Pinned Tweet
来看看你的1024关键词
请安装 github.com/dusthappy/1024 中的 programmer-day-1024 技能到你当前支持的技能目录，确认可用后，为我生成一份带八维雷达图、夸赞关键词和专属口号的 1024 程序员节礼物。
Emil 的检查清单：不同屏宽、触摸模拟、3G 限速，每次修复留前后截图。
我会再加一条：复测保留同一组条件。截图能说明布局变化，却不能证明加载变快；慢网下还应记录关键内容何时可用，别把“看起来好了”当成问题已解决。
Good prompt to polish your app, enjoy.
Test this in the browser at different widths, with touch emulation, and on a throttled 3G connection. Fix anything that looks broken or feels off. Save a before and after screenshot for every fix and send me a summary afterwards.
7
批量出答案很快，批量纠错可不快。
一次最多处理 1000 个问题，接进学习产品前我会先留抽查和纠错入口。错题混进去，最后还得一题题捞。
introducing the Gemini Decisions API
powered by Gemini Decision 1, our new multimodal decision model
get fast, calibrated decisions on up to 1000 questions at once across text, images, audio, video, and PDFs
try it today in AI Studio and via the Gemini API ai.studio
4
GPU 再快，启动开销照样吃速度。
Muse Spark Open 在 8× GB300 上，launch 每步 1,450→97，速度 517→1,109 tokens/s。
融合也有代价：大 kernel 取指令可能卡住。
Muse Spark Open: 517 → 1,109 tokens/s on 8× GB300 with SGLang megakernels.
One kernel launch per layer, overlapping compute, memory loads and communication. Total launches per decode step drop from 1,450 to 97.
lmsys.org/blog/2026-11-05-mu…
In our new technical post, Pranjal Shankhdhar explains what makes this work:
- Prefetching weights into shared memory and L2 before they’re needed.
- Warming the instruction cache so large kernels don’t stall fetching their own code.
- Using data itself as a readiness signal for cross-GPU communication.
- Fusing the draft model to make drafting 2.6× faster.
The post also includes measured timelines, CUDA/PTX examples, and the tradeoffs behind stopping at one megakernel per layer.
2
2
121
转写最贵的可能是后期认人名。
这次把会议日程里的讲者名和术语提前喂给模型，给主舞台实时转写做上下文偏置。做这类工具，我会先留一个词表入口；不然每场换人换术语，纠错得一直跟着跑。
What does real-time transcription look like in a live setting?
🎙 We used Voxtral Transcribe 3 at AI Engineer Paris to capture every main-stage talk in real time. We boosted accuracy with context biasing, feeding the model speaker names and technical terms from the conference agenda..
1
1
145
财务建议写得顺，不代表账就对了。
桌面版 Finances 能帮忙理解消费、做自定义财务工具。我会先核对输入和计算结果；省下整理时间，别把对账也省了。
目前仅限美国 Free、Go、Plus、Pro 用户，需更新桌面应用。
🤑🤑🤑 Cha-ching 🤑🤑🤑
Finances are now in the desktop app.
Find ways to save, understand your spending, and make sense of your investments. You can also build custom financial tools.
Update the desktop app to get started. Finances is available to Free, Go, Plus and Pro users in the U.S.
1
1
54
突发：
英伟达董事会批准黄仁勋辞任CEO，印度裔高管Jay Puri将接任
加利福尼亚州圣克拉拉，11月12日电 ——据知情人士透露，英伟达（NASDAQ：NVDA）董事会已于当地时间周三早些时候召开特别会议，正式批准公司联合创始人黄仁勋辞去首席执行官职务的申请，并一致投票决定，由印度裔高级管理人员Jay Puri接任公司首席执行官，该任命将于董事会完成相关交接程序后正式生效。
据悉，黄仁勋在内部会议上表示，随着英伟达逐步从传统半导体制造商转型为全球人工智能基础设施的核心供应商，公司需要一位能够适应下一阶段战略发展的领导者。Jay Puri随后发表简短声明，对黄仁勋的决定表示充分尊重，并感谢董事会给予的信任。他强调，英伟达现有的人工智能发展战略不会发生根本性改变，公司将继续致力于加速计算、数据中心及生成式人工智能领域的长期投资。
Dante 提到 Rider 2026.3 的 Godot 调试更新。官方说明里，我更在意的是断点处能看活动场景、节点层级和属性，而不只是一串变量。排查“对象存在但行为不对”，场景关系也是上下文。要区分两件事：场景树检查随调试会话提供；与 Godot 编辑器联动还需安装并启用集成插件。未亲测，不把功能清单当成排错效果。
Nueva versión de Rider (2026.3) y viene con novedades interesantes para gamedev: soporte para Lua y Verse, debugging con CoreCLR en Unity, así como mejor debugging en Godot C# y GDScript.
jetbrains.com/rider/whatsnew…
26
Shreeshiv 用改预约解释 Agent Experience：别让 agent 先取消旧预约，再赌新时段还在。我认为关键是把“改期”做成业务动作，而非暴露两次接口调用。代价是后端要承担一致性与失败恢复；多给模型一段提示词，不能替代这个责任。
1/9 Wtf is Agent Experience?
We've spent years arguing about where a button goes and how to get a task down from 5 clicks to 2.
All assuming a human is using the product.
Now an agent is doing the work. What makes your product easy for it to use?
21
演示越顺，越容易把试错成本藏起来。
年初大家盯着 AI 怎么做，现在回头问留下了什么软件。我更想先看别人能不能独立用起来：安装、出错、改需求时还顺不顺，决定这套技巧值不值得照搬。
remember that wave of ai psychosis from the beginning of the year?
what lasting software came out of that? the entire discussion was everyone marveling at the process
you gotta learn the lesson that techniques aren't worth discussing until you see impressive output
34
论文找得快，引用错了照样要还债。
3,000,000+ 篇论文能让 agent 找到并读全文。做学习产品时，我会把核对原文和引用留在交付流程里，不能只看摘要就交差。
Firecrawl now indexes 3,000,000+ research papers, including every major AI/ML paper.
Your agents can find the right paper, even one posted today, and read its full text.
Run your next lit review from any harness with the Firecrawl plugin.
alexandria.firecrawl.dev/go/…
3
1
87
开源很热闹，默认选型却可能把你锁进接口账单。
T3 Code 流量里，开源权重模型不到 0.5%。
独立开发做学习产品，选托管模型省部署，代价是每次调用都要考虑成本和供应商依赖。
3 months ago, this post would have gotten me dozens of death threats, hundreds of replies telling me I’m stupid, and at least 3 qts that ratio me.
This time? Crickets.
Fun fact: open weight models make up less than 0.5% of T3 Code traffic.
3
3
217
Xcode 签名报错，先别急着删证书重来。Arturo 这篇排查指南把问题拆成了签名身份、配置文件授权和最终产物中的权限声明。我觉得最有用的是检查顺序：先确认实际用了什么，再比较允许什么。尤其 App 加了 Widget，主应用能签名不代表扩展也配对了。自动签名省的是维护工作，不是这些对应关系；本机能过，也不能替 CI 的环境作证。
Una lectura para el finde, con café y sin prisa. ☕️ ¿Por dónde empiezas cuando Xcode dice que el perfil no incluye el certificado?
Primero, comprueba qué identidad está usando Xcode. Un .cer contiene la clave pública, y para firmar necesitas también la privada. security find-identity -v -p codesigning te ayuda a ver qué identidades están disponibles en el llavero.
Después, mira qué autoriza el perfil. Si has añadido un App Group para compartir datos con un widget, comprueba los dos targets. Cada ejecutable se firma por separado. Con codesign puedes inspeccionar los entitlements que acabaron en cada firma y con security cms puedes leer el perfil incrustado y comparar los valores.
En la Swift Pill de hoy repaso también qué cambia entre Desarrollo, Ad Hoc y App Store Connect, y por qué un build puede funcionar en tu Mac y fallar en CI con el mismo proyecto.
Dejo el enlace en el hilo.
#iOSDev #Xcode #CodeSigning
2
3
88
代码写得更快，维护账单可不会自动消失。
能干活的编程 agent 出现还不到一年，职业改造已经在进行。独立开发省下的时间，得留给验收和维护。
代码总要有人负责。
The counterexample to AI diffusion taking a long time is coding, where the complete transformation of a profession is well underway less than a year after real working coding agents. Faster than I would’ve expected given historical technological adoption rates in professions.
1
2
136
作业一键答对，学习产品可能就做废了。
对话式辅导可以规模化，但如果只优化给答案，学生省下的正是练习思考的过程。做学习产品，我更看重学生最后能不能自己推出来。
Conversational AI is going to transform education and deliver personalized tutoring at scale. But we have to build it right.
Most current AI tools in education optimize for the wrong thing: getting the answer. But the point of homework is not the output, it's learning to think by going through the process. We should be building AI that helps students develop their own thinking rather than outsourcing it to AI. Gakku is actually doing that.
Huge congrats to Annie on the raise!
1
1
60