4

[呆崽说 AI @shanke9527]

开源模型GLM 5.3自己动手，在离线沙箱里给6个最新版Firefox全写出了能用的漏洞利用，每个成本约1000美元。Firefox有2亿多用户，补丁还没推到用户手里，攻击就先造出来了。我觉得补丁窗口期这事以后得重新算：以前挖洞要专家，现在1000美元就能批量开洞。
https://t.co/agZwhuZYtK

[呆崽说 AI @shanke9527]

东南亚那批租英伟达卡的中介，快被债压死了。据报道，Megaspeed 20个月买了约48亿美元的卡，4月68.4兆瓦电力被收回。Litong负债率89.82%，一季度利息3500万人民币，比净利润3026万还高。说实话，三年租约、六年折旧，第四年旧卡没人要，前三年利润一把抹掉。
https://t.co/1cYSmay6Pl

[X.PIN @thexpin]

https://t.co/K7IEcN3EZu

[呆崽说 AI @shanke9527]

Google 用手表数据算出一个“能改的健康年龄”：SensorWM 世界模型拿 228 万人周的可穿戴数据，模拟你未来 12 周换个作息身体会怎样。盲测里，教练 71.4% 的时候选了它给的方案。我觉得这才是手环该干的活——不是报你昨天走了几步，而是告诉你改了会怎样。你信吗？
https://t.co/9MxEjp65Bg

[呆崽说 AI @shanke9527]

LangChain 让自家销售彻底不用手写材料了：GTM Agent 12 周里生成了 1540 份销售物料，省了 337 小时。有物料的交易 47% 推进到下一阶段，没有的只有 16%。我觉得这才是 Agent 该有的样子——不是聊天框，是直接把活干完还能量化。你会让 Agent 碰你们的客户材料吗？
https://t.co/8yRBl0VlSv

[呆崽说 AI @shanke9527]

Meta 管基建的老将要走了：Santosh Janardhan 2009 年加入，一路做到基建负责人兼工程联席主管，现在交棒给在亚马逊干了 19 年的 Dave Brown，明年正式离开。我觉得时间点很微妙——Meta 正砸钱扩数据中心，换帅等于换打法。接下来看 Brown 怎么接这摊子了。
https://t.co/CKygOvddZK

[呆崽说 AI @shanke9527]

10 月美国企业 AI 支出又创新高，但增速只剩 1.65%，是放缓了。分化很夸张：头部 1% 的企业人均每月花 10200 美元，普通企业才 16.32 美元。说实话，AI 花钱这事现在就是头部游戏，1% 的企业贡献了 OpenAI 和 Anthropic 80% 的收入。
https://t.co/KQf5qyO9Zq

[呆崽说 AI @shanke9527]

Meta 的 Muse Ember (Max) 刚登上 Arena Agent 榜高效模型第一，在 6000 场真实智能体会话里净改进 +1.7%，每任务成本才 0.02 美元，大概是竞品的四分之一。我觉得高效模型这条赛道要变天了，便宜又能打的模型一多，Agent 落地成本直接砍一截。
https://t.co/8rB5xh4Jtf

[Arena.ai @arena]

Muse Ember (Max) by @AIatMeta lands in Agent Arena as the most capable and efficient : +1.7% net improvement at about $0.02 per task, roughly 4x less than GPT-6 Luna (Max) and Claude Haiku 5.5, across 6k real-world agentic sessions from our global community.

How it stacks up against other efficient models
- #25 Muse Ember (Max): +1.7% | $0.02/task | 0.5 min/task | 11.6K output tokens
- #26 GPT-6 Luna (Max): +1.6% | $0.09/task | 6 min/task | 41.2K output tokens
- #29 Claude Haiku 5.5 (High): +0.5% | $0.08/task | 1.8 min/task | 24.3K output tokens

Performance is on par across the three, but Ember's median task costs roughly 4x less, finishes about 12x faster than Luna (Max) and 3.5x faster than Haiku 5.5, and uses 72% and 52% fewer output tokens.

Congrats to the @AIatMeta team, @alexandr_wang and @finkd on this impressive release!

[呆崽说 AI @shanke9527]

Google 新出了个专门做判断的模型 Gemini Decision 1，一次能判 1000 道是非题、多选题，150-250 毫秒就出结果，还带置信度。每百万输入 token 0.1 美元，输出 token 0 元。说实话，这就是冲着替代自训分类器来的，以后路由、审核、打分这种活，可能都不用再训小模型了。
https://t.co/rsDFnwpGBZ

[呆崽说 AI @shanke9527]

CoreWeave 三季度营收 36.6 亿美元，同比涨了 168%，但净亏 7.09 亿美元，一个季度利息就 8.46 亿美元。现在它改主意了：不靠发债，靠客户预付款撑扩张，未来三个月预计能收约 90 亿美元预付款。我觉得这模式就是在赌，订单不断就没事，断了就很悬。
https://t.co/XCEH7RundA

[呆崽说 AI @shanke9527]

英国芯片初创 Fractile 刚融了 7.2 亿美元，估值超过 50 亿英镑，折合约 70 亿美元，英国主权 AI 基金也投了。领投的是 Thrive、Founders Fund、Lightspeed。说实话，欧洲终于有人想在 AI 芯片上跟英伟达掰手腕了，就是不知道这 7.2 亿够烧多久。
https://t.co/XttOhogv5O

[呆崽说 AI @shanke9527]

AI 编程没让人少干活，反而越干越晚。Marc Andreessen 在播客上说，硅谷出了个词叫“AI 吸血鬼”：效率太高，写代码根本停不下来。顶级 AI 程序员年薪已开到 5000 万美元。我觉得这行正在分化，会用 AI 的越卷越狠。你最近是变轻松了还是更晚了？
https://t.co/7tym19wFiQ

[Rohan Paul @rohanpaul_ai]

Marc Andreessen (the guy who called it when “software ate the world”) pointed out what AI is doing to coding.

"Everyone assumes AI coding means fewer hours… or leaving the profession. But almost everyone I know is working more hours."

"There’s a new term in the Valley: the ‘AI vampire.’ You’re up all night AI coding because you’re so productive you can’t shut it off."

"The top AI coders make $50M a year. They’ve found the philosopher’s stone."

"Every company has a thousand projects they’ve wanted to build but never had the bandwidth. Now they can. This isn’t a blip, it’s going to intensify."

except the real time sink flipped

developers aren't wasting hours writing syntax anymore

they're wasting hours manually switching models, babysitting reasoning depth, and burning token budgets across 4 tabs

the engineers actually compounding don't manually choose models on every turn

they let a decision model route each prompt to the right engine automatically

stop babysitting models at 3am

---

(Full video from PowerfulJRE" (Joe Rogan) youtube channel (link in comment)

[呆崽说 AI @shanke9527]

AWS 刚开源了 AWS Emulator：给编码智能体用的本地版 AWS，能离线跑测试。
覆盖 200 多项服务、400 多个 API 模型，代码 CC0 开源，而且全是智能体自己写、自己测的。
说实话，写代码的是智能体，测代码的是智能体，现在连测试环境都是智能体写的，这闭环快得吓人。
https://t.co/xLHTmkFnRA

[呆崽说 AI @shanke9527]

AWS CEO 泼冷水：太空数据中心，至少还得等 10 到 20 年。
Matt Garman 的理由很实在：一个 GPU 机架重达一吨，发射又贵又慢，太空散热难、辐射遍布，建和维护都比地面贵得多。
我觉得这等于给炒上天的太空算力判了缓刑，技术上都能解决，但钱包和时间都等不起。
https://t.co/C3d9zugqAD

[Rohan Paul @rohanpaul_ai]

AWS CEO Matt Garman: Data centers in space are at least 10 to 20 years away.

"I think that it is highly unlikely to be economically feasible or practical in the next decade or two.

It sounds good. You get free energy and it’s up there somewhere. Launches are still relatively infrequent and expensive.
One person is not going to fix it. We as a world don’t do one launch a day today.

We’re landing in a data center today one GPU every 14 seconds. So we don’t do a launch a day for the world, and that’s just us doing one GPU.

A rack of GPUs might weigh a ton. I don’t even know if you can stick that on a launch vehicle or how many you can get up there.

People talk about gigawatt data centers and megawatts of racks. Most solar panels drive kilowatts, not gigawatts. So you need these gigantic things that are up there.

They’re really hot. Everyone’s like, space is cold. Space also doesn’t have air. Dissipating heat in space is actually quite hard.

It’s just a really hard environment. There’s radiation everywhere. It’s going to be way more expensive to build servers, from a serviceability point of view.

These are all technically solvable problems, and it’s just going to take a while. I’m not saying any one of these things is impossible. I just think that it’s way far out."

----
Full video on "Sources Podcast and and Alex Heath (@alexeheath ) " YouTube channel, (link in comment)

[呆崽说 AI @shanke9527]

有人把 Windows 11 塞进iPhone 里本地跑起来了，不是串流，是完整桌面 ARM 系统。
Ammaar Reshi 用 Gemini 4 Carbon 做的 iPhone Duo，展开就是口袋 PC，还能跑 Antigravity 写代码。
说实话，手机变电脑喊了多少年，这次看着真能用了，就看续航顶多久。
https://t.co/7PDiJgBUu3

[Ammaar Reshi @ammaar]

I used Gemini 4 Carbon to get Windows 11 running on my iPhone Duo! 💻

Unfold the screen and it's a pocket PC, including Antigravity running coding agents on my phone.

This is the actual desktop ARM OS running locally with touch controls.

Demo + repo below!

[呆崽说 AI @shanke9527]

UCSD 搞了个能给 LLM 基建做形式化验证的工具 vstack，顺手在 PD 分离推理服务里揪出 3 类 RDMA 和 KV cache 的 bug，还提了 6 项上游修复。
说实话，AI 基建以前全靠跑测试蒙，现在开始上数学证明了，这路子一旦跑通，出事的成本会低很多。
https://t.co/E6p1tQwe6c

[Hao AI Lab @haoailab]

Towards Verifiable LLM Infra 🚀

Suppose coding agents can build a fast inference engine from scratch, and one engineer orchestrates 10,000 agents in parallel: how do we know the engine they wrote is actually correct? Having agents review other agents mostly wastes tokens compared to empirical and formal verification.

We introduce vstack, a 5-stage formal verification workflow for coding agents, and veriserve, a verifiable PD-disaggregated serving engine: uncovering 3 classes of real RDMA and cache races in production serving engines, and backing veriserve's KV-page lifecycle design model with a 226-module, 516,973-obligation TLAPS safety proof at +0.18% mean p50 TPOT overhead on GB200.

Blog: https://t.co/SdbINvSVht
veriserve: https://t.co/ugHLToReCN
vstack: https://t.co/GZJebbjdpL

[呆崽说 AI @shanke9527]

AI 最赚钱的环节居然不是卖模型：Tomasz Tunguz 扒了 CoreWeave、Anthropic 的公开文件算了一笔账，应用层和 Neocloud 加起来拿走 68% 的毛利。
我觉得这账算得挺扎心的——大家都在卷模型，钱却被离用户最近的那层和租算力的那层赚走了。
https://t.co/0k8pOfzvCU