🐶puppy

06 185 ♂ 大二在读 | 找固炮 | 无偿 | 可长期/短期 | m | 💗🦶
may be situationship?
人机/有偿勿扰
可私

[Roxy @DrRabbitAI]

本期分享 AI 在课题申报中的���用场景。 https://t.co/K3XU0F2yR7

[Axiom @axiommathai]

Most think vibecoding at scale is slop. Here’s a counterexample.

Today we release a 16M-line formalization of the entire Stacks Project, critical infra underlying modern math, with optimized code quality and its highly interconnected structure preserved.

A mathematical city 🌆

[Michael Guo @Michaelzsguo]

世界上最聪明的大脑之一，普林斯顿大学理论物理学家 Edward Witten，居然也开始用 Claude 帮自己做前沿物理研究了。

Witten 是现代弦理论最重要的奠基者之一，也是历史上第一位获得菲尔兹奖的物理学家���他在数学和物理学之间建立了大量深刻联系，尤其是推动了 M 理论的发展，改变了人类对弦理论和量子引力的理解。

11 月 3 日，Witten 发表了一篇 36 页的新论文（我表示完全看不懂），研究困扰理论物理学界多年的难题：如何在离散的时空格点上定义包含手征费米子的规范场论。

有意思的是，论文里三次提到了 Claude Opus 5.5。

这可不只是用 AI 润色论文。Witten 明确写道，Opus 5.5 提出了一种新的数学构造变体，用于将离散格点上的规范场平滑地延拓到连续空间。它还提出了另一种基于线性插值的方案。

在论文致谢中，Witten 特别写道：

“Claude Opus 5.5 在许多细节上提供了帮助。”

当一位站在理论物理学最顶端、自己就能推动整个学科发展的科学家，开始在正式研究论文中明确感谢 AI 对数学构造的贡献，这件��本身就很值得关注。

AI 正在从帮助科学家写代码、整理文献，走向直接参与前沿科学研究中的技术推导和方案探索。

[小马 | 商业生存指南 @VibePonyAI]

表达能力差，真的会让你错失很多机会。

2019 年，MIT 教授帕特里克·温斯顿用一场 1 小时的演讲，
讲透了如何清晰表达、抓住听众注意力。

时隔多年再看，依然值得认真学一遍。

[Nikiniki @Nikinik58681513]

今天是我在Bytedance的Lastday～
谢谢字节圈的大家对我的祝福。
我重新介绍下自己，我是niki，马上快到了我27岁生日，也是我做后端开发的第四个年头，我做了一个非常大的决定，离开了大厂去startup尝试一种新人生。

过去几年，我逐渐熟悉了大公司的游戏规则，怎么扩大 scope、争取资源、证明技术价值，怎么把项目变成对团队和业务的贡献。我乐在其中，享受这份成就感，却也逐渐感到麻木。大厂不会死，即使是不赚钱、无聊的业务，只要能找到切入点去创新与协作，也能创造内部价值。

今年我慢慢的开始思考组织带来的认可和社会市场需要的价值差异在哪里。
我开始好奇如果离开这套评价体系，我能不能做出有价值的东西？

另外，AI。
不可否认，AI 已经在降低知识和软件工程的门槛，甚至未来 AI 也会逐渐改变 research 的方��。
所以当技术越来越容易获得，什么东西会变得更有价值？
过去，能够把产品做出来，本身就是一种竞争优势。而当所有人都具备这种能力时，找到真实的问题、理解用户、判断市场，以及把产品带给用户，就会变得更加重要。这些所有都会到一个问题GTM。
Uber、Owner都是很有趣的市场策略，在一个已有成熟玩家的行业里，如何找到切入点，改变用户习惯背后的某个关键环节，最终建立自己的市场。
但这不意味着技术不再重要。相反，AI 让我们有机会用更少的资源去解决过去不敢想象的问题。

从 2022 年开始，我们这一代软件工程师几乎完整经历了 LLM 如何一步步改变软件行业。每次看到新东西的时候都感到焦虑FOMO，但也让我感到兴奋。
过去积累的专业知识遭到了迅速贬值，但能尝试的事情也远超几年前。
不知道未来会变成什么样子，但我感觉与其努力守住自己已经掌握的知识，不如学习如何在不断变化的世界里持续创造价值。

我很喜欢那些大胆得近乎疯狂的想法。比如 Musk 对商业航天和Roadster。

上周一个朋友和我说：“我欣赏Musk但是又嫉妒他，为什么改变人类世界的是他，而不是我？”

听上去很ego但是有感觉很能共鸣，我也想改变世界，被时代记住。但是很天真，绝大多数人终其一生，都不会成为被历史记住的人。
所以我想如果做不到的话，我也希望自己能够站在技术发展的最前沿，见证人类社会的进步。

我并不认为创业比在大厂更有意义，也不认为冒险本身就���得赞美。

相比继续沿着一条已经相对熟悉的道路往前走，我更想知道，在一个充满不确定性的环境里，自己究竟能做成什么。
可能会失败，可能会发现很多判断都是错的，也可能几年以后重新回到大公司。
但我希望多年以后回头看，至少能吹自己曾经认真参与过这个时代，而不只是旁观。

人生是一件很有意思的事情，每个人都只有一次独一无二的经历。既然未来本来就无法被准确预测，那就保持好奇心、思考能力，以及重新做出选择的勇气。
焦虑是真实的，但也会带来机会，希望大家都能顺利活得精彩～

[Dobby @lesliedqbb]

最近看到一堂斯坦福公开课，用大约两小时，系统梳理了大语言模型从文本输入到训练、推理和生成结果的完整链路。
它不会让你两小时造出一个 Claude 或 ChatGPT，但能帮你建立一张非常重要的“大模型知识地图”。
课程重点：
00:00｜LLM分词机制 文本进入模型前，为什么必须先被拆成 token？不同分词方式又会怎样影响训练效率与模型表现？
25:44｜模型如何解码提示词 大模型并不是一次“理解”完整句话，而是根据上下文计算概率，逐个预测接下来最可能出现的 token。
35:40｜大模型训练流程 从训练数据、预训练到参数更新，了解模型能力究竟是怎样一步步形成的。
1:16:47｜从零理解LLM架构 拆解嵌入层、注意力机制、Transformer 模块和输出层，看清语言模型内部的数据流动方式。
我最大的感受是：
很多人每天都在研究提示词和新工具，却没有真正理解模型底层是怎样工作的。
当你理解了 token、注意力、训练和推理之后，很多原本模糊的概念都会连起来：
为什么上下文长度有限？ 为什么模型会产生幻觉？ 为什么推理速度如此重要？ 为什么训练一个大模型这么昂贵？
建议先完整看一遍建立框架，再根据时间点回看不懂的部分。两小时不能让你成为大模型工程师，但足以帮你���断下一步应该学什么。

[Yacine Mahdid @yacinelearning]

alright folks tomorrow 11:00 AM EST until whenever we are done we are going to be hosting a workshop led by francesco on robot learning!

in this workshop we're going to cover:
- Classical robotics
- Pre-training vs. post-training
- Real-world RL
- VLAs / VAMs / WAMs

and more!

[Xi Yin @XiYin137]

Today is my last day at OpenAI. It’s been an awesome ride, but I’m getting off a little early to pursue a new direction. I feel fortunate to have worked alongside some of the most incredible people I know. I’m especially grateful to @ALupsasca and Boris Alexeev for helping me navigate my first corporate job, and to @danintheory for his constant support. I’ll be rooting for you all!

[Samip @industriaalist]

Humans are ~1,000,000x more sample efficient than LLMs, and there are no good explanations.

We’ve found that a big part of the answer is that LLMs are *massively* underparameterized by many orders of magnitude. The brain has 100T synapses learning from ~200M tokens of language. Chinchilla prescribes a 1T-parameter model on 20T tokens. That’s 7 OOMs apart in parameters per token.

After a lot of painstaking work, we present new data-constrained scaling laws and learning algorithms to scale to (extrapolated) quadrillion-parameter transformers.

w/ @akshayvegesna, @EthanBaron78553, @asatheesh6561
Full paper led by @EthanBaron78553, with @ShmuelBerman, @andrewgwils

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Mustafa. @oprydai]

tensor calculus is one of those subjects that completely changes how you understand geometry and physics. ordinary calculus tells you how quantities change, but tensor calculus goes a step further. it asks how quantities change when the coordinate system itself is changing, especially when you’re working on curved surfaces or spacetime. a vector might have different numerical components depending on how you describe it, but the physical quantity it represents remains the same. tensors give us a mathematical language to preserve that underlying structure regardless of the coordinates we choose.

the interesting part begins with the metric, which tells us how to measure distances and angles. then comes the connection, which tells us how to compare directions at neighboring points. on a curved surface, you cannot simply move a vector from one location to another and assume its direction remains unchanged. parallel transport gives us a rule for moving it, while the covariant derivative measures how a tensor changes after accounting for the changing geometry. once you understand these ideas, curvature stops being something you only visualize and becomes something you can calculate. this is the mathematical machinery that allows us to describe curved spaces consistently.

and this is where tensor calculus becomes fundamental to general relativity. einstein’s field equations connect the geometry of spacetime to the energy and momentum of matter. gravity is no longer treated simply as a force pulling objects toward one another. it becomes a consequence of the geometry through which objects move. the same mathematical ideas also appear in continuum mechanics, fluid dynamics, elasticity and differential geometry. what i find beautiful is that tensor calculus is really about separating physical reality from the coordinates we use to describe it. the equations may change their appearance, but the underlying relationships remain invariant. that’s one of the deepest principles in mathematics: a good description of reality should not depend on the perspective from which you observe it.

[om jambhale @omjambhale3]

if i were starting from zero in physical ai, these are the people i'd follow

software (policy):
@DrJimFan gr00t, nvidia gear lab
@ericjang11 robot learning, world models
@julianibarz optimus ai lead, tesla
@svlevine robot rl, physical intelligence
@xiao_ted rt-2, ex-google deepmind robotics
@JasonMa2020 eureka, reward design, dyna robotics
@chelseabfinn robot learning, physical intelligence
@hausman_k pi0, physical intelligence
@chris_j_paxton ai lead, agility robotics
@yukez co-leads nvidia gear lab, ut austin
@xiaolonw humanoid teleop + learning, meta
@kevin_zakka mujoco playground, berkeley
@pulkitology dexterity + sim2real, mit
@TairanHe99 humanoid whole-body control (asap)
@LerrelPinto dexterous manipulation, meta
@chichengcc diffusion policy + umi
@ankurhandos dexterity team, nvidia
@GuanyaShi agile humanoid control, cmu
@xbpeng4 motion imitation (deepmimic, amp), nvidia
@KarlPertsch openvla + fast tokenizer, pi

hardware:
@BerntBornich 1x, designs its own motors
@tonyzzhao built aloha, now sunday
@_milankovac_ ex-optimus engineering lead
@GoingBallistic5 actuator + humanoid teardowns
@zipengfu built mobile aloha
@alexkoch_ai koch open-source arm, tau robotics
@dhanushisrad clone, artificial-muscle androids
@philippswu gello teleop arms, xdof
@therobotstudio so-100 arm + open humanoid hands
@kenny__shaw leap hand, low-cost dexterous hands
@Raunaqmb anyskin, robot touch sensing

companies (mostly for updates):
@Tesla_Optimus optimus
@BostonDynamics atlas
@Figure_robot figure humanoids
@UnitreeRobotics g1, h2
@NVIDIARobotics isaac, gr00t
@1x_tech neo
@physical_int pi models
@agilityrobotics digit
@GeneralistAI gen models
@sundayrobotics memo
@LeRobotHF open-source robotics
@pollenrobotics reachy

who else would you add?

[Zhanyi Sun @s_zhanyi]

Can more pretraining improve a robot policy’s success rate, yet makes it worse for porttraining? 🤔

In our new work, we find that it can. Continued supervised pretraining of generative robot policies can improve reliability at the expense of coverage, and this tradeoff affects downstream posttraining.

Read on to see why this happens and how we can preserve coverage. 1/n 🧵

[matharium @matharium]

Every inner product space is normed, every normed space is metric, and every metric space is topological. Each level adds one piece of structure.

A topology says which sets are open. A metric adds a distance, a norm adds size with d = ‖x − y‖, and an inner product adds angles with ‖x‖ = √⟨x, x⟩.

Completeness gives the named classes: a complete normed space is a Banach space, a complete inner product space is a Hilbert space, and ℝⁿ with the dot product is a finite-dimensional Hilbert space.

[DAN KOE @thedankoe]

https://t.co/aOjXzzvlv8

[cat @_catwu]

Friday was my last day at Anthropic.

Two years ago, when Claude Code was a small internal tool, I asked it to build an RL env. I steered it for tens of hours but never manually wrote a line of code. It worked, I was hooked, and I joined the team.

When we first launched, most people approved every permission prompt by hand and mostly used Claude to answer questions and write tests. 

Today, the default is trusting Claude's judgment to auto-approve. Claude does large migrations and builds new products. 

Helping build Claude Code into an industry-defining product has been the journey of a lifetime. 

Thank you to the incredible team behind Claude Code, Cowork, and Tag. These products are the sum of thousands of great ideas from you. I'll miss working with you all.

Thank you to our users for embracing a new way of working. Your constant feedback shaped so much of what we shipped and your stories made the work so meaningful.

Excited to see what you all make next!

[Praxis @praxiss]

I feel this article is well worth reading. It helps dispel the myth that critical thinking can be applied to all situations. Many educational programs and business leaders have advocated for learning “higher-order thinking skills,” assuming these skills will then be++

[Pablo Vela @pablovelagomez1]

Splatting is going to be a critical part of the ego collection pipeline IMO, so I'm working on improving our current official support.

code here - - <https://t.co/BaB9WPtWyg>

The current shipped version does well for less than 1M splats, but not so much for larger ones such as the iconic garden scene (5.2M splats). I ported over brushes kernels to @rerundotio and now on my M5 I'm getting ~40-50fps when recording or a solid 60fps when not. This is basically a 10x improvement from the current version and this huge improvement that wouldn't have been possible without all the verification loops we've added with the rerun skills and MCP. I'll break down what was required to make this work ⬇︎

[Tony Dunn @tony873004]

Happy November! This month Voyager 1 will become the first spacecraft to reach 1 light day from Earth. 🔭 #astro

[An Xuan @anxuanng]

Opus 5.5 built a live warehouse from NVIDIA's free SimReady models