听大内密谈聊孙宇晨，把他的发家过程顺了一遍。十年里他做的是同一件事，先拿一个名头，把人聚过来，再把这批人换成钱。

codex 里的 agent 之前全跑 Astra/high，用量烧得很快。GPT-6 的 Sol 和 Luna 出来以后， 我拿七种模型和 effort 组合在真实 repo 上跑了三轮：自己的 monorepo、工作上三个 repo 的跨项目题、 还有盲测复现历史 PR。代码能写对的比想象中多，但在 agent 自己的集成 runner 里把真实调用路径也跑通的少很多， 好几份自测全绿的提交其实没修好。文末是最后定下来的分工表和几个派活规则。

opus 5.5 出来当天，我拿自己 11 个 repo 里的 22 个真实提交做了一轮 medium 对 high 的配对测试： 两档完成度一样，high 慢三分之一、多花三成多 token，但盲评里多抓出几个测试查不出来的真 bug。 再对一下 Artificial Analysis 的 Intelligence Index 数据，最后把默认档定成了 high。 文末按不同用法给了推荐。

Every Codex agent I ran was on Astra/high, and usage burned fast. When GPT-6 Sol and Luna landed, I had seven model and effort combinations run through three rounds on real repos: my own monorepo, cross-project tasks in three work repos, and blind replays of historical PRs. More routes wrote correct code than I expected; far fewer also got the real caller path passing in their own integration runner, and several submissions with green self-tests weren't actually fixed. The final routing table and delegation rules are at the end.

On launch day I replayed 22 real commits from 11 of my own repos, running Opus 5.5 at medium and high effort side by side. Both finished every task; high took about a third longer and 35% more tokens, but blind judges credited it with catching several bugs the tests couldn't see. Cross-checked against the Artificial Analysis Intelligence Index, I made high my default. Recommendations by usage profile at the end.

Reading Anthropic's September misuse report through the economics of agent workflows, the limits of account bans, and the gap between observed activity and proven harm.

An analysis of an AI-energy podcast: the value of getting power on time, the limits of flexible computing, and who carries the risk when infrastructure outlasts demand.

读 Anthropic 九月滥用报告，聊 agent 为什么能让小团队持续做复杂的坏事，以及模型拒绝、人工审批和账号封禁各自能管到哪里。

从《十分吸引》的电力访谈出发，聊按时通电为什么变贵，异步 agent 能给电网多少弹性，以及谁承担基础设施等不到客户的风险。

From a glass fiber pulled over an alcohol burner in 1976 to seven of the world's top ten module vendors, China built its optical industry in reverse order — cable first, then equipment, then modules, and only now the chips at the top. The module business is won; the high-end silicon above it still isn't.

Pluggable, LPO, NPO, CPO, OIO — five packaging schemes that are really one siting problem, how close to the office tower to build the courier depot. The optical DSP is the module's brain and its biggest power draw, and it sits at the center of the fight.

Kevin compresses nearly two centuries of wired communication into three shifts. Copper losing to glass was settled by physics; the fifty years since have been the boundary between them moving toward the chip, one step per speed generation.

Networking has its own Moore's law, and AI's appetite for bandwidth outruns it. Copper's loss climbs with frequency while fiber's loss ignores speed but charges a fixed conversion toll — the two cost lines cross, and every generation the crossover slides toward shorter distances.

The blinking box on your wall is a slow optical module. The 800G ones filling switch faceplates in an AI cluster are the same symmetric signal chain, just run a few dozen times faster.

从赵梓森用酒精喷灯拉出第一根光纤，到光模块全球前十占七席，中国是先做光纤光缆、再做设备、再做模块，如今在攻最上游的光电芯片。模块这块做起来了，最上游的高端芯片还没做下来。

可插拔、LPO、NPO、CPO、OIO，五种封装方案其实是同一道选址题，快递站建在离写字楼多近。光模块的大脑 ODSP 是功耗大户，也是这场仗的争议中心。

Kevin 把一百多年通信史压成三次跃迁，铜到光的胜负是物理性质写定的，之后五十年光的边界一直往芯片端推，推到今天的机柜边上。

通信界也有自己的摩尔定律，AI 的带宽需求比它还快。铜线的损耗跟着频率往上飙，光纤的损耗和速率没关系但要交一笔固定的转换费，两条线一交叉，光模块要占的就是交叉点右边那段。

你家墙上一闪一闪的光猫就是一只低速光模块。从光猫到数据中心里插满面板的 800G 模块，拆开看是同一条对称的信号链。

和朋友争论 Mio 该走疗愈树洞还是恋爱模拟,争到最后发现两条路都不对。市场数据说这个品类的用户和角色平均只处 5 到 7 天,差评第一名是失忆。我觉得毛病出在关系是单方面的。这一篇讲 Mio 的新方向,双向奔赴,附一份完整的产品主线 deck。