Joined September 2014
同一个词在不同人、不同团队语境下,可能也有微妙的意思差别。有时候人也不知道描述一个东西应该用什么准确术语,这时候就得跟Agent商量方案。Matt Skill很强调用经典术语,不仅是从标准角度考虑,也是为了充分激活模型的能力,相关文本叫leading words。也用到了ASD-STE100。
Matt 有几个方法论,到目前为止,我不太赞同 1. 不赞同和 AI 讨论好术语 其实我们要的并不是和 AI 讨论好术语,而是让 AI 不擅自发明新术语。商定好固定的术语在时间久了之后,依然会遗忘。所以更有价值的是标准术语。 所以我们在与 AI 沟通时应该强调使用 established industry standard terminologies 2. 不赞同通过 Design Tree 展开无休止的询问 以 Opus 5.5 为例,目前模型的世界知识以及跨领域运用世界知识的能力已经达到了空前的高度。除非你是某一领域拥有 10 年以上经验的专家,否则你做的决策大概率不会比 Opus 5.5 好。 所以我们要做的是从 top-down 层面描述好需求,让模型通过世界知识自行展开设计。而不是让模型去问你一些很细节的东西,反过来倒逼你学习这些细节相关的知识。
1
1
3
2,304
OrangeCLK retweeted
ByteDance’s overseas short drama app PineDrama has surpassed 100 million downloads in less than a year. The rapid growth of the Chinese-pioneered format is intensifying global competition and driving rival platforms to consolidate caixinglobal.com/2026-10-16/…
2
2
996
OrangeCLK retweeted
Authorities have taken away Huang Ranfei, son of former Chongqing Mayor Huang Qifan, who is currently being investigated by China’s corruption watchdog. The younger Huang previously worked at foreign investment banks, including Morgan Stanley. caixinglobal.com/2026-10-15/…
4
31
4
114
39,721
OrangeCLK retweeted
旷视十五周年,在旷视的六年真是一段难忘的经历。有个师弟比喻得好,“中国 AI 的仙童”。
74
1
1
116
38,736
OrangeCLK retweeted
This is the most important chart in AI right now because it shows where AI spending is actually going by model capability. Of Anthropic’s models, Fable (dark blue bottom) is only 11% of revenue. Very few companies want to pay for the best model. The majority of spending is going towards Opus 4.8/5. This means most businesses want very strong models but don’t feel it’s worth paying for the best models. Opus 4.8/5 is basically the performance of Chinese open-source models like Kimi K3 and GLM-5.3. But the Chinese models are cheaper to use. This means there’s likely very strong underlying demand for Chinese open-source models, but it takes time for US businesses to make the shift and some are worried about regulatory or other risks. Overall, there’s a big capabilities overhang in AI that makes it harder to justify massive investment in ever more powerful models from a sheer commercial standpoint. Businesses are still just trying to find interesting ways to use AI in general, much less push the limit with absolute frontier models. Chart source: ft.com/content/5ee49718-c258…
52
120
26
584
262,222
OrangeCLK retweeted
财新最新一期封面,报道了柬埔寨太子集团二号人物组建的如蛛网般复杂的国际化产业迷宫。让我想起以前的「三个互联网形态」的总结。太子集团这种,就是深藏地下的互联网形态。如今的报道,也只能是盲人摸象,不知道其背后还有多少没被掀开的内幕。
在中国,存在三个互联网形态。一种是媒体给人灌输的互联网,以海外 IPO 为目标的;一种是草根互联网,低调掘金,如迅猛龙般彪悍;一种是深藏地下的互联网。 @Fenng 写于 2012 。
9
70
33,177
Claude Code, OpenClaw, Codex, WorkBuddy这些更晚的产品,设计思想都比Manus更为落后。
Manus 从产品形态看基本就是朝着「终局」去的: 1. 本地无打扰完全运行在云上 2. 2025 年中引入 wide research 增加任务拓扑方向上智能调用密集度 3. 2025 年年底透露在做主动式 agent 拓展时间维度智能调用密集度 4. 很早就建立了 evaluation 团队 只能说季逸超牛逼。老天让这个团队运气再好点吧。
1
279
Manus领先太多
i think this is what tibo is talking about. openai bought ona to move codex from local machines into secure, persistent cloud environments. the announcement literally calls it “the next phase of codex” and says agents will keep working after your laptop is closed. tibo himself is quoted in that announcement. then look at the role openai is hiring for right now. “software engineer, cloud agents.” the team is building orchestration, sandboxing, storage, identity, observability and cost controls for long-running agents. the job says they will design systems for “orchestrating agents at scale” across codex, chatgpt and the api. so when tibo says codex will look primitive in 2 to 3 months because the next generation needs more than your laptop, i think he means one model will be able to create the cloud organization required to finish a goal. separate computers. separate agents. shared memory. tools. permissions. days of uninterrupted work. your laptop becomes the control screen for a temporary ai company. First image is ona second is the job.
1
81
OrangeCLK retweeted
做了开源几年,关了几万个 Issue,合了很多几万个 PR 之后我觉得我还是能说上两句的。目前开源社区的主要矛盾就是:无限扩张的 AI PR 供给,同开源维护者有限而宝贵的注意力之间的矛盾。
51
4
67
8,513
Loop这个想法是经不起推敲的
/loop简单说就是把AI拉的shi,再让AI吃回去。指望多拉几轮后shi变好吃 但是shi永远是shi,就像说出这句话的人的开源项目openclaw一样,又臭又大
1
1
69
Caixin's @caixin piece on this is worth reading slowly, and to my mind one of the best explanations of the new rule to come out of China. Setting the examples aside, here is what is most worth watching in the regulation itself, including the parts that are easy to miss. 1. It is China's first dedicated, top-level law on outbound investment. Rules that were scattered across several ministries are now consolidated into a single administrative regulation with real legal force, which lifts the whole area from departmental practice to national law. This is the step a maturing system takes once the activity it governs grows large enough to need one clear rulebook. 2. The framework now covers individuals, not only companies. Personal offshore investment, and the red-chip and VIE structures behind many Chinese tech firms, used to be handled piecemeal, mostly through foreign-exchange registration. Bringing them into one framework closes a long-standing gap and gives investors a single set of rules to follow. 3. It covers technology, not only money and equity. Licensing core technology, sending engineers, giving cross-border technical guidance, training staff abroad and building offshore R&D can now be assessed. The aim is to keep controlled key technology from being moved out through indirect routes that the old, equity-focused rules never reached. 4. National security review becomes a standalone safeguard. It used to be one consideration folded into the approval; now it is a dedicated review run jointly by several ministries, so security questions are weighed on their own terms. For projects touching key technology, important data, strategic resources or critical industries, that review applies in its own right. 5. Oversight now runs the full length of an investment, before, during and after, and it follows the money through. It also covers reinvestment by entities already based offshore, which closes the gap where setting up abroad first used to sit outside the rules. The point is a complete framework with no blind spots. 6. Its center of gravity is protective and inward-facing. Where American-style review is mainly about cutting off money to a rival's industries, this rule is more about safeguarding China's own strengths, rare earths, battery technology, manufacturing know-how, from leaving the country. Protecting core capabilities that took decades to build is something every major economy does. 7. It includes countermeasures, and they are defensive by design. The rule connects to China's anti-sanctions and foreign-trade laws, giving Beijing a clear legal basis to respond when its companies face discriminatory or extraterritorial measures abroad. It also lines outbound investment up with export controls and data security, so the parts of the system work together. 8. A good deal of the detail is still to come, and that is deliberate. The catalogs of encouraged, restricted and prohibited industries, the specific rules for individuals, and the security-review procedures will follow. The framework and its direction are set, and the operational detail is being built on top of it. One last thing. It is easy to read this as a wall going up. The more accurate read is that China, now the world's third-largest outbound investor, is doing what every major economy has already done, putting clear and unified rules around capital that used to move under scattered departmental notices. Openness on this scale needs a framework to hold its shape, and that is what this rule builds. #China #USChina #Trade #ODI #SupplyChain
2
10
54
6,309