where zichang writes

1. Apollo published a blog post [AI is not driving productivity outside of the tech sector] noting that, four years past the ChatGPT moment, there is still no evidence that AI is driving productivity across the economy. No noticeable increase in value-add measured by GDP, nor the profit margin improving.
2. Other sources make similar observations; see the St. Louis Fed post [here] and an MIT study [here].
3. This aligns with my experience implementing generative AI applications at one of the largest US banks. While AI has made many coding tasks much easier, the way day-to-day work gets done hasn’t fundamentally changed yet.
4. Why aren’t we seeing the same productivity boost from AI in other areas that we’ve seen in coding?
5. The first reason is the interface. Chat interfaces are far removed from how work is usually done, and the terminal isn’t friendly to a large share of human workers.
6. The second reason is that humans have a great deal of knowledge and context, built up over years of experience, that isn’t yet readily available to AI. In the compliance space where I work, veterans know the nuances and edge cases, and that knowledge hasn’t yet been fed to AI in a way that makes it useful. This may be solvable with the right interface, since this knowledge can technically be captured and used to train specialized AI.
7. The third reason is that AI fundamentally doesn’t yet have the accountability or trust-earning ability of humans. AI also still can’t sense the politics, read the room, and contribute to a meeting the way a person does. Maybe that’s what Anthropic wants to build with its “virtual colleague,” but such colleagues will still need to earn people’s trust in human ways.
8. Right now, it looks more like a very impressive intern. 
9. I’ve also realized that jobs are much more than a collection of tasks. With the right combination of task design, model capability, and interface (how AI reads data, stores memory, and takes actions), a growing share of those tasks may become doable.
10. When I was at the airport in Washington, D.C., watching people eating and drinking at the gates, I realized that commercial air travel didn’t exist until the 20th century. Society reorganized itself to make it work: building airports, creating rules, regulating airspace, and so on. Maybe society and organizations will need to reinvent themselves to make the most of this technology.
11. It probably doesn’t depend solely on building out data centers, but also on rethinking how work gets defined and done to give the models some airspace.

Oct 11 edits: Financial times reported that China, is similarly seeing high central debt to GDP rate. The tax revenue is going to be slower because of low domestic spend. If the engines of the world, US, Japan, China, EU are all entering a stage with debt issues: how will the world get out of this? Can we "grow out of the debt", like Bessent said?

I read about WSJ report on a looming French debt issue, and decided to look deeper into this problem. The numbers research are assisted by AI. This is, at its best, just an exercise jogging down raw numbers and trying to understand the problem of government debt, using France as an example. 

Size of the problem

1. The magnitude of French government bonds maturing on 2030 tops 285b EUR. It will rise sharply from 60b EUR in 2026 to 203.7b EUR in 2027 [4]. Totaling 1 trillion USD in debt due by 2030. [1] [4]
2. A deeper look into French income statements [2]

• 2a: Revenue by categories (in 2025). 

1. Total = 1,561B EUR.
2. Taxes = 882B EUR (57%)
3. Social contributions = 446B (29%).
4. Sales and other production revenues = 130B (8.4%). 
5. Property income = 23.4 (1.5%). 
6. Other revenue = 86.1B (5.5%)

• 2b: Spending by categories (in 2024)  [3]

1. Total = 1,672B EUR. 
2. Social benefits = 694B (41%, pension accounting for 394B)
3. Health care = 261B (15.6%)
4. General public services and debt service = 181B (10.9% in 2024, 64.7B dedicated to debt service in 2025) 
5. Economic affairs=166B (6.7%)
6. Education = 149B (8.9%) 
7. Defense = 54B (3.2%)
8. Public safety = 52B (3.1%)

• 2c. Who owns the government debts? [4]

1. Non-resident investors: 57.5% (rising from 50% from December 2022)
2. French “others”: 20.6% (Banque de France, ECB bond-buying programmes, takes ~18%)
3. French banks: 10.5%
4. French insurers: 9.6%
5. French investment funds: 1.8%

1. continuous crises (pandemic, the war in Ukraine) increased government spending.
2. The parliament is not set up to reduce government spending - politically painful and dysfunctional. 
3. The economy has been slow - lower tax revenue than expected.

1. There will be a 59% surge in debt service by 2030. The debt to GDP ratio is 119.0% (€3,595.5bn) at end-Q2 2026. [5]

1. Cut spending: will the government pull this off?
2. Increase tax revenue: will the economy be stronger to increase the tax base?
3. Default: will this happen, under the framework of ECB?
4. Bail out from European Central Bank

Source: 

[1] Wall Street Journal 'After Years of ‘Magic Money,’ France Becomes a Debt Bomb'.

[2] INSEE https://www.insee.fr/fr/statistiques/8956575

[3] INSEE  https://www.insee.fr/fr/statistiques/8735252

[4] AFT Monthly Bulletin, August 2026 (holders, debt stock, redemption schedule)

[5] INSEE https://www.insee.fr/fr/statistiques/9053525

1. 人类似乎总是担心自己会被无限的事物碾压取代。Navier-Stokes Equations被 [OpenAI] 解决了，大家以后或无事可做?
2. The advantage of AI is in obviously better compute, and near-infinite parallelization 
3. Parallelization is key - it shortens the time epoch needed to solve a problem. It is the hidden gift of time - there is almost no restriction in parallelizing search - the challenge is how to manage these parallel processes. Map Reduce solved this issue for mass compute. OpenAI claims the problem is solved by 10k+ agents, it would be a curious questions on how the agents are curated and organized. What’s the solution for AI agents? 
4. 苏轼在[《赤壁赋》]里描述了同样的焦虑：“哀吾生之须臾，羡长江之无穷。” ——人类相比天地，相比世间许多东西，本身就很小，自古如是。
5. 人类5000年前就建了金字塔。汽车发明了，没有人跑得过；起重机发明了，没有人比它强；原子弹出现了，人类无法控制的力量。但是卡车司机出现了，起重机师傅出现了，核不扩散的国际秩序出现了。
6. 有理由相信人类会继续向前走。although a bright future is not guaranteed.
7. 苏轼通过感知世界，开解了对无穷的焦虑：“惟江上之清风，与山间之明月，耳得之而为声，目遇之而成色，取之无禁，用之不竭。是造物者之无尽藏也。“
8. 对于AI通过parallelization所能遍历的近乎无限的时间，也许人类可以御风而行：告诉AI什么是对的，什么问题，值得找答案。
   

1. so far enterprise AI projects I worked on are generation: given X, output Y. 

2. to make the predicted Y better, prompt engineering and context engineering are performed. this spans across projects with evidence extraction, tags generation, and recommendation.

3. the next step is execution: given a process P, deliver outcome O. This is basically the agent systems. 

• How do we build enterprise-level agent systems? 
• What are the right workflows to target? 
• How do we define that workflow, executable by AI, and make sure guardrails and safety nets are in place?

4. As I mull over the next project to work on, agent systems seems like an interesting candidate.

体会未见的生命。可以看到生命的另一种可能性，历史的，虚构的，超脱于现实生活的局限。

理解现实的视角。看到社会上的人，第一层反应之后，自然将其回归到小说里刻画过的某一类角色里：鲁迅的孔乙己，王小波的领导和人肉人瑞。于此，便有抽离，获得一些广泛的视角。可以看到到达他们现在人格的草蛇灰线，不至于不理解。

1. New technologies will not be uninvented - generative AI will have a place in corporate OPEX. There are projects where it is rather cost-effective to justify the incrementally adapting AI can bring. 
2. But it won’t justify the tokenmaxxing strategy, indefinitely. It simply doesn’t make financial sense - not all tasks are cost effective to use the SOTA model on it.
3. In a lot of ways, LLMs right now are like expensive Swiss knives. They are great, but question mark on whether everyone in the company needs it.
4. A justification for tokenmaxxing strategy, is that it serves as a priority discovery tool. People don’t know what will be the most productive use case of AI, and would be worried that early curtails hurts potential innovations. So they let the puppy fly for a while, collect the data and estimate the distribution of values/cost. Then finance will come in, and there will be cost management coming in.
5. The market will find an equilibrium. But the AI Capex, building of data centers, will take years to complete - and in cases already facing delays and challenges. https://www.wsj.com/tech/ai/americas-data-center-build-out-is-falling-way-behind-schedule-e408a9a8
6. When these projects are not yielding cash flows early enough, what would happen? Things will turn ugly, pressures will come in, cost cutting underway, assets sales may happen to raise liquidity, etc.
7. Adaptation of innovations like this, however, is interesting that it is relevant to everyone, not just pockets of work force. The last innovation like this is probably PC, or internet. 
8. Prudence. The best is yet to come.

1. Principle is, the economics of AI investment need to make sense. ROI needs to make sense.
2. I am seeing equilibrium being established.
3. On one end, running inferences, even large scale, is relatively cheap, it is a cost that corporates should be able to foot. 
4. One the other hand, “token-maxxing” strategy is raising concerns and make CFOs questioning their ROI: Uber already burnt their annual budget for AI in April, anecdotes coming everywhere that they can’t get enough of their tokens to complete their jobs (/ambitions/ideas).
5. But eventually, these AI projects need to be valuable products. Need to contribute. Otherwise, it would be an expensive hackathon.
6. So an equilibrium is being figured out - we can pay for it, but not infinitely. 
7. Reminded me of the case of dark fiber. In the early ages where fiber was introduced to the world, there was a avid investment in building out fiber networks throughout the country. Not all of them, are commercially viable to be “lit up”. As such, there are miles and miles of built-and-ready sleeping beneath the ground, waited to be waken by new players willing to pick them up.
8. Maybe compute will be like that. Yes, it will not be “uninvented”, demand for HPC will always be there. But there may be excess.
9. Will these “dark HPC” be relevant and usable in 20 years? That, we don’t know.  My understanding is likely not. 
10. If they will somehow still be in the game in 20 years, there will be a market for it.  

我在火车上坐下，看见身旁的女人已经在喝第二瓶香槟。好奇顿起。

原来她是要北方的城市见一位朋友。“我已经二十年没约会过了，”她说。在女儿上大学、自己刚完成了离婚之后，她开始了新的生活。

“今早我六点就起床了。”她有点紧张，又抿了一口香槟。

“你看着很棒，别担心，”我说，“享受今晚就好了！”

人的感受、害羞、紧张，不被时间磨去。

• Went to the “Raphael: Sublime Poetry” at Met today. The exhibition started with his father and his masters work - you can see his years of pupil, and what led to his style and crafts.

• He used rulers and compasses as tools to design his composition - there are solid methodical approaches backing the mastery and genius of composition - a theme that recurs. You can see his precise measurement in his sketches. Impressive and maybe a bit demystifying that there are traces of engineering in perfection in beauty. 

  

  

  

• However, rulers and compasses are just technique - but the balance and harmony are truly genius.

  

  

  

• very methodical - he would start with simple red chalks for sketches and build from there - but even simple sketches are captivating.

  

  

  
• Seeing his finished work reminds me seeing the first retina screens from Apple. It was meticulously complete. 

  

  

• He was a very prolific painter.  He started with a ferocious, ambitious, and extremely diligent study of the past masters, and later the raining orders from his patrons including the Popes. The sheer magnitude of his work was foundational of his impact. It was impressive to see how greatness unfolded and were practiced, and humbling to be reminded that diligence was the foundation of humankind’s finest. 
• He seems to be a humble student from the greats of his time, adapting his work and practices with advices from Leo Da Vinci and studies from Michelangelo (who were more direct rivalry of Raphael). 
• The surging projects from the Palace and the Pope - albeit led to the bankruptcy of the papacy of Leo X eventually - may signal the splurging activity and power of the renaissance age.
• His success, and managing to replace an earlier generation of painters at the Vatican palace involved his artistic virtuosity, but his personal charisma, savvy, and political support from his friends from Urbivo played a key role in earning the keen of the Pope. Raphael himself never shies away from getting in with the culture - in his sketches you can see lines of sonnet reflecting his attempts to rhyme with the poetry making culture of the court at that time.
• Renaissance artists were supported by patrons - thus their ability to serve different clientele is crucial. Raphael’s style and emphasis changed from Julius II to Leo X - partly because of the different preferences of the clients. 
• Reputation and impacts are an results of timing, culture, personal dedication and interpersonal impact (天时地利人和.） It’s hard to say that without the scale and popularity of Raphael and his studio among his contemporaries, or without the demands and cultural support of his clients, he would have so many prolific work and thus be this impactful. He had indeed succeeded in scaling his work.
• Key to his success of the scale, with some degree of controversy, was the partnerships with his fellow artist and pupils. He may design the composition, and leave the work to his studio and teams to complete, often leading to controversy of ownership. A modern enterprise.