POP Art Style Drawing Model＆Artist : @KateHew03 Outlook on life: "If you cannot stand a situation, Quietly turn around and leave, If you cannot get along with a friend, Let the relationship end according to conditions." 看不惯的场景, 淡然转身而去, 聊不来的朋友, 随缘散去。

OpenAI 跟 Cerebras 那笔 750MW 的合同今年 1 月签的，1 月份报道的合同价值超过 100 亿美元，后来 S-1 招股书披露的 Master Relationship Agreement 总价值超过 200 亿美元（包含到 2030 年的扩展选项）。第一个产品 2 月 12 日就上线了，叫 GPT-5.3-Codex-Spark。这个产品本身就是 Cerebras 推理引擎实战表现的一份成绩单，可以从几个方面做具体判断。 吞吐方面，Codex-Spark 在 WSE-3 上跑出 1000 tokens/秒以上，标准 GPT-5.3-Codex 跑在 GPU 集群上约 65 tokens/秒，速度差距 15 倍。Cerebras 自己在 gpt-oss-120B 这个开源模型上能跑到 3000 tokens/秒，是当前所有商用推理服务里最高的吞吐。第三方 benchmark 横向对比，同样 gpt-oss-120B 模型在 Cerebras 跟在普通 GPU 推理后端的吞吐差大约 10 倍。这个 10 到 15 倍的实测差距，跟物理层面 Cerebras SRAM 21 PB/s 对比 H100 HBM3 3.35 TB/s 那约 6300 倍带宽鸿沟之间，还有几百倍的 gap，被软件栈、工作负载、batch 处理这些系统级因素吃掉了。 延迟方面，1000 tokens/秒对应每个 token 1 毫秒生成间隔，这个速度的硬件能力反过来暴露了 OpenAI 自己推理 stack 的瓶颈。OpenAI 在 Codex-Spark 上同时引入了持久 WebSocket 连接和 Responses API 重写，每次客户端服务端往返开销减少 80%，per-token 开销减少 30%，time-to-first-token 减少 50%。原来 OpenAI 的整套推理服务架构是按 GPU 推理"几十毫秒一个 token"那个延迟尺度设计的，跟 Cerebras 1 毫秒每 token 的硬件能力对不上，软件栈的开销反而成了瓶颈。Cerebras 的速度让 OpenAI 必须重做整个 web 层，硬件能力升级倒逼软件栈重构。 成本方面，Cerebras 的硬件成本优势目前还没有被公开定价验证。Codex-Spark 只对 ChatGPT Pro 200 美元/月订阅用户开放，API 定价至今未敲定。OpenAI 选择高 ARPU 订阅档而不是开放 API 定价，意味着单位推理成本仍然显著高于普通 GPU 推理服务，需要靠 Pro 订阅的高客单价摊薄 Cerebras 的硬件固定成本。如果 Cerebras 真的便宜过 GPU，OpenAI 应该敢把 Cerebras 后端的 API 价格直接公开。延迟披露 API 价格这件事，可以理解为 cost-per-token 还没真正击穿 GPU 推理的成本曲线。作为参考，Cerebras 自己的推理云上 gpt-oss-120B 定价是 $0.69/M 输出 tokens，GPT-4o 的 API 售价是 $10/M 输出 tokens，但两者模型规模和能力差距很大，不能直接对比。 但 Cerebras 的这套东西有几个关键的约束条件，框定了它现在能做什么、做不了什么。 第一是模型大小限制。Codex-Spark 是蒸馏后的小模型，被显著压缩才能跑进 WSE-3 的 SRAM。WSE-3 的片上 SRAM 只有 44GB，frontier model 的参数量远超这个容量，必须做剪枝或蒸馏才能装得下。换速度的代价是损失约 19 个百分点的模型能力，Terminal-Bench 2.0 上 Spark 约 58% vs 标准 Codex 77.3%。这是 weight streaming 在工程上仍然不够好的旁证，OpenAI 选择用蒸馏小模型而不用 weight streaming 跑大模型，说明前者目前在工程上更可行。 第二是容量限制。Codex-Spark 目前是 research preview，OpenAI 明确说明"由于运行在专用低延迟硬件上，使用受独立速率限制约束"。这个表述基本等同于承认 Cerebras 容量稀缺。750MW 全部部署完需要数千台 CS-3，对应 Cerebras 的整机产线是巨大的扩产挑战。这也是招股书里 RPO 约 246 亿美元有 85% 要到 2028 年之后才能确认收入的根本原因，硬件交付的物理速度限制了营收节奏。 总结来说，OpenAI 这笔交易对 Cerebras 是阶段性胜利，但远没到终局。技术验证层面，Codex-Spark 证明 WSE-3 在中小模型推理上确实能做到 GPU 做不到的速度，这一点产业内已经形成共识。但商业兑现层面，Cerebras 还要解两个核心问题。一是大模型推理的 SRAM 容量限制，OpenAI 用蒸馏小模型而不用 weight streaming 跑大模型，说明 weight streaming 当前还撑不起完整的 frontier model。二是规模化部署的速度，每台 CS-3 都要一片完整的 5nm 晶圆和一套非标准化的整机集成流程，年产几千台 CS-3 是相当具体的工程瓶颈。 对国产推理芯片的同行来说，这件事还有几个很具体的工程信号。 端到端延迟优化的价值正在被产业重估。OpenAI 重写整个 Responses API 说明硬件再快，软件栈跟不上的话推理体验仍然受限，专用推理芯片的价值要靠端到端的延迟优化才能完全释放。 蒸馏在产业部署里的实际重要性远超学术讨论的程度。OpenAI 这种规模都要为了跑 Cerebras 专门蒸馏模型，说明专用推理硬件 + 专用蒸馏模型这个组合是接下来一段时间的主流形态，单纯的"通用推理芯片"在竞争中会比较吃亏。 1000 tokens/秒正在成为 Agentic AI 工作流的新基准。慢于这个速度的推理硬件在交互式 agent 场景下会被边缘化，这个速度天花板对国产推理芯片是一个相当严峻的目标。 Cerebras 当前展示的能力是真实的，但商业化释放节奏被产能和工程瓶颈卡住了。研究 Cerebras 的真正关键，一是看它在 2027-2028 年能不能把 750MW 真的部署完，把 RPO 真的转成营收；二是看它能不能在 OpenAI 之外签下新的推理大客户。AWS 的 CS-3 上架是一个信号，但目前还没有第二个 OpenAI 量级的合同出现。从 G42 依赖到 OpenAI 依赖，客户质量在提升，但集中度的风险结构没有改变，只有客户组合真正分散了，估值里的风险溢价才能消化。

这就是东海边，真实的赶海风景。 承蒙关注，让我们在这个地球上有着一丝关系，十三妹已经在一批一批回关了。 This is the real scenery of the East Sea. With your attention, let us have a trace of relationship on this earth. The thirteenth sister has been returned one by one.

我进币圈前是做美股的，22年SpaceX好多员工股在卖，我什么币圈朋友和资源也没有，就帮卖SpaceX，想和大家交朋友。 结果币圈所有人都说太贵了，一毛也没有卖出去。 当时73b。 Before entering crypto, I came from the public equities world. In 2022, when many SpaceX employees were looking to sell their shares, I tried to help broker some secondary transactions. At the time, I had zero crypto network and was really just looking for ways to meet people and build relationships in the industry. Everyone in crypto thought SpaceX was ridiculously expensive, so I didn’t manage to sell a single share.

《无人着装 / Worn by No One》 币安九周年线上装置艺术展 An online installation exhibition for Binance's 9th Anniversary 这是一批被重新观看的币安官方周边。帽衫、球衣、围巾、包袋、鞋帽、运动装备……它们被从日常使用中抽离，置于白色展厅、绳索、冰、尘土、树脂与光线之间，成为一组关于物件、空间与关系的装置。 This exhibition reimagines official Binance merchandise. Hoodies, jerseys, scarves, bags, caps, shoes, and sports gear are removed from everyday use and placed into white spaces, ropes, ice, dust, resin, and light—becoming installations that invite a different way of seeing. 我们的大部分生活已经发生在线上。朋友、对话、工作、身份与信任，都停留在网页、软件、账户和屏幕之间。也因此，这些真实存在的衣服、包袋、帽子和道具，反而提醒我：身体、触感、空间，以及人与人的连接，依然值得被重新看见。 Today, much of our lives unfold online. As more of our relationships, identities, and memories move into digital spaces, these physical objects begin to carry a different meaning. They remind us that touch, space, and human presence still matter. 如果其中某一幅让你感到一点错愕、惊讶、不适、快乐，或者只是停留片刻，那也许就是它的价值。 If even one piece leaves you feeling surprised, unsettled, amused, or simply makes you pause for a moment, then perhaps it has already fulfilled its purpose. 线上展厅已上线 · The exhibition is now live 👉👉👉 @binance @binancezh @heyibinance @cz_binance

这是我和我老婆晚餐时候的对话， 你认为她说的对吗？ My Wife: You know, I've been reading about this Chinese concept of "face" lately. Me: Really? What brought that on? My Wife: Honestly? Living with a Chinese husband made me curious. Me: Uh oh. And what did you find? My Wife: Well, I used to think face just meant pride or reputation. But now I think it's much more complicated than that. Me: That's already better than most explanations. My Wife: What surprised me is that face isn't really something a person owns by themselves. It's something that exists between people. Me: Go on. My Wife: If I understand it correctly, giving someone face is a way of showing public respect, recognition, and consideration. And when you give someone face, there's an expectation that they'll give you face in return. Me: That's actually pretty accurate. My Wife: So face is reciprocal. It's not just about protecting your own image. It's about maintaining a relationship where both sides acknowledge each other's dignity. Me: Exactly. My Wife: And if one side refuses to give the other face, it's usually not seen as a small social mistake. It can be interpreted as disrespect, and sometimes even as hostility. Me: You're sounding more Chinese every minute. My Wife: Don't worry, I'm still American. I just finally realized that "face" isn't the same thing as self-esteem. It's more like a social obligation to recognize and respect the other person. Me: That's one of the best explanations I've heard from a non-Chinese person. My Wife: What really made it click for me is this: Americans often see respect as something personal. Chinese culture sees respect as something relational. Face is what helps keep those relationships in balance. Me: I couldn't have said it better myself. My Wife: So, to me, face isn't about ego. It's about reciprocity. It's the understanding that respect has to flow both ways. When people stop giving each other face, the relationship starts to break down.

和Quant Alex @StochAlex07 讨论： SABR Theta与Spot Theta+Vol Theta+Cross Theta的异同与应用，以及SABR模型自洽性分析。 **English Summary of the Chat** **SABR Theta vs Spot Theta + Vol Theta + Cross Theta** The conversation between **Alex Wu** (white bubbles) and **Jeff Liang** (green bubbles) is a technical discussion focused on **SABR Theta versus Total Theta** (i.e., Spot Theta + Cross Theta + Vol Theta), model self-consistency, PDE residual, and the correct definition of SABR Greeks. ### Key Points Discussed: 1. **SABR Gamma = Spot Gamma** (first major question, raised by Jeff) Jeff asked whether SABR Gamma (\(\partial^2 P / \partial F^2\)) is identical to Spot Gamma and whether it includes the dependence of \(\sigma_B\) on \(F\). He also provided the full chain-rule expansion of SABR Gamma in terms of Black-76 Greeks. Alex confirmed the understanding and **later affirmed in code** that this is exactly how SABR Gamma is implemented in their system. 2. **SABR Theta vs Total Theta and Model Self-Consistency** (main topic, led by Jeff) Jeff shared a clear 3-point understanding: - SABR Theta is computed directly via the SABR approximation formula to obtain \(\sigma_B\), then applying the Black-76 chain rule: \(\partial P/\partial t =\) BS_Theta(\(\sigma_B\)) + BS_Vega \(\cdot \partial\sigma_B/\partial t\). - Total Theta is the exact decomposition from the SABR PDE (Spot Theta + Cross Theta + Vol Theta). - When the model is **fully self-consistent** (Residual = \(\partial P/\partial t + \mathcal{L}P = 0\)), SABR Theta = Total Theta; otherwise the difference is the unexplained PnL caused by the approximation error in the Hagan formula (especially pronounced in long-dated, high vol-of-vol, or high-skew options). 3. **Practical Implication – Theta Decomposition Decision** (comment by Alex) Alex noted that whether to perform Theta decomposition depends on the risk-management approach: - Without decomposition → use SABR Gamma vs. dP/dt. - With decomposition → SABR Gamma maps to Spot Theta, Vanna to Cross Theta, and Volga to Vol Theta. **Overall Tone**: The discussion is highly technical and collaborative. Jeff drives the conversation by asking clarifying questions and presenting a well-structured 3-point summary of his recent study. Alex provides confirmations, practical insights, and code-level validation. Both participants demonstrate a strong command of SABR model nuances, particularly the relationship between approximation error, PDE residual, and real-world risk management.