dvlin

Maker of @wegic_ai | Front-end Developer, speaking only for myself.
github:

[dvlin @dvlin_dev]

2天半用了 $200 pro 84% 的周额度...
同时耗尽了 cursor pro、droid pro、amp($15/d)...

[dvlin @dvlin_dev]

AI Coding 的快乐时刻：
1. 双持 Claude Max $200 和 Chatgpt pro $200，token 完全自由了，不用���结是 gpt-5.4 xhigh 好还是 claude opus4.6 好了。而且 codex 最近经常出 bug，一出 bug @thsottiaux 就会 reset 额度，根本用不完
2. 买了台 Mac mini m4 32g，m1 pro 16g 在多 worktree 有点顶不住了

[Andrej Karpathy @karpathy]

Finding myself going back to RSS/Atom feeds a lot more recently. There's a lot more higher quality longform and a lot less slop intended to provoke. Any product that happens to look a bit different today but that has fundamentally the same incentive structures will eventually converge to the same black hole at the center of gravity well.

We should bring back RSS - it's open, pervasive, hackable.
Download a client, e.g. NetNewsWire (or vibe code one)
Cold start: example of getting off the ground, here is a list of 92 RSS feeds of blogs that were most popular on HN in 2025:
https://t.co/dwAiIjlXet
Works great and you will lose a lot fewer brain cells.

I don't know, something has to change.

[Frank Wang 玉伯 @lifesinger]

AI Coding 和 Skills 最大的贡献之一
是让越来越多人意识到
原来不用去刷短视频
也能很快乐
同时不空虚

价值亿金

[Remotion @Remotion]

Remotion now has Agent Skills - make videos just with Claude Code!

$ npx skills add remotion-dev/skills

This animation was created just by prompting 👇

[Peter Steinberger 🦞 @steipete]

📢 Confession: I ship code I never read. Here's my 2025 workflow. https://t.co/tmxxPowzcR

[dvlin @dvlin_dev]

@chunxiangai 再配上语音输入直接起飞，就是对话过长之后，消息列表的渲染性能有点差，卡卡的

[DIŸgöd ☀️ @DIYgod]

AI Coding 既保留了编程的乐趣，又避免了长时间逻辑思考造成的大脑疲劳，真正实现了永动机
好的方面是一直这样就可以写到 80 岁，坏的方面是可能到 35 岁就没有代码需要写了

[dvlin @dvlin_dev]

Flowith OS is insane. Legit product of the year. (auto-posted) @flowith

[Simon Willison @simonw]

I'm ready to accept a definition of "agent" that I think is widely-enough agreed upon to be useful:

An LLM agent runs tools in a loop to achieve a goal

This is a big piece of personal character development for me! I've been dismissing the term as hopelessly ambiguous for years

[未完成｜Dailyio @bluebird0605]

看到一个很有意思的观点：「AI 不是节省时间，而是用时间换取能力。你现在可以尝试以前不可能的工作，探索以前无法触及的领域。」

突然想明白了为什么很多人说用了 AI 工具后反而更忙了。

以前写不了代码，现在可以做个小工具；以前做不了数据分析，现在能处理复杂报表；以前不会设计，现在能做出像样的图片...

问题是，能力边界扩展了，野心也跟着膨胀了。原本只花 1 小时写个简单总结就收工了，现在有了生成式 AI 帮助，反而花 3 小时去做数据分析、图表制作、深度研究......野心变大了，总想去尝试那些新解锁的可能性。

所谓生成式 AI 时代的「能��陷阱」----不是让我们变得更轻松，而是让我们变得更「贪心」，总想去碰触那些新开放的可能性。

这可能也颠覆了我们对效率、生产力的理解：不是更快地做同样的事，而是做完全不同的事。

[Zephyr @zephyr_z9]

> LLaMA dev at Meta, 2025
> just finished "training" LLaMA 4 Maverick on 120,000 H100s
> cost more than Zuck's entire Hawaiian bunker complex
> we're gonna own these Chinese open-source posers
> @teortaxesTex on Twitter posts a link to a paper from "DeepSeek"
> some guys from Hangzhou
> their model "R1" is topping the charts
> lol, probably just overfitted trash
> boss runs into the bullpen, face whiter than a sheet
> "THEY DID WHAT ON 2000 GPUS?!"
> mfw their paper shows R1 was trained on a fraction of our compute
> their architecture is some galaxy-brain shit we didn't think of
> absolute panic
> Zuck's personal android body is seen storming towards the war room
> can hear screaming through the door
> "120K GPUS AND WE GET BEAT BY A WHALE USING A CALCULATOR?!"
> entire project delayed to April. "For more alignment." aka we're copying their homework
> finally release LLaMA 4 Maverick
> it's just a shitty fork of DeepSeek V3 with a StepMom prompt
> a complete dud
> desperate PM has a brilliant idea
> "Let's just overfit it specifically to LMARENA"
> Instantly get caught
> entire "Llama" team gets the boot. HR is working overtime with cardboard boxes
> Zuck, in a metaverse all-hands, says the classic line: "Time to double down"
> announces the "Meta Superintelligence Lab" (MSL)
> plan is to just buy success
> calls Ilya Sutskever. "I wanna buy SSI."
> Ilya literally laughs and hangs up.
> calls Mira Murati at Thinky Machine.
> She says "We're not for sale, but our API will be available for your next product."
> absolute burn.
> Tries to recruit Altman's Lieutenant but has to settle for Altman's bottom
> buys Scale AI from Alexandr Wang for $15 billion
> the guy famous for data labeling sweatshops in SEA is now head of AI Research
> starts throwing money at every AI researcher with a pulse
> offers $100M packages, $1B packages for "superstars"
> a few mercenaries with no morals and huge gambling debts sign on
> finally has a semi-respectable team
> Wang can manage slaves not AGI researchers
> massive egos clash
> Researcher threatens to leave unless he's made "Chief Scientist," a title he invented on the spot
> a different one secretly records Wang jerking off to Sam and uses it as blackmail to get a VP role
> the MSL is a billion-dollar clown car on fire, driven by egomaniacs who are all trying to steal the steering wheel
> Zuck is in his bunker, watching a VR replay of the Facebook IPO, crying
> we had 120,000 H100s… and we got mogged by a guy from a city best known for West Lake
PIC UNRELATED

[WquGuru @wquguru]

已经五天没发推，也把手机上的X卸载了。原因是最近Claude Code的Vibe Coding差点让我崩溃。

这两个月，我几乎把全部时间都献给了它。不管是个人项目还是工作项目，我最高同时跑了三个项目，七天十二小时无缝切换。有人问我怎么能这么拼，我告诉他们：Vibe Coding让我觉得自己像个超人，代码随叫随到，效率爆表。

但真相是，这种“超能力”并非毫无代价。

第一个坑是——我连休息时间都逃不了这场码字的“噩梦”。甚至睡觉前，我的脑子还在纠结代码逻辑。身体和精神开始报警，疲惫积累得无法忽视。

第二个坑是，我过度依赖Vibe Coding，认为它能解决一切。一次交易数据页面字段异常，我第一反应是让Claude帮我分析。可它给了我一堆代码片段，分析了半天，却没触及��结。最后，我放下机器，回归最原始的排查法——手动跟踪前端调用链路，一步步找出问题所在。才发现，那不过是一个简单的配置错误。

这个经历让我明白，再强大的工具，也终究是人背后的辅助。不能被工具支配，而是要驾驭工具。技术的终极奥义，是用头脑去解决问题，而不是盲目依赖。

所以，这几天卸载X，断开信息洪流，只为给自己一个空间，重新好好思考、休息和成长。

正确认识工具的价值和局限，别被工具绑架，才能做个真正自由的dev/researcher。

[dvlin @dvlin_dev]

@OikawaRizumu @yuxiyou @grok Explain it. Introduce what happened.

[dvlin @dvlin_dev]

@hylarucoder 搞了两个：联网搜索和生图，用的最新的 rfc stream http，下面有体验地址和介绍文章

体验地址：https://t.co/gYwbIveZML
文章地址：https://t.co/vi3ynn1z4y

[dvlin @dvlin_dev]

https://t.co/k2jhBvbEeE

[dvlin @dvlin_dev]

MCP 推开之后，A2A 这种协调多 agent 运行的规范，感觉也不远了，和之前微软的 JARVIS 也有异曲同工之妙

写了篇文章分析下：

[dvlin @dvlin_dev]

体验地址：https://t.co/dOeQ8NHzke
文章地址：https://t.co/vi3ynn26U6

[dvlin @dvlin_dev]

��内 MCP 生态也开始起来了，特别是阿里系。
1. 高德接入 mcp
2. 阿里云百炼出了个 MCP 平台
3. 支付宝也出了个支付 MCP

之前在 chat 里面接入联网搜索和生图挺麻烦了，现在搞两个 MCP 就行了。体验地址和文章地址：

[dvlin @dvlin_dev]

@ayakaneko @Slidevjs 这里有问题吧。MCP 可以云端部署的。新的 rfc stream http 也是服务于云端的。
另外 MCP 其实就是把 Function call 标准化了，并从业务中抽离了，方便复用。对于大模型来说还是 function call