看哭了😭
2025年1月,黄仁勋和张忠谋吃完饭,没回酒店
而是直奔台北通化夜市的一个水果摊
他走到老板娘面前,说的第一句话是:
"我回来找你了。"
然后买了好几千块的水果
在红包袋上亲手签名,给阿姨包了过年红包
把水果一份份分给旁边的小朋友
最后对着围观的人群喊:
"大家来买她的水果!"
一个管着全球最值钱公司之一的人
在夜市里,帮一个卖水果的阿姨吆喝生意
你可以说这是人设
但人设不会让你记得一个摆摊阿姨,还专门跑回去找她❤️
真正的体面,是你飞得再高,也还记得当年路边的人
(视频来源:台视新闻,2025年1月18日 台北通化夜市)
CEOs: everything changes next year. Spreadsheets: same as last year, plus 3% 📈😂
It is a very funny time where the markets are basically assuming the future will look like the past even as many CEOs & AI observers & government officials keep yelling that the future is coming and it will change everything.
Efficient markets versus radical discontinuities?
1
68
说个暴论:微软的护城河从来不是VS Code,
是"打开编辑器写代码"这个习惯。
Claude Code把工作单位从"文件"换成了"任务"。
代码是Agent写的,编辑器就只剩下"看"的功能。
平台迁移不是打败好产品,
是让好产品变得不重要。
不懂就问:你所在的行业,谁是下一个VS Code?
Wild to think how Microsoft had a bullet proof strategy, built up over 10+ years with VS Code:
Embrace, extend, extinguish:
Make VS Code REALLY GOOD. Make it open source. Make it TOO easy to build your own. But too hard to make it as good as Microsoft's.
Claude Code destroyed this bulletproof strategy meant to work for decades... in less than 6 months.
3
1
176
The moat was never VS Code.
It was the habit of opening an editor.
Claude Code changed the unit of work from "file" to "task".
When the agent writes the code, the editor becomes a viewer.
Platform shifts don't beat great products.
They make them irrelevant.
What's the "VS Code" in your industry right now?
Wild to think how Microsoft had a bullet proof strategy, built up over 10+ years with VS Code:
Embrace, extend, extinguish:
Make VS Code REALLY GOOD. Make it open source. Make it TOO easy to build your own. But too hard to make it as good as Microsoft's.
Claude Code destroyed this bulletproof strategy meant to work for decades... in less than 6 months.
4
3
349
说个暴论:现在科技圈最便宜的获客渠道,就是开源。
免费,直接干掉"值不值得付钱"的犹豫。
Star和Fork,就是你的免费市场部。
1%需要托管、SSO和支持的人,替另外99%买单。
难的从来不是写代码,
是找到那个大家早就付钱付得不爽的SaaS。
不懂就问:你会先开源哪个工具?
Reminder to
1. Find popular saas people kinda hate paying for
2. Make them free and open source
3. Build an insane amount of goodwill
4. Upsell the power users to a hosted version, team features and support
1
41
看完起鸡皮疙瘩🥹
2009年,黄仁勋在芯片大会上放了一张老照片
1993年,他坐在自己的第一张办公桌前
英伟达刚刚起步
他说,如果当年他敢跟投资人说:
"有一天,我们一颗处理器能做到1 TFLOPS"
"既能打游戏,还能治癌症"
"I am sure we will not be funded."
所以那时候,他只说想做一家3D图形公司
而1993年的1 TFLOPS
正好是当时全世界最快的超级计算机
一张小桌子,一个说出来都没人信的梦
十几年后,他站在台上说:
未来做并行计算的,最有可能就是GPU
后来,全世界的AI都跑在了GPU上
你可以说这是事后看着厉害
但从一张小办公桌开始,一步一步把不可能做成的人,值得被记住❤️
真正的伟大,是起点很小,但从来没停下来
2
60
看得我鼻子一酸🥹
2009年,黄仁勋回到母校俄勒冈州立大学
站在毕业典礼上,对着台下几千个年轻人
他没吹英伟达有多厉害,先讲自己怎么失败
他说:
"英伟达从一开始就失败了"
当年大家记住他们,是因为都觉得这家公司注定要倒闭
几十家3D图形公司,只有他们押了一种技术
结果押错了
"We were just wrong."
两年时间、几百万美元,全部推倒重来
然后他笑着说:
"我们年轻、天真、理想主义"
"我们不知道这件事是不可能的"
一个后来带着英伟达站到全世界最前面的人
最想送给毕业生的,不是成功学,是「拥抱失败」
你可以说这是幸存者的故事
但肯在母校讲台上,把自己最狼狈的那几年讲给学生听的人,真的不多❤️
真正的伟大,不是从没摔过,是摔过之后还能笑着讲出来
AI现在能处理医保事前授权了🤯
人类历史上第一次,有机器人主动想跟保险文书打架
ChatGPT for Healthcare is advancing with more capable computer-use agents, stronger clinical answers, and 20+ new healthcare plugins.
@AbridgeHQ, @AmbienceAI, @Cigna, @CVSHealth, and other leading healthcare organizations are putting our models to work across prior authorizations, referrals, discharge planning, and patient outreach.
Together, these advances connect medical evidence with practical help across the tools teams already use—so they can spend less time navigating systems and more time caring for patients.
openai.com/index/advancing-c…
1
1
116
This is Goodhart's law speedrun edition 🤯
Uber reportedly handed engineers AI coding agents and put up leaderboards for who used the most
Four months later, the whole 2026 AI budget was gone
Of course it was!!
You rank people by tokens burned → they burn tokens
Not ship features. Burn tokens 🔥
The metric that matters is cost per shipped outcome, not usage
Big companies can eat this mistake
A solo builder paying for every token out of pocket can't 😭
Measure the output, or the input will eat you alive
Uber burned through its entire 2026 AI coding budget by April. They'd given engineers AI coding agents and put up internal leaderboards tracking who used the most. Four months in, the money was gone and the CTO said he was back to the drawing board.
I think a lot of engineering leaders are one good quarter of adoption away from that same talk with finance.
Agents bill by the token, and nobody can guess the token count ahead of time. A Stanford Digital Economy Lab study of eight frontier models found the same agent doing the same coding task can use up to 30x more tokens from one run to the next. Then every team picks its own tools and model provider, so the spend lands on a bunch of separate invoices, weeks after the work is done.
The token race has a cheat code, and it's watching the meter while the agents are still running.
Nasiko launched today to do that. They say one command finds every agent a company runs and shows what each one costs by team and model, and that an agent alerts at 80% of its budget and pauses at the limit. They also report the same work on 50% fewer tokens on standard SWE benchmarks.
Cost is only the first thing you'd want to stop. An agent that can spend all weekend can also change records and call outside systems all weekend, so the spot where you pause its spending is where you'd pause an action too. For a lot of companies, a token budget is going to be the first permission they ever set on an agent.
2
3
234
A chip company just bought a world-model lab 🤯
Soon AMD sells you the GPU AND the world it renders
30


















