maxime_zhang

ex-amazonian, amateur investor, touring China with me
visit China easily with

[kejun @kejunz]

传统 UI 注定会被「生成式 UI」取代，只不过它不是消息流中的一张「卡片」，这两天做了一个探索（数据都是Mock的）

1. 交互模式：用户的顺序式操作 / 固定的界面 → Agent 驱动的事件流，界面按需生成 / 销毁
2. 状态管理：前端单向状态管理 / 后端仅持久化 → 前端与 Agent 共享状态，双向同步且可冲突调和
3. 输出方式：静态页面 + 局部异步刷新 → 动态组件树、流式语音/视频、思考链可视化、可中断提示
4. 任务模式：规定流程，用户必须遍历步骤 → AI 自主规划子任务、多步执行与人机中断/回退
5. 可观测性：仅记录用户行为→ 全面记录 Agent 思考、规划、工具调用、审批链，支持审计与重放
6. 安全模式：基于用户权限的接口调用 → 基于 Agent 能力边界的沙盒渲染与操作指令校验

[Arnaud Bertrand @RnaudBertrand]

This is a really fascinating paper that everyone interested in China's industrial policy should read.

It destroys so many myths (see below), and is written by deeply credible people who conducted over three years of fieldwork in China and interviewed 60+ Chinese officials, entrepreneurs, and engineers. When it comes to China studies, it literally doesn't get more rigorous than this.

First myth it destroys: contrary to popular belief, Beijing's industrial policy didn't build the companies that became China's EV champions. They rose largely **despite** it, through its cracks.

For sure, Beijing did favor EVs as an industry and pushed hard for it but their big bet was SOEs (State Owned Enterprises): research grants, pilot programs, licenses, cheap credit - virtually all of it flowed to state firms.

The result? China's actual EV champions - BYD, Geely, NIO, XPeng, Li Auto, etc. - are overwhelmingly private firms that succeeded despite Beijing favoring their SOE competitors.

How so? Because, when favoring SOEs, the central government didn't just pick winning companies, it picked winning cities, each SOE being anchored in a specific city: Shanghai (SAIC), Changchun (FAW), Wuhan-Shiyan (Dongfeng), etc.

Which means that every city not on the list, that wanted a piece of the auto boom, had only one option left: team up with private entrepreneurs who were equally excluded from central government favor.

That's what truly fueled China's EV miracle: an alliance of the excluded, between local private entrepreneurs and local mayors.

This is the biggest misconception this paper destroys: the reality is that the "Chinese state capitalism" that many in the West think powered the EV boom actually tried to block many of these companies from existing. In effect, it was closer to an obstacle course that local actors (mayors and provinces) learned to game.

Geely - now the third largest automaker in China - is a fantastic example of this.

First of all, it started off illegal since, to build passenger cars, you had to have a central government license and they couldn't get one. Zhejiang Province told them to go ahead regardless because the province had hundreds of auto parts suppliers but no carmaker of its own.

It's only a couple of years later, recognizing the fait-accompli that Geely was producing cars and was competitive, that the central government admitted them to the National Sedan Catalog - effectively legalizing them retroactively because there were facts on the ground.

Then there was the Volvo acquisition in 2010, which is fair to say - looking back - proved to be the most strategically valuable acquisition in Chinese automotive history. Despite it being presented at the time (and still described this way today) as "China buying Volvo", all 3 major state-backed banks in China (Export-Import Bank, China Development Bank, Bank of China) refused to finance the deal. The only state-bank money Geely managed to get was a $200 million loan from a provincial branch of China Construction Bank - a tiny fraction of what the deal required.

Geely actually did the deal with Goldman Sachs money via Hong Kong plus loans and equity from four local governments (Chengdu, Zhangjiakou, Daqing, Shanghai's Jiading district), each of which bought in by securing a Volvo plant or headquarters for itself.

In effect, the doors that Beijing controlled were largely closed to Geely, but it made it because the doors subnational actors controlled were opened.

Which all means this paper destroys another very common myth: the big merit of the central government in all this was to be relatively chill about it, to NOT be dictatorial.

I just imagine if that had happened in France and you had - say - the mayor of Lyon or Marseilles open, fund and promote an unlicensed carmaker against Renault: the préfet would shut it down within weeks, and the mayor would be lucky to escape prosecution.

That's the irony: on industrial policy, the supposedly "totalitarian" Chinese state proved more tolerant of local defiance than most Western liberal democracies would be. Beijing's greatest contribution to the EV miracle wasn't the plan - it was looking the other way while the plan was being violated.

To be sure, the paper doesn't hide the costs of this system: ferocious local competition also produced what's known today in China as "involution" (内卷-Neijuan, basically a hypercompetitive price war), as well as some spectacular failures. For instance one county lost 6.6 billion yuan on a carmaker that never really made cars.

But that's precisely the point: this is a high-risk, high-reward model of decentralized experimentation, the very opposite of the careful central planning Westerners imagine.

I've repeated this countless times but it bears repeating again: the single greatest misconception people have about China is - probably because we wrongly associate communism with centralized control - that it is a monolith run from Beijing. Some even say it's run by "one man."

The reality is the exact opposite: China is, in practice, one of the most decentralized countries on earth. Roughly 85% of government spending in China happens at the subnational level - against about 30% in the average OECD country (and even less in France, which is actually one of the most centrally controlled countries on earth). A Chinese mayor commands fiscal resources, land, investment funds and policy latitude that virtually no Western mayor could dream of.

Last but not least, I'd be remiss not to mention what the paper has to say on the positive legacy of Mao and its role in the rise of EVs (given I myself wrote an article titled "Mao's economic record wasn't bad, actually": https://t.co/1NZgHqBHwg).

When it comes to China myths, none is more entrenched than the idea that Mao left behind nothing but ruins.

This paper confirms a key argument of my article: Mao's deliberate dispersal of industry across China (during the Great Leap Forward and Cultural Revolution decentralizations) left dozens of cities with their own small auto works. Inefficient, yes - but these scattered factories survived into the 1990s and became the seed stock of everything that followed: the industrial base, the engineers, and the production licenses that EV startups would use to enter the market.

The paper even says it outright: the fragmentation that industrial policy "sought to eradicate" is "precisely" what "ironically enabled" the EV sector's rapid rise.

This is exactly the mechanism I described in my Mao article: structures built in the Mao era - communes becoming township governments, commune enterprises becoming TVEs, Third Front factories seeding interior industrialization - became load-bearing foundations of the reform miracle.

Fittingly, the spark for China's first municipal carmaker adventure was literally a TVE (Township and Village Enterprise), the institutional descendants of Mao's commune enterprises: Tongbao, a kit-car maker in Wuhu whose success stunned local officials into building what became Chery (one of China's biggest carmakers today). You can't tell the story of China's EV miracle without crediting the legacy of Mao.

What's the biggest lesson in all this for Western policymakers?

The obvious one is that the part of industrial policy that most people assume China does and that they sometimes want to copy - i.e. the state picking winners - is actually the part that failed.

The part that did succeed is the China nobody in the West believes exists: a radically decentralized system with a high degree of tolerance for disobedience and experimentation.

We imagine China as a country where nothing happens without Beijing's approval when the reality is closer to the opposite: China's EV miracle happened precisely because localities asked for forgiveness rather than permission.

All in all, and this is the lesson I often come back to, this is yet another illustration of the importance of understanding China for what it is as opposed to the caricature we've built of it. This matters whichever "camp" you're in. If you see China as a rival, you can't compete with someone you don't understand. If you see them as a source of lessons, you can't emulate what you've misunderstood. Whatever you want from China - to compete with it or learn from it - the entry fee is the same: genuinely understanding it.

[Amelia @queerzoi]

What chinese apps look like to a westerner

[maxime_zhang @maximezy]

@JunaidAckroyd Let’s see

[maxime_zhang @maximezy]

@TheHumanoidHub Welcome!

[maxime_zhang @maximezy]

@thalogiguy Let’s hang

[maxime_zhang @maximezy]

@nathanlawkc Debating freedom 😂😂😂you are patheticly funny

[Vincent | 信号＞噪音 @VincentLogic]

刚看完这个访谈，既被治愈了，也有点危机感。

这个山东妹子曲晓音太猛了：27 岁，斯坦福辍学，前 Facebook PM，现在创业还拿了 OpenAI 和 a16z 的投资。

她有个观点特别戳我：

AI 时代工作不会变少，反而会变多。

但给“老登公司”打工的机会会变少。

以后每个人都要把自己当成一个“一人公司”来经营，做通才比做螺丝钉重要太多了。

不管你是想创业、想转型，还是只是怕被裁员，这期都值得看。

很多话听着扎心，但确实是这个时代的真实变化。👇

[Xiao Ma @maxiaoalex]

🚗🔋 Many think Beijing masterfully planned China's EV takeover. Fengming Lu (@ANUBellSchool ) and I spent 3 years and 60+ interviews finding out what actually happened in our latest article @TheChinaJournal. A thread 🧵

[diva @divaagurlxw]

As an AI Engineer. Please learn

>Harness engineering, not just prompt engineering

>Context engineering, not just long prompts

>Prompt caching vs. semantic caching tradeoffs

>KV cache management, eviction, reuse, and memory pressure at scale

>Prefill vs. decode latency and why they optimize differently

>Continuous batching, paged attention, and throughput optimization

>Speculative decoding vs. quantization vs. distillation tradeoffs

>INT8, INT4, FP8, AWQ, GPTQ, and when quantization hurts quality

>Structured output failures, schema validation, repair loops, and fallback chains

>Function calling reliability, tool contracts, argument validation, and idempotency

>Agent guardrails, loop budgets, tool budgets, and termination conditions

>Model routing, graceful fallback logic, and degraded-mode UX

>RAG architecture: chunking, embeddings, hybrid search, reranking, and freshness

>Retrieval evals: recall, precision, grounding, attribution, and citation quality

>Evals: golden sets, regression tests, adversarial tests, LLM-as-judge, and human evals

>LLM observability as a first-class discipline: traces, spans, tokens, latency, errors, and drift

>Cost attribution per feature, workflow, tenant, and user journey not just per model

>Safety engineering: prompt injection defense, data leakage prevention, and permission boundaries

>Multi-tenant isolation, cache safety, and cross-user context contamination prevention

>Fine-tuning vs. in-context learning vs. RAG vs. distillation and when each is the wrong tool

>Latency, quality, cost, and reliability tradeoffs across the full inference stack

>Production failure modes: hallucinated tool calls, malformed JSON, stale retrieval, runaway agents, and silent eval regressions

[maxime_zhang @maximezy]

Is Australia more Sinophobia or Europe? My feelings are the liberals are the most Sinophobic and Europeans are mostly liberals

[Andrej Karpathy @karpathy]

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

[xIA @xiathis]

En vez de 2 horas de Netflix esta noche, mira esta clase magistral de 40 min del fundador de una empresa china de IA valorada en más de $20B

La explicación más clara que he visto sobre enjambres de agentes y sistemas de IA a gran escala.

[maxime_zhang @maximezy]

Only self hate people would visit Japan

[Masa @masanews3]

迷惑系中国人、コンビニの前でキショいダンスを披露😱🦁

コンビニに入れない！邪魔すぎる！と批判殺到😱🦁

[maxime_zhang @maximezy]

@AryaAphrodi Nobody respects India when Indians can’t even respect honesty

[maxime_zhang @maximezy]

Can’t believe so many Indians got mad under this honest post lol

[Options.India @Options_IndiaAB]

I recently spent 2 weeks in China.

6 cities: Shanghai, Beijing, Xi’an, Zhangjiajie, Chongqing and Chengdu.

I went there with curiosity.

Like many Indians, I had heard a lot about China through media, social media and conversations. I expected to see progress, maybe discover some business ideas, and understand what the country is actually building.

I came back with a very uncomfortable feeling.

Not because I found a business idea for myself.

But because I saw 100 things that governments can do when infrastructure, tourism, transport, urban planning and civic systems are treated seriously.

I travelled within China by flights, trains, cars and local transport. The infrastructure was honestly stunning.

Clean cities. Smooth roads. High-speed trains. Well-managed traffic. Public spaces that actually feel designed for people. Tourist destinations that are built, maintained and promoted like national assets.

And then I kept thinking about India.

We keep comparing ourselves to China. Our media keeps telling us how India is catching up, how China is restrictive, how we are better in so many ways.

After spending time there and speaking to people, I realised how much of that narrative is just comfort food.

China is not perfect. No country is.

But on infrastructure, execution, tourism, civic discipline and quality of urban life, they are not 5 years ahead of us.

They are decades ahead.

The saddest part for me was the currency.

Everything felt expensive. Not because China was insanely expensive, but because the rupee has weakened so much that even normal spending starts feeling heavy. As an Indian taxpayer, that genuinely hurt.

We pay taxes. We work hard. We talk about becoming a global power.

But where is the quality of life?

Where is the civic sense?

Where is the infrastructure that makes daily life easier?

Where is the tourism vision beyond religious tourism?

I met travellers from other countries who were excited to visit China because they wanted to see its progress. When I asked about India, many had no real desire to visit. Not out of hate. India simply was not on their aspirational travel list.

That should bother us.

Even the so-called “closed internet” surprised me. We are told people there are missing out because they don’t use Google, Instagram, WhatsApp or Facebook.

But China has built its own digital ecosystem. Payments, maps, transport, messaging, shopping, everything works inside their own infrastructure. People did not seem to feel deprived. They seemed adapted.

Again, this is not a hate post.

I love India. That is exactly why this trip bothered me.

Patriotism cannot only be about saying we are great.

Real patriotism is having the courage to admit where we are falling behind.

China made me realise one thing very clearly:

India’s potential is not the problem.

Execution is.

And unless we stop comforting ourselves with comparisons and start demanding better infrastructure, better governance, better tourism, cleaner cities and a higher quality of life, we will keep celebrating the idea of progress instead of actually living it.

[Max For AI @MaxForAI]

https://t.co/81hzTi5U9T

[AB Kuai.Dong @_FORAB]

彭博今天发了篇大长文，终于解析了中国为什么要收紧去海外炒股，原因是有大量的非法资金外流，其中炒美港股是重灾区。

文章指出，有大量的投资者，用每年 5 万美元换汇额度，以旅游、留学名义，换完汇后，把钱转到富途、老虎这类 App 里买海外股票。

如果金额更大，有些人会用场外交易，在国内把人民币交给中间人，中间人在海外，让另一个人给你对应的外币。

另外，还有一种方法是买香港保险，比如用人民币买香港保险，然后退保，拿到外币退款。

所以这波监管收紧，不光只是媒体们猜测的限制开户，还将涉及整条业务链，仅仅是个开始。

中信证券估计，内地投资者通过富途和老虎，持有的香港资产大约有 2000 亿到 2500 亿港元，所以中国不仅是在打击跨境券商，也是在系统性收紧内地资金。

未来绕道出海炒股的灰色通道，将会持续处于高压状态。

[Kirill @kirillk_web3]

instead of watching 2 hours of Netflix tonight, watch this 40-minute masterclass from the founder of a $20B China AI company

it's the clearest explanation I've seen of how Agent Swarms and AI systems actually work at scale

useful whether you've never built an agent in your life or have been using Claude every day for the past year

I took the key ideas and turned them into a practical guide on how to actually build with Kimi

find it below