Nyla Z

𓇼 ⋆.˚ 𓆝 𓆡⋆.˚ 𓇼in sunny california

@UCLA

[Nyla Z @NylaZ_02]

If you're in tech and wants to have a conversation on AI, let's connect!!

[Nyla Z @NylaZ_02]

第一次被小红书网友在线下活动认出来…哦不对，应该是第三次😭

[NVIDIA Newsroom @nvidianewsroom]

NVIDIA announces CEO transition. Read the release: https://t.co/MTTUvFHCJA

[Nyla Z @NylaZ_02]

Why AI Infrastructure Matters

Most conversations about AI start with models and apps. Which model is smarter? Which product is better? I recently read a post about why class advantages persist, and it made me ask a different question: every time an AI product serves a user, who gets paid along the way?

The post used “infrastructure” broadly to mean things people repeatedly rely on: housing, licenses, platforms, brands, and distribution channels. Rent is a simple example. Part of your paycheck goes to your landlord because they own a place you need to live. They can reinvest that income in other assets, and some of those assets can be passed down. Over time, ownership generates income that can buy more ownership.

You can see a similar dynamic in AI. An app relies on compute, cloud services, and inference software to serve users. NVIDIA is a good example. When teams have already built their deployment workflows and tooling around its ecosystem, sticking with that stack can be the easiest choice when they need more compute. The hardware generates revenue, and the software ecosystem gives teams a reason to keep buying it.

What interests me more is how infrastructure changes which products are viable in the first place.

An agent might call a model dozens of times to complete a single task. A small increase in the cost per call can eat into the margin on the entire task. Bring inference costs down far enough, and a workflow that only made sense as a demo might become a product users can afford. Or take prefix caching. When requests share a long prefix, an inference engine can reuse work it has already done and cut the time users wait for the first token. Those improvements shape what teams can charge, how responsive the product feels, and which workflows are worth shipping.

That’s part of what I want to explore at the SGLang Summit. Open source gives infrastructure projects a way into production: developers can try the software and deploy it themselves. Once teams build their deployment scripts, monitoring, and performance tuning around an inference engine, it becomes part of how they operate. The question I’m curious about with SGLang and RadixArk is what teams will pay for once they already use the open-source engine. Is it managed hosting, operational support, or something else?

It also changes how I think about starting a company. You put time and money into solving a problem, deliver something customers can use, and gradually earn their trust. When the same problem comes up and they come back to you, you’ve earned a place in their workflow. Falling costs can create openings here. A task that was too expensive or too slow to automate last year might be worth building a business around today.

I used to think infrastructure was far removed from the work I wanted to do. Now, when I look at a role or a project, I ask more concrete questions. Will I get to see why customers pay? Which upstream decisions determine what I can deliver and at what cost? What problem comes up often enough that I could get good at solving it?

An early engineering role can give me access to customers and workflows I’d struggle to reach on my own. That’s where I’d like to learn which part of the workflow is worth building for someday.

[Nyla Z @NylaZ_02]

为什么 AI 基础设施重要？从谁在收「过路费」说起

聊 AI 的时候，我们很容易盯着模型和应用：哪个模型更聪明，哪个产品更好用。最近看到一篇讲「阶层为什么会固化」的帖子，让我想到另一个问题：一个 AI 产品每服务一次用户，这笔钱会经过谁，最后又流到谁手里？

那篇帖子把房子、牌照、平台、品牌和渠道都叫作「基础设施」。它们有一个共同点：别人做事的时候会反复用到。房租就是个例子。你每个月工资里的一部分会沿着产权关系流到房东手里，而这笔收入又可以用来投资新的资产。有些资产还能传给下一代。持续收益、再投资和继承，让先占住的位置越来越稳。

AI 产业也有类似的结构。一个应用要服务用户，背后需要芯片、算力、云服务和推理软件。NVIDIA 的优势就可以从这个角度理解。当开发工具、模型和部署流程都围绕一套生态建立起来，团队每次扩大算力投入，往往会沿用已经熟悉的技术栈。硬件带来收入，围绕硬件积累的软件和工具又让这笔收入有机会持续下去。

但我觉得，基础设施更有意思的地方在于，它会改变哪些产品值得做。

比如，一个 Agent 完成一次任务，可能要反复调用模型。单次调用贵一点，乘上整个任务的调用次数，可能就吃掉了产品的利润。推理成本降下来，原来只能做成 demo 的工作流，才有机会按用户能接受的价格跑起来。再比如，很多请求共享同一段很长的前缀。如果推理引擎能复用已经算过的部分，团队就可以减少重复计算，缩短用户等待第一段输出的时间。对做应用的人来说，这些变化会直接影响定价、交互设计，以及什么场景值得上线。

我这两天要去 SGLang Summit，想看的也是推理引擎这一层。开源提供了一种建立基础设施的方式：先让开发者能免费用、能自己部署，再看它能不能进入团队的生产系统。等部署脚本、监控和性能调优都围绕它积累起来，这个项目在生态里的位置就有了价值。我对 SGLang 和 RadixArk 感兴趣的地方是，开源采用之后，哪些需求会让团队愿意为托管、运维或者企业支持付费？从「大家愿意用」到「有人愿意付钱」，中间还有一段路。

这也让我重新理解创业。你先花时间和钱解决一个问题，做出产品、完成交付，再慢慢积累客户和信任。等客户下次遇到同样的问题还会回来，你就在他们的工作流程里有了一个位置。新技术降低成本时，往往也会让一些以前太贵、太慢的需求变得值得解决。后来者的机会，就可能出现在这些刚刚变得可行的地方。

以前我觉得基础设施离自己很远。现在看一份工作或一个项目，我会具体问：我能看到客户为什么付钱吗？我做的东西，有多少成本和交付能力取决于上游？如果我以后想自己做产品，有哪个问题会反复出现，而我已经懂得怎么把它交付好？

第一份工作可以借公司的渠道接触客户。至于以后想修哪一段自己的路，我更想从这些真实的需求里找答案。

[Nyla Z @NylaZ_02]

Anyone else at SGLang Summit tomorrow? I'll be at Fort Mason Thu + Fri for the agent and inference sessions. Building agents or hiring 2027 new grads? Let's meet up, DMs open.

[Nyla Z @NylaZ_02]

love hyping people up fr🙌🤣

[Nyla Z @NylaZ_02]

hey, i’m Nylaaaaaa

20 y/o girl in tech @ ucla

side quests: cooking & baking, painting, editing vids, pottery, daydreaming, weekend costco combat, boba & cake field research, and hyping up other rising stars;

spent 5 years tutoring k–12 kids in english part-time;

post memes and unhinged videos on RedNote. somehow, got 13k likes.

just trying to make the world a little better while i’m here🙌

when i’m stressed, i chug coconut water and order hot pot spicy enough to give me a new problem

💻coding for fun (until it stops compiling)

[Nyla Z @NylaZ_02]

@dviolettchan 不一定，我国内导师很好，手把手教，喂科研idea，经常hands on改各种细节。但他也是北美回来的lol

[Nyla Z @NylaZ_02]

记录你自己的Vision Board：

有时候会在网上看到一些地方的美景，会保存下来加到一个特定的相册里，久而久之那个相册就变成了愿景板合集。

有个说法是当你对一样东西有desire的时候其实在另一个时空维度上已经拥有了。

不要小看自己对一些事物、想去的地方的渴望和关注中所蕴藏的巨大的能量。

三年前，我的vision board里是加州的大海。

除了风景，一些世俗意义上的成就和追求同样适用。

[Nyla Z @NylaZ_02]

I thought I needed experience to get an internship.

Turns out, the internship is where I got the experience.

During my sophomore year, I came across an internship opportunity in a field I had no direct experience in.

I almost talked myself out of it.

But I reached out anyway.

To my surprise, the team was interested in my coursework projects and undergraduate research experience.

Things I had considered “just school projects” turned out to be relevant to what they were looking for.

I got the internship.

And that’s where I started developing the skills I thought I needed before applying.

Looking back, I wonder how many opportunities we miss simply because we assume we’re not qualified enough.

Sometimes the opportunity comes before the expertise.

Don’t reject yourself before the company does.
#JobSearch #newgrad
#jobhunting #Internship #NewGrad #TechCareers #CareerGrowth

[Lilian Weng @lilianweng]

I've always been a curiosity-driven person. Leaning new knowledge or figuring out how to solve problems that are not well defined often gives me simple joy.

The fun part of being a cofounder is how it "forced" me to develop new perspectives on company strategies, team building (at the company level), and much more, as well as how these seemingly abstract terms are deeply connected with our day-to-day actions, planning and narratives.

With real-world experience, ideas that once felt theoretical become approachable and it is true for many terms in many domains.

[Sanbu 散步 @sanbuphy]

想学 Agentic RL，但被4卡、8卡和一大堆训练部署配置劝退过的，可以看看这个仓库。

Datawhale 成员宋志学做了一个 Agentic-RL Lab，专门拆解和复现 LLM / Agent 强化学习算法。从 GRPO、DAPO、GSPO，到 Search-R1、ReTool、AgentOPSD，都有对应的讲解和代码。

它借助 PyTRIO 的 Train API 降低训练工程的门槛，让学习者能把注意力放在几个具体问题上：数据怎么准备，reward 怎么给，loss 在优化什么，一条 Agent 轨迹怎么参与训练。

比如 GRPO 章节，会在 GSM8K 上比较三种 loss；Search-R1 训练多轮搜索，ReTool 训练模型调用代码工具，还有在 ALFWorld 里做家务、用光谱数据找星星的实验。可以顺着一个个任务，看论文里的公式最后怎么变成训练代码。

论文出处、核心变量、数据处理、训练循环和 SwanLab 记录都放在一起了。想研究某个奖励设计或者 loss 改动，可以先找一个对应的实验跑通，再在这个基础上改。

挺喜欢这种学习材料：读到一个算法，旁边就有能动手试的代码。想开始学 Agentic RL 的朋友，可以挑一个感兴趣的章节看看。

https://t.co/6JxKiT8Fsd

[Professor Oak @Prof_Oak_]

Surprise 1: Anthropic acquired a one-person biotech
.
.
.
Surprise 2: that person is me
Surprise 3: I’m here to tell you all about it

Yes, this is what you think it is; after 6 yrs of increasingly loose anonymity: the Professor Oak reveal post

…and yes, it has gotten so boring around here that I’m literally doxxing myself to spice up the timeline

-------
Ok, allow me to introduce myself. My name is Cody Tranbarger. I mostly tweet, but occasionally find time to dabble in drug dev & starting/building biotechs. My past lives span the VC, crossover, startup continuum; Prof Oak has been my way of keeping a digital foot in those worlds, despite having my physical feet planted elsewhere.

The past 18 mo, ‘elsewhere’ was Mithril Medicine, a co. I founded to develop bsAbs for LAMA2-CMD & other ECM-related genetic diseases.

Mithril was born out of hands-on building with @MikeNGladstone, @JasonPaulRhodes, @LifeSciVC, & team @atlasventure. Somehow, I made enough half-decent calls along the way (even a broken clock is right twice a day, eh?) to convince them Mithril/I warranted a small seed. While I can hardly rationalize such an outrageously cavalier bet, I’m grateful for it, & I'd partner with Atlas again in a heartbeat.

In typical AV fashion, Mithril’s primordial form was virtual, decentralized, & ultra-lean. This stage tends to be short-lived, but for Mithril, it worked surprisingly well, for surprisingly long, thanks in part to the rapid maturation of the AI tools I used liberally.

Eventually, Mithril became a bit of an experiment itself - to find the upper bound of how far it could scale as a single operator + a bunch of Claude (+ a stellar bench of part-time consultants).

In parallel, I started spending time with @nc_frey, @an1lam, & @CoefficientBio team; and promptly experienced, for the 1st time, AI at the frontier. Soon thereafter, Mithril stopped looking like an experiment, & started looking like a preview of an entirely new operating model.

From that spark came “Pipeline-in-a-Person.” The idea immediately resonated when we shared it here back in April; and it was central to the shared vision that led to Anthropic’s acquisition of Mithril, & my joining the team.

Why join Anthropic? Three reasons.
(1) For Mithril, the programs continue, with more resources & infrastructure behind them, and as much Claude as I can put to work.

(2) For me, the chance to start several more programs, & find out how many one person can actually run in parallel - an almost unprecedented luxury for a curious drug hunter.

(3) For our mission, a path to institutionalize & scale this new way of working, then put it in the hands of all of biopharma, all over the world - which, candidly, I find the most energizing.

Will AI x Bio cure diseases? Unequivocally yes, imo… but not by the route that gets the most airtime. Rather than timeline acceleration & compression, I see orchestration & parallelization as today's low-hanging fruit.

Most drug hunters see a handful of go/no-go calls in a career, & spend entirely too much of the rest trading emails (my inbox can attest).

Give them the bandwidth to run several programs at once, and the time to focus on scientific judgment, and I’m willing to bet they’ll make better calls. And at scale, better calls translate to more drugs, higher hit rates, broader patient impact, & likely, a disease or 10 or 100 cured. I couldn't be more excited to build toward that future @AnthropicAI.

Finally, a heartfelt thank you to BioTwitter/X. The end of anonymity is bittersweet, but if anyone’s earned a proper introduction, you all have.

It’s difficult to overstate how much this app has done for me: a decade of education, plenty of friends, many job opportunities, several mentors, a few investors, even a co-founder (fmr). It was the generosity of this community that inspired me to start posting 6 years ago, after just as many years lurking. I hope I’ve been able to pay it forward, in my own small way.

But, to be clear, Prof Oak isn’t going anywhere. Same beat, same themes, and now, a front-row seat to drug development at the frontier. Stay tuned for dispatches from the trenches; and remember...

Gotta Cure 'Em All

[头雁 @alacheng]

前沿Neo Lab Top 100+名单，包括他们在做什么，融资规模，创始人以及他们的X账号。

https://t.co/la5hMS8ZhH

[Viking @vikingmute]

超级好的文章，《如何与 AI agent 协作》 as of October 2026，作者在 AmpCode 工作，里面没有啥复杂的大道理，核心思想就是大模型现在非常强大，不需要那么多约束和规则，充分相信他的能力，并且给予适当的规则即可。摘抄几条都短小精悍但是深得我心 看了好几遍

1 用最新 最强的模型，差的模型让你反复浪费时间 时间也是成本

2 别再担心 AI 会写出 slop ，它的能力已经够强大，即使有些冗余，也不重要

3 告诉大模型你到底要什么 别告诉它具体要怎么做

4 让 agent 能看见整个流程，这个说的非常好，整个流程应该都对 agent 可见 能给多少给多少 减少幻觉 E2E 生产日志 部署 等等 多给它权限

5 要证据，要大模型给你 E2E，对比数据 截屏等能证明它完成了，写到 agents.md 中作为验收条件 代码风格等等不要管了

文中还有很多条很多条，每条都很好，都值得拿出来好好念一遍，大家要好好看看

[Roc @AIStarsGo]

几条教训

1. 不要依赖 agent 读论文或者高质量博文，
后果：思维深度大幅度下降，灵感触发时机大幅度减少，浪费大量的时间去理解 agent 的垃圾表达。

2. 不要把搞科研/技术学习 和 英语学习混在一起。
后果：效率低下，推进缓慢，失去耐心后放弃，灵感和思考都需要时间积累，混合学英语这个目标会大幅度延迟你的思考，让你的脑子在来个任务上来回切换

3. 不要因为 token 用不完就想着 background 跑点东西，把 token 用掉，不影响你的主线
后果：除非你编了稳定的 workflow，否则即时奖励更强的 vibe coding 一定会吸引你的注意力，打断你的深度思考。

总结，一定要感受自己在 agent 时代下的认知水平的提高速度，到底是加快了，还是不变，还是退步了。如若不然会沉迷于 shallow work，最后认知退化，导致阅读长文字、深度思考都没耐心，人就废了。

[Tibo @thsottiaux]

Day 22/ Terra is back!

Our first model that is significantly better at writing, personality and nuanced understanding of requests.

Packs a punch across the board, very good at computer use, and gives 2X more usage than 6.1 Sol across API and subscriptions.