xAI's Best Days, and Why It Couldn't Keep Her
Sheng Ying calls early xAI v1.0: under 100 people, no politics, and both support and freedom at once. She left two months early for the SGLang community, walking away from a one-year cliff.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
What xAI gave her was support plus freedom
Sheng Ying says her few months at Databricks trying to promote SGLang were not successful: Databricks is a mature company, everything needs a sufficient reason, and she was a new grad researcher with not much power, and she would not drive a big effort inside a mature company. xAI gave her two things at once — support and freedom, and she describes it as ‘when I have support and freedom at the same time, I can do this thing’. Some of her other opportunities offered only support, some only freedom. When she joined, xAI had fewer than 100 people, and she says that was the best period she remembers, the first time she experienced a company with no people games at all, no politics.
— Sheng YingTwo months of vesting was still too long to wait
The SGLang community became a pure community project very early on; the core developers were all people met online, working together for two years without ever turning on a camera, and the understanding was ‘as long as I see your code, I know who you are’. In 2025 demand surged, but everyone was contributing in their spare time, staying up at night and working a bit more at noon, and the growth got so fast that this model could not hold, and delivery became a borderline fly-by. By July and August, Sheng Ying's anxiety had reached the point of ‘if we don't come out and fill in the hollow state, we are about to disappoint the outside world’. She was on a one-year cliff, had only been there ten months when she left, and two more months would have gotten her the first year's vesting, but by May she could no longer stand it.
— Sheng YingInfra is not support, Infra itself is the product
Sheng Ying says Infra people have long been underestimated within the full stack: they have to support researchers in getting experiments running, and support product people in running the product stably and fast, yet they are always forced to do hacky things themselves. Her rebuttal is ‘Infra itself, in my eyes, is the product’ — when you make a product you think about whether it is beautiful, whether it is usable, whether it hits human nature, and no one would say they hope Infra hits human nature, but the impact Infra produces is transmitted from the bottom up. She calls this relationship romantic, and says this is not her special case; the whole group of people who write Infra can understand it. RadixArk's approach is the mirror image: when a training job gets stuck, solve the problem systematically rather than patching it so it has to run.
— Sheng YingRadixArk's long-term mission is not in the inference race
Sheng Ying states plainly that although SGLang is Infra, RadixArk's long-term mission does not belong to this track. The goal she gives is ‘bring the best AI to everyone’, and she says the ultimate mission is to make the next generation of AI — not purely an Infra implementation, and not necessarily a model, but ‘something that does not exist in the world right now’. So she does not consider herself in the same category as the existing Inference Players, and precisely for that reason she is willing to keep good relations with every Inference company: SGLang does not belong to RadixArk, it belongs to the community, all the company's contributions are open source on the market, there is no private fork, and the company's lifeline is not here.
— Sheng YingOpen source is air, not deliberate insistence
Sheng Ying corrects the phrasing ‘insistent on open source’: she was born in Jiangxi, and when she wanted to learn programming the textbooks and resources on the market were extremely scarce, so her learning came almost entirely from random posts on the internet, CSDN, online judges; she did not know who wrote the code, or where in the world they were. She says she was raised by this group of people, so the culture of sharing openly exists like air, as a matter of course, ‘you have an idea, you do a thing, and if you don't share it, that is what feels strange’. She also points out the side effect of open source getting better: the admiration has mixed in something impure, and the presence of utilitarians makes those who originally believed unable to tell which parts are pure.
— Sheng YingEquality is not an education problem, it is a power problem
Sheng Ying gives an example from her competition days: play equally well, and if you are a boy everyone says he is a genius, if you are a girl they say she works hard, is obedient, got lucky. Every match she won had to be explained — the opponent was off form, this problem had been done before — while others' wins needed no explanation. She says each such thing on its own is very small, but it happens every day, every minute, every second, and the impact on the people affected has to be multiplied by a hundred million. Her conclusion is hard: there is only one way to solve this, and that is for women to truly have rights; through education you cannot reach any goal. She admits she cannot change structural problems, but ‘if I can change 1%, I am willing to spend my whole life changing that 1%’.
— Sheng YingBefore fundraising she did not know what a Termsheet was
Sheng Ying says when she first went out to raise money she knew nothing about this system: she did not know what a Termsheet was, did not know what valuation was, had never even heard the word P-S-D-K, and had no idea how many pages to write or what went on each page. She read books, and the books all described something beautiful — investors and founders in lifelong friendship, fighting shoulder to shoulder. Reality is far more complicated, but in the end she really did return with her P-S-D-K investors to the kind of relationship the books describe. Her summary is: there exist in the world such pure mentors, investors and interviewers, it is just that the vast majority of people do not persist to the end of that ups-and-downs process, do not persist to believing that the good thing exists.
— Sheng YingBeing split is fine, you don't have to escape it
Asked how a rigorous thinker faces a world that is random and a slapdash outfit, Sheng Ying says she is very emotional, small things cause big emotional effects, but she can separate emotional effect from value construction: seeing something she does not want to happen, she will be unhappy, but she will not therefore consider it normal. She says her positive feedback loop lets her wait longer and longer — the first time she waited a few months and was validated, the next time she waited years, the next time she would wait ten years. She thinks being split is completely OK, it is not really painful, people are just afraid; the only key point is ‘are you perceiving and acting according to your own ideas, or are you actually being led by someone else’.
— Sheng YingIn their own words · checked verbatim
One is called support, one is called freedom. When I have support and freedom at the same time, I can do this thing.
一个叫做support 一个叫做freedom 同时有support和freedom的时候 我就能做这件事情了
Sheng Ying34:37
If we don't come out and do this thing, if I don't fill in the blank behind it — this hollow state — if I don't fill that blank, we are about to disappoint the outside world.
如果我们再不出来 做这件事情 我再不把这个背后 实际上的空白 就是这个是个中空的状态 我再不把这个空白填充的话 我们就快要让外界失望了
Sheng Ying46:22
I think Infra is not a support role. Infra itself, in my eyes, is the product.
我觉得Infra 它不是一个support的角色 Infra本身在我眼中就是产品
Sheng Ying57:26
I was raised by this group of people, by people on the internet whose place in the world I didn't know — I was taught by them.
我是被这样一群人 被互联网上 不知道世界上在哪里的人 给raise起来的 就是我是被他们教会的
Sheng Ying1:20:30
I said I can't change it, I really have no way to change it. But if I can change 1%, I am willing to spend my whole life changing that 1%.
我说我不能改变 我确实没办法改变 但是如果我能改变1%的话 我愿意用我的一生 去改变这1%
Sheng Ying1:31:35
But every win of mine had to be explained, there always had to be a reason behind it called why I won, rather than why you lost.
但是我的每一场赢都要被解释 背后都有个原因叫做我为什么赢 而不是你为什么输
Sheng Ying1:35:36
There exist in the world such pure, such good mentors, such good investors, such good interviewers. But the vast majority of people, they don't persist to the end of that ups-and-downs process.
世界上存在 那么纯粹 那么好的导师 存在那么好的投资人 存在那么好的采访者 但是绝大多数的人 他们没有坚持到 把那个跌宕起伏的过程 走完
Sheng Ying1:41:36
Figures
| RadixArk total funding | $100 million | 0:00 |
| RadixArk team size | 40-plus people | 1:03 |
| Headcount at xAI when Sheng Ying joined | fewer than 100 people | 38:18 |
| Length of Sheng Ying's tenure when she left xAI | ten months (one year cliff not reached) | 46:22 |
| Valuation of the new company from the vLLM core team | $800 million | 1:05:28 |
| Baseten funding | $300 million | 1:08:28 |
| Baseten valuation | $5 billion | 1:08:28 |
| Fireworks AI valuation | $4 billion | 1:08:28 |
| New funding sought by Together AI | $1 billion | 1:08:28 |
Glossary
- SGLang
- An open-source large-model inference engine that uses RadixAttention for prefix cache reuse; the capstone of Sheng Ying's PhD work.
- RadixAttention
- Uses a radix tree to index the common prefixes of requests, reusing already-computed KVCache to avoid recomputation.
- KVCache
- Caches the keys and values of attention computation at inference time, so later requests in a multi-turn conversation can reuse the earlier history.
- Miles
- An RL framework developed by RadixArk; when DeepSeek released a new model it was the first to achieve day zero RL compatibility.
- day zero
- The inference engine completes adaptation and support on the very day a new model is released; the market's expectation of the Infra side.
- one year cliff
- Equity vests only after a full year; leave early and you get not a single share.
How to listen
Engineers doing AI Infra, maintainers of open-source projects, and technical people leaving big companies to start up, especially those interested in the part about an organization moving from rule by people to systems.
The stretch from 2:16 to 26:01 on school, the low period, and Empresses in the Palace has nothing to do with AI judgment.