Sean Zhou

[fin @fi56622380]

另外云厂host开源模型卖token，毛利是比直接卖GPU hours高一大截的

开源的兴起，对O和A的定价和毛利确实有一定影响，但对上游半导体来说基本是平替，除了deepseek，其他开源模型推理资源消耗并不少

这个趋势其实6月开始已经非常强劲了，开源模型ARR增速已经很惊人，但是目前被算力所限

以后O和A相当于是苹果A和苹果B，开源模型就是安卓，安卓还是要占大部分市场的

[Sean Zhou @seancz11]

@aleabitoreddit @twisterchessur CPO？华为半导体首席科学家廖恒：温度影响光介质折射率/光信号，所以CPO远不如NPO。对此你怎么看

[Sean Zhou @seancz11]

@jukan05 It's purely market behavior; no one is forcing you to buy with a gun.

[JangaFX Software @JangaFX]

A underwater peaceful vegetation setup, made in IlluGen.
The tiler's shapes are oriented to the UV direction and then stretched with an advect.

#IlluGen

[Sean Zhou @seancz11]

@jukan05 Apple and PrismML have pointed the way forward for on-device AI, and MediaTek will likely follow suit quickly.

[Xudong Han @Xudong07452910]

Claude Code 有个很值得养成的收尾习惯：

每次关掉一个长会话之前，先让它写「交接文档」。

直接输入：

「这个会话要结束了。请写一份交接文档存到 HANDOFF.md：我们在做什么任务、已经完成了什么、当前卡在哪、下一步计划是什么、有哪些踩过的坑绝对不要再踩。写给一个完全没有上下文的新会话看。」

下次开新会话，第一句话就是「先读 HANDOFF.md」。

这一步特别适合跨天推进的大任务。

因为 AI 的记忆以会话为界，你的项目以周为界，中间的落差，每次都靠你口头补，又累又漏。

聪明但每天失忆的员工，最需要的不是变得更聪明，是一本工作日志。

这其实是给 AI 建「跨会话记忆」的最低成本方案：不装任何插件，一个 md 文件就够。

[Sean Zhou @seancz11]

@jukan05 The demand gap is still huge, but Hyperscalers have to limit their Capex ambitions because of cash flow, so for now we just need to patiently wait for profitability on the application side to improve.

[Sean Zhou @seancz11]

@BadBrainCode @mlech26l @grok The twisted network cable slots on the network switch on the left reveal that this is an AI-generated image.

[Sean Zhou @seancz11]

@TheOmnius @jukan05 🤣

[Sean Zhou @seancz11]

@jukan05 This is old news from March. The launch of Gemini 3.5 Flash on May 19, 2026, significantly eased the compute capacity constraints. Meanwhile, the recent release of GLM 5.2 has been drawing even more user attention away.

[Sean Zhou @seancz11]

@EpicShaders It looks like High GI just adds a more aggressive layer of AO compared to New Medium GI.

[Sean Zhou @seancz11]

@jukan05 Can devices using CXMT memory be supplied to non-US markets? Especially devices produced for the Chinese market.

[vLLM @vllm_project]

🙏 Thanks to the @NVIDIAAI team for highlighting DFlash support on vLLM!

With DFlash speculative decoding, swapping EAGLE-3 for a DFlash checkpoint is a config-only change — no code edits needed.

It runs through the open-source Speculators library, which links the DFlash drafter to the target model's hidden states in the vLLM inference path.

On Gemma-4 31B on a single Blackwell Ultra GPU, this delivers up to 5.8x higher throughput at the same concurrency over autoregressive decoding:
🧮 Math500 — 5.8x
➕ GSM8K — 5.3x
💻 HumanEval — 5.6x
🐍 MBPP — 4.4x

Read the blog here! 👇

[Sean Zhou @seancz11]

@karminski3 有没有可能MacStudio卖二手还能赚一笔内存理财的钱🤣

[soli @solisolsoli]

Art by Adipocere

[Sean Zhou @seancz11]

@tphuang 本地建厂带动当地就业和税收，就像以往的合资丰田那样

[Sean Zhou @seancz11]

@NoamShazeer Excellent. Crisis averted—OpenAI’s poor performance was about to crash the market. Wall Street and the whole market will be thanking you. 😂

[Ning @totheagi]

We're the first to make the full GLM-5.2 (FP8) run on RTX 4090s.

GLM-5.2 is the new 753B SOTA open-weights model, and it officially ships for datacenter GPUs only: H100, H200, B200. We ported its sparse-attention kernel stack to consumer hardware.

A frontier open model, off the scarce GPUs and onto the abundant kind.

https://t.co/08rCk3b3h5

@Zai_org

[Sean Zhou @seancz11]

@jukan05 Two viewpoints both align with Cook's intentions, and they don't conflict (raising prices and prepaying to lock in capacity).

[Sean Zhou @seancz11]

@jukan05 半导体知名爆料人Jukan 👏