
🇺🇸 🌏 U.S. Think tank studying #5GW (5th gen warfare) & a more interesting reality
Joined April 2021
- Tweets5.9K
- Following517
- Followers8.7K
- Likes1.6K
>*IT comes in the office angry* "We're looking for a DeepSeek!"
>Us: "that the hell is a deepseek?"
>*IT shows photo of a DGX spark*
>Us: you mean a spark?
>IT: What's a spark?
5
428
海外硬件媒体 Tom's Hardware 与硬件监测机构 UNIKO's Hardware 连续曝光了中国显卡改卡工厂的流水线实况:车间里正在把消费级 RTX 5090(含公版 Founders Edition 与非公旗舰)成批拆下 GB202 GPU 核心和 GDDR7 显存颗粒、重植球,改造成能塞进高密度 AI 机架的单槽/双槽涡轮散热计算卡(Blower Card)。
车间负责人在直播中边吹焊枪边拆穿了市面上流传的 RTX 5090 48GB / 96GB 大显存魔改卡 骗局:因为英伟达把 Blackwell 架构的签名验证 BIOS 锁得死死的,没有底层解锁固件,哪怕在定制 PCB 上硬焊双面 GDDR7 显存颗粒也根本点不亮,目前号称现货量产大显存版的全是骗定金,工厂只接客户自带原装 32GB 核心与显存的来料加工。
最绝的是直播间的弹幕问答。网友问为什么他们改出来的 16-Pin(12V-2x6)供电口从来不烧,老板说因为放弃了全自动 SMT 贴片,全靠老师傅手工插件加锡焊死。
海外观众问:黄仁勋允许你们这么拆 5090 改涡轮卡吗?老板头都不抬:不归他管。
再问:那老黄要是来找工作你们招吗?老板:招啊,门口保安队长挺适合他的。
We tweeted this on Friday. Nobody deploys a software update on a Friday.
Listen, we still don’t have the complete picture on this FAA ground stop, but who in their right mind runs updates on key ATC systems during peak times when loads on ATC are highest? If that is actually what happened, or some form of it, whoever approved that shouldn’t be in their position. It’s a life safety thing above all else. Sorry you have to work a night or a weekend. Join the club,
317
DM32s are awesome go-to field radios for non sensitive work. The Mayhem firmware looks like a lot of fun
Instead of banning Chinese printers, we should ban European printers. This thing has may be five hours on it @Prusa3D. We need a US Bambu
The core one is trash pile compared to @BambulabGlobal (even considering the Bambu network plugin is spyware)
107
Agentic first radio networks are seriously the future.
Although, as we like to tell the Department of War, if your agentic network cannot order you a pizza from a military base in Indiana fully autonomously, you’re not going to win against the Chinese.
Qwen is anticipated to drop a new open weight model that will outperform everything else using their new engram architecture. We got a taste for this in 3.8 flash next, and because of cost, the Qwen 4 architecture will result in Chinese dominance in AI in developing regions
IT/Legal at work: "You're not using Deepseek 4.1 1017 at 2,500 tok/sec are you? You are sheltering the weights."
In reality, running Chinese inference FAST provides a major strategic advantage that needs to be studied, tested and copied by the U.S. and its allies.
So ban Chinese models? Absolutely NOT. These are the greatest support for American industry (just like cheap solar). We know of many US aerospace companies running Chinese models in containers to replace Claude / Open AI for work and it can be done safely.























