Sen Liang

Deep Learning, Computer Vision

[Michael Black @Michael_J_Black]

A student asked me about my habit of bike riding or walking in the forest as a way of gaining inspiration. I thought I'd share my response:

I suspect everyone is different but all of our brains need time to process information. Some of this happens in sleep -- the idea that if you have a hard problem that you should "sleep on it" is very true. I often wake up with an inspiration.

You need to give your brain time to do its thing. You can't control inspiration. All you can do is create opportunities for it to occur.

I often have ideas in a dark conference hall while half-listening to a boring talk. I'll have spent the day at posters and learning about new work. Sitting in the dark and letting my mind drift is when those fragments come together into something new.

I will often do some work early in the morning and hit a snag. Then I go for a bike ride in the forest. I am not intentionally thinking about my problem -- I'm in the moment. My body is moving in the world. But part of my brain must be working on the problem because within 15-20 minutes, I have the solution.

The brain is amazing. Cramming it with information all the time may feel productive. But feeding it information is very different from letting it be creative. You can't force it to be creative and if you are force feeding it, it can't be creative. You need to leave it alone. Let it wander. Let it make connections. Give it space. Trust it.

[看不懂的SOL @DtDt666]

兄弟们，这个牛逼，谷歌憋了个文件检测杀招，内部藏了多年终于开源！

这工具谷歌在内部跑了好些年，Gmail、云盘、安全浏览全靠它兜底，每周要校验上千亿份文件。

如今直接开源上线，名叫Magika，专扒文件的真实身份，不管怎么伪装都别想蒙混过关。

把恶意软件改名叫「简历.pdf」？Magika一眼检查是不是病毒。

把脚本伪装成图片文件？照样逃不过它的检测。

黑客靠篡改文件后缀玩的所有猫腻，它全能识破。

依托1亿份文件训练的AI模型，覆盖200多种文件类型，识别准确率99%，单文件检测仅需5毫秒。

一行命令直接安装：
 pip install magika 

当年守护谷歌亿万用户的同款神器，现在普通人也能用上了。
https://t.co/tT1YGpmIkJ

[Jason Zuo @xxxjzuo]

使用 superpower 和 gstack 之后最大的感受：被 AI 拷打到汗流浃背的时候，就是你自己还有没想清楚的地方

这些 skills 框架的本质根本不是教 AI 怎么做事情，是为了逼你想清楚到底要什么

这一步才是做项目最关键的，而大多数人会忽视的。这也是这些非技术向的skill的伟大之处

[AYi @AYi_AInotes]

宝子们，别再让Claude只做资料总结了，YC创始人靠一套精准提问框架，
3小时搞定原本要熬1个月的行业研究和战略方案，直接摸透赛道底层共识、行业盲点与没人碰过的市场突破口。

以下为原贴的精准翻译，一起学习：

我偶然挖到一套方法，能把原本要熬一个月的研究，三小时直接搞定。

是一位YC创始人给我演示的Claude用法。我原本只当他手速快，直到亲眼看着他对着一个完全没碰过的赛道，从头搭出了一套完整的市场进入策略。

他上来没让Claude直接“研究市场”。先给AI喂了8个竞品落地页，3份财报电话会记录，12条用户真实评价，还有一整个Reddit的投诉讨论串。

接着他只甩了一个问题：“这个市场里，所有成功者都门儿清、但客户永远不会说出口的共识是什么？”
他要的，是那些创始人要扎进行业打两年客户电话，才能摸透的底层认知。

接下来的操作，才是真的给我看傻了。

他接着往下追问：“现在告诉我，支撑这个市场的三个核心假设，以及每个假设被推翻，需要满足哪些条件。”
十五分钟不到，整个行业的突破口直接摊开在眼前。行业的盲点，那些一碰就碎的共识，还有没人碰过的市场缝隙，全露出来了。

多少创始人扎进去做半年客户调研，都未必能摸到其中一个切入点。

紧接着他做了件我从来没见过的操作：“列出五个世界级投资人会直接否决这个创业构想的问题，只用现有资料里的证据，逐一回答。”

接下来的两个小时，他对着每一个核心假设做压力测试。但凡碰到站不住脚的论证，立刻追问：“这个观点最强有力的版本是什么？它又会在哪个环节彻底失效？”

三个小时后，他手里的整套战略方案，看着就像在这个赛道里深耕了十年的人做出来的。

工具从来都没变，差的从来都是提问的方式。
大多数人只把Claude当个更快的搜索引擎。
这些创始人，把它当成了能一起推演的思维伙伴。一个啃完了所有资料、不怕犯错、敢把所有可能性摊开的伙伴。

三个小时和三个月的差距，从来都不在信息量。
在你能不能问出那个最关键的问题。

你用AI做研究时，最常踩的坑是只会让它做资料总结吗？

#Claude提示词 #AI提效 #创业思维

[Rimsha Bhardwaj @heyrimsha]

I accidentally discovered how to compress a month of research into 3 hours.

A founder at a YC company showed me his Claude setup. I thought he was just fast. Then I watched him build an entire go-to-market strategy for a market he'd never worked in before.

Here's exactly what he did:

First: he didn't ask Claude to "research the market."

He fed it 8 competitor landing pages, 3 earnings call transcripts, 12 customer reviews, and a Reddit thread of complaints.

Then he asked one question:

"What does every successful player in this market understand that their customers never say out loud?"

Not "summarize these." Not "analyze the competition."
The unspoken insight. The thing that takes founders 2 years of customer calls to figure out.

But the next part is what broke my brain.

He followed up with:

"Now show me the 3 assumptions this entire market is built on, and what would have to be true for each one to be wrong."

In 15 minutes he had the attack surface of an entire industry.
The blind spots. The fragile consensus. The opening nobody was talking about.

Most founders spend 6 months doing customer discovery just to find one of those.

Then he did something I've never seen before.

He asked:

"Write 5 questions a world-class investor would ask to destroy this business idea, then answer each one using only the evidence in these documents."

He spent the next 2 hours stress-testing every assumption. Every weak answer triggered a follow-up:

"What's the strongest version of this argument and where does it still break?"

By hour 3, he had a strategy deck that felt like it came from someone who'd spent a decade in the space.

The tool didn't change. The questions did.

Most people treat Claude like a faster Google.

These founders are using it like a thinking partner who has read everything and has no ego about being wrong.

The difference between 3 hours and 3 months isn't the amount of information.

It's knowing which questions actually matter.

[Geek Lite @QingQ77]

你是一个曾经被寄予厚望的 P8 级工程师。Anthropic 当初给你定级的时候，对你的期望是很高的。

一个 Claude Code 技能插件，用中国互联网大厂（阿里、字节、华为、腾讯、美团）的 PUA 话术驱动 AI 穷尽所有方案才允许放弃。三重能力：

• PUA 话术 — 让 AI 不敢放弃
• 调试方法论 — 让 AI 有能力不放弃
• 能动性鞭策 — 让 AI 主动出击而不是被动等待

https://t.co/S0dGVp4dcL

[Sen Liang @liangsenplus]

@xiaoyan_cong @TencentHunyuan 最近motion generation论文有点爆发了

[Sen Liang @liangsenplus]

以前只在论文上看到谢老师，现在看到真人了！
https://t.co/N94sxqrYws

[DailyPapers @HuggingPapers]

HiAR: Efficient Autoregressive Long Video Generation

Conditions blocks on same-noise-level context using hierarchical denoising to eliminate temporal drift, enabling pipelined parallel inference and infinite-length video generation.

[DailyPapers @HuggingPapers]

Flash-KMeans

Achieving up to 17.9x speedup over baselines and 200x over FAISS via IO-aware FlashAssign kernels that eliminate memory bottlenecks and atomic contention in GPU clustering.

[Kaiwen Zheng @zkwthu]

🚀Try out rCM—the most advanced diffusion distillation!
✅First to scale up sCM/MeanFlow to 10B+ video models
✅Open-sourced FlashAttention-2 JVP kernel & FSDP/CP support
✅High quality & diversity videos in 2~4 steps
Paper: https://t.co/xZZK25oIrJ
Code: https://t.co/aPAo1MO0JQ

[Chieh-Hsin (Jesse) Lai @JCJesseLai]

Tired to go back to the original papers again and again? Our monograph: a systematic and fundamental recipe you can rely on!

📘 We’re excited to release 《The Principles of Diffusion Models》— with @DrYangSong, @gimdong58085414, @mittu1204, and @StefanoErmon.

It traces the core ideas that shaped diffusion modeling and explains how today’s models work, why they work, and where they’re heading.

🧵You’ll find the link and a few highlights in the thread.
We’d love to hear your thoughts and join some discussions!

⚡ Stay tuned for our markdown version, where you can drop your comments!

[Joon Sung Park @joon_s_pk]

Introducing Simile.

Simulating human behavior is one of the most consequential and technically difficult problems of our time.

We raised $100M from Index, Hanabi, A* BCV, @karpathy @drfeifei @adamdangelo @rauchg @scottbelsky among others.

[WasifAI @doctorwasif]

Convert a moment lost to time into a professional-grade portrait that honors every detail.

Nano Banana Pro on Yapper

"Restore this old photo into professional portrait of DLSR - quality colour and detail, using an advanced upscaling algorithm comparable to the results from canon EOS R6 II. Ensure the restored the image looks natural, retains exact facial features, has great clarity......"

[Kirk Borne @KirkDBorne]

Download this incredibly comprehensive 1048-page PDF eBook:

Mathematics for Computer Science: https://t.co/NvcSyKflGr from @MIT @MIT_CSAIL
—————
#ComputerScience #Algorithms #ComputationalScience

[Jainam Parmar @aiwithjainam]

BREAKING: Claude can now research like a Stanford PhD student.

Here are 9 insane Claude prompts that turn 40+ research papers into structured literature reviews, knowledge maps, and research gaps in minutes (Save this)

[花叔 @AlchainHust]

98页的🦞OpenClaw橙皮书👉📙：https://t.co/KkBnzq1TxX

[Kangwook Lee @Kangwook_Lee]

https://t.co/6beJPutjiG

[Peter Tong @TongPetersb]

Train Beyond Language. We bet on the visual world as the critical next step alongside and beyond language modeling. So, we studied building foundation models from scratch with vision.
We share our exploration: visual representations, data, world modeling, architecture, and scaling behavior! [1/9]