1. While U.S. state activity visibly ticked up around the releases of GPT-3.5/4, Chinese local AI policymaking rose more steadily and then accelerated around the release of DeepSeek R1.

2. Most Chinese policies cluster in Shandong, Shanghai, Guangdong, Zhejiang, Anhui, and Sichuan, indicating an uneven provincial development that likely reflects pre-existing economic capacity and AI ecosystems.

3. Local rules emphasize sectors where AI is readily deployable or labor-substitutable (research, IT, manufacturing), alongside a notable push on education aimed at smoothing labor market adjustment and guarding against equity concerns.

   Share

The Number of Policies

First, the cumulative number of local government AI policies in China has been steadily increasing. Looking at Figure 1, there was a slight slowdown from 2021 to 2023, and the pace picked up around 2025.

I also downloaded AI policies of U.S. state governments from the AI Governance and Regulatory Archive (AGORA)2. Note that, as stated on its website, the data is neither comprehensive nor representative, so please take it with a pinch of salt. Nevertheless, I find this trend very close to my expectation: there was a visible, rapid increase in the number of U.S. policies around the release of GPT-3.5 and 4.

Leave a comment

Comparing the two, China did not have this increase around the release of GPT-3.5 and 4. Instead, it slowly increases its number of local policies over the years. Why is that?

I have two tentative hypotheses here. First, China ended its zero-COVID policy on December 7, 2022, and local governments were slow to regain their footing. Provinces like Shanghai, Guangdong, and Zhejiang, which usually post strong growth and shoulder much of the country’s AI development later, experienced a marked deceleration in GDP growth early in 2022. The zero-COVID policy’s disruptions appeared to have compressed both supply and demand at once, affecting both commercial activity and household incomes. Given the situation, it’s plausible that these provinces had limited bandwidth to engage with the rapid progress in AI abroad.

Second, ChatGPT.com has been blocked by the Chinese government via the “Great Firewall” since around March 2, 2023. As a result, the “tech shock” did not transmit to the general public as much as it did in the U.S. Thus, there wasn’t a sudden surge of local citizen/firm demand that would force immediate municipal rule-making to regulate. In support of this argument, I do see a slight acceleration in the number of AI policies after the release of DeepSeek R1, when AI became a popular tool in China as well.

Subscribe now

Topics of Policies

I also analyze the topics of Chinese AI policies. Somewhat contrary to my expectations, local governments seem to mention AI research a lot more than chips when it comes to AI policies.

Another observation is that after the release of DeepSeek’s R1, there was an acceleration in developing policies relevant to AI research, AI safety, and chips, especially in the first two areas. This further shows that DeepSeek’s R1 was the “proto-shock” for Chinese local governments, prompting them to care more about AI development. This is also in contrast to many people’s view that China does not care about AI safety at all.

Probably in response to more policies on AI safety, frontier AI safety research in China has increased rapidly around the same time. As Concordia AI reports, “From June 2024 to May 2025, Chinese scholars published ~26 papers/month, more than double the previous year’s output.”

Leave a comment

Industries and Collaboration Fields

I also check which industries are mentioned as those that can be assisted by AI. Intuitively enough, research, IT, manufacturing, and education are recognized as industries that can be assisted by AI the most. On the contrary, finance, hotel and catering, mining, and real estate seem to be the least targetable when it comes to AI assistance.

Zooming in on the top industries targeted, scientific research and IT make a lot of sense — occupations in these industries are also among those with the highest AI exposure per Eloundou et al. (2023). Manufacturing also makes sense, as it includes AI-assisted cars, robots, and phones.

I find education very interesting. Some provinces have targeted AI education as early as primary and middle school to ensure students can catch up with the new technology. For example, Implementation Plan for ‘Artificial Intelligence + Education’ in Shandong Province mentions “offer general AI literacy courses across all grade levels” and “Achieve system-wide coverage of AI education in primary and secondary schools.” Moreover, Zhejiang Provincial Higher Education Teachers’ Artificial Intelligence Literacy Framework (Trial) takes a step further to talk about AI ethics in education, stating:

Fairness ethics requires teachers to understand fairness issues arising from algorithmic bias and safeguard the rights of vulnerable groups. Benefit ethics requires teachers to ensure that technology applications serve the fundamental goals of education, guard against technology abuse, and avoid ignoring ethical requirements due to a focus on benefits.

I think the focus on education comes from two places. First, the government worries about unemployment stemming from AI replacing labor. Familiarizing students with AI has become increasingly important to prepare them for the job market. Second, Chinese tradition places a strong emphasis on equality in education. People are concerned that AI could further educational inequality between families of different social statuses.

Subscribe now

Lastly, moving to fields where AI can be used for collaboration, I use the six fields outlined in the Opinions of the State Council on Deepening the Implementation of the ‘Artificial Intelligence+’ Initiative (2025).

We see that most policies mention industry development and science & technology as potential fields of collaboration. By contrast, increasing the quality of consumption and global cooperation are mentioned less.

Conclusion

China’s local AI governance appears to respond more to domestic market demands than global AI progress milestones. The spatial concentration also means there could be deepened unequal AI development (and its spillovers to other industries) in the country in the future.

I hope these facts help people think about the U.S.-China arms race not as two identical players in a prisoner’s dilemma, but as two players, each with their unique conditions and situations, trying to achieve their unique economic and political goals.

Share

Picture this: you're a fresh college graduate in China, aspiring to one of the nation's most esteemed careers – a government official. This role is famously known as the “iron rice bowl.” Much like academic tenure in the United States, it signifies exceptional job security, ensuring you'll always have “rice in your bowl.” Beyond stability, government positions often come with superior benefits. Moreover, Chinese culture has a long-standing tradition of holding government service in higher regard than many other professions, including business. Since the job is appealing, landing a government post is nicknamed ‘going ashore’ (上岸).

The appeal of a government career is undeniable. And so, you've made the decision, joining the more than 3 million young people across China who are diligently preparing for the National Civil Service Exam (the number goes to 8.5 million if you include Provincial Civil Service Exam applicants as well). Here are the questions that you will encounter in order to “go ashore.”

Subscribe now

Are You Qualified?

Before taking the exam, we should look at the requirements for the applicants:

1. Hold the nationality of the People's Republic of China;

2. Generally be between 18 and 35 years of age (born between October 1988 and October 2006). For 2025 graduates with a master's or doctoral degree, the age limit is relaxed to under 40 years old (born after October 1983);

3. Support the Constitution of the People's Republic of China, support the leadership of the Communist Party of China, and support the socialist system;

4. Possess good political qualities and moral character;

5. Have the physical and psychological conditions required to perform duties normally;

6. Possess the work abilities required for the position;

7. Hold an associate degree or above;

8. Meet other qualifications required for the intended position as stipulated by the central civil service authorities.

The Written Test

The first part you will encounter is the Written Test.1 This test is split into two sections — the Administrative Aptitude Test and the Essay Writing Test. The Administrative Aptitude Test usually has 110-135 questions in ~120 minutes covering 5 areas: Common Knowledge Judgment, Verbal Comprehension and Expression, Quantitative Relations, Judgment Reasoning, and Data Analysis. Each question can be worth 0.5-1 point with a final total score of 100.

The second section, the Essay Writing Test, is subjective, including summarization and generalization questions, questions requiring the proposal of countermeasures (or solutions), comprehensive analysis questions, official document writing, and a major essay, among others. Given the length of this piece, I will cover that section in a separate article.

Share Great Divergence

Administrative Aptitude Test

Common Knowledge Judgment

The Common Knowledge Judgment questions include several questions about the CCP (Chinese Communist Party)’s important speeches or history, and some common-knowledge questions. You will see several questions like this:

The answer is C. This is mainly just memorization.

Then you will encounter some more general common-knowledge questions like this:

The correct answer is C.

Verbal Comprehension and Expression

The next few questions are on Verbal Comprehension and Expression. These questions are like GRE verbal tests, where you fill in the blanks with words that fit the sentence the best. Since this relies on knowing Mandarin, I will not discuss it here.

Quantitative Relations

Next is the Quantitative Relations section. The first few are usually about finding the patterns like this one:

The answer is D. Observe the odd-positioned terms in the given sequence: -2, 0, 26, 124. Notice that if you add 1 to each of these numbers, they become cubes of consecutive odd numbers, -1, 1, 3, 5 and the next should be 7^3 -1 = 342.

The answer is B. Firstly, the second hand tip of the clock has gone 5 full rotations, meaning that 5 minutes passed. Therefore, the area swept by the minute hand is:

30/360×π×3^2=0.75π.

To summarize, these questions seem not too hard to solve (comparable to GRE questions, in my opinion), except that it is a little wittier.

Share

Judgment Reasoning

These questions always confuse me the most. The weirdest part is pattern recognition. I often find the pattern somewhat ad hoc. Some online comments even refer to this part as “you know, the government needs people who are just purely lucky.” Anyway, let’s dive in.

The answer is C. Why? Well, the explanation is that, observing from left to right:

• Figure 1 has two white sections, with white dots numbering 5 and 5 respectively;

• Figure 2 has two white sections, with white dots numbering 6 and 4 respectively;

• Figure 3 has two white sections, with white dots numbering 5 and 5 respectively;

• Figure 4 has two white sections, with white dots numbering 6 and 4 respectively;

• Figure 5 has two white sections, with white dots numbering 5 and 5 respectively.

Therefore, the figure to be placed in the "?" position should also have two white sections, containing 6 and 4 white dots, respectively.

Let’s do one more since those questions are so “fun”:

The answer is D. You have to count the number of straight lines in each graph. There are, from the first graph to the last, 5, 6, 7, 8, 9 straight lines respectively. The figure in “?” should have 10 straight lines and thus, D is the answer.

Now, try one on your own!

Leave your answer in the comments!

Leave a comment

Of course, Judgment Reasoning also includes other questions without figures, such as this one about definition:

The answer is D.

Another type of question in this section is analogies:

The correct answer is D, because the two roles in D have a subject–object relationship.

Lastly, you will also encounter logic judgment questions like:

The answer is D.

Data Analysis

The last few questions are about analyzing some financial or operational data. Here is one example:

The correct answer is B.

Share

Did You Pass?

Remember, the exam is 2 hours long, so 1 minute for each question! Usually, you will need around 60 out of 100 points on the Administrative Aptitude Test to have a chance to apply for popular positions. For example, to apply for provincial-level or above positions in the 2019 National Civil Service Exam, candidates must achieve at least 60 in the Administrative Aptitude Test. Other lower-ranked positions or positions in Western China have lower requirements for the Administrative Aptitude Test, but a minimum score of 45 is still required.

Below, I also plot the overall level of competition in each province by using the total participants / total recruits for positions in each province. However, please also note that this is already the result of an equilibrium where some provinces accept applicants with lower scores. The numbers of recruits and applicants are particularly high in Guangdong, Shandong, and Beijing.

What do you think? Could you pass, and which province would you target?

Leave a comment

Aside: AI Models Still Struggle With Pattern Recognition

One thing that I find interesting is that AI models seem to do pretty badly on the Pattern Recognition part of the exam. For example, for this question:

GPT o3’s answer was:

Because of AI models’ inability to solve those visual questions, we also see similar Pattern Recognition questions in US firms’ screening tests. (If anyone wants to make a Chinese Civil Service Exam benchmark out of this, let me know!)

Subscribe now

I think this is a serious mistake, the economics of AI labs is affected by both the supply and demand for their services. Yes, how good the technology gets is important for understanding the demand side, but the brutal economics of AI labs is the lack of moat AI labs have on the supply side. As I discuss below, this can severely constrain AI labs profits even if they are developing incredible technology. In this article, I want to review a bit why running an AI lab is brutal economics and how I think this will change things on the path to AGI.

Subscribe now

Review of the Economics of AI Labs

Running an AI lab is brutal economics because you are trying to invent better AI, but they have little to no moat – that is, no effective way to ensure only you can use your discoveries via patents, trade secrets, network effects etc. Look how competitive the industry is now - the best model went from OpenAI’s o1 to Claude Sonnet 3.5 to Grok 3 to Claude Sonnet 3.7 to Gemini 2.5 Pro, and most sophisticated players switched between models with one-line changes in their .env file. Why is that such a problem? It means when an AI lab makes some breakthrough, they can extract monopoly profits for only a short period of time to make up their R&D costs until a fast-follower lab invents similar technology and undercuts their prices. This undercutting will continue until eventually prices are at inference costs.[2] We saw this dynamic with Deepseek R1 copying OpenAI’s o1 after just 1-2 months at a massively cheaper price.

This puts major AI labs in a tough position. Either LLMs mature or scale indefinitely. If they mature, then any lead will evaporate and fast followers will compete prices down, putting AI labs in the red with their high R&D costs. If LLMs scale indefinitely, then labs will need to continually re-invest huge amounts in compute and R&D like they currently are, which puts them perpetually in the “losing money as a fast-growing startup” mode that they find themselves in right now. The problem is that fast-growing startups need to mature at some point and extract profits somehow, but without a moat it’s not clear how to do that

The main retort people have with concerns about LLMs having no moat is that online search like Google/Bing/etc also has no moat but Google makes a ton of money. I think this claim is just wrong on the substance - search has a very clear quality moat where most people strongly prefer Google. Most economists think this quality moat comes from network effects, the more people use Google, the better their data is and thus the better their search is.

I don’t think this is true in AI, at least not yet. We have seen multiple new AI labs (xAI and Deepseek) go from 0 to arguably at or beyond the frontier OpenAI was at. There are fairly minimal network effects with data when the only real inputs you need are Internet Scale Data, GPUs and researchers, which anyone can buy.

Let’s take seriously the chance that LLMs will scale to AGI, something that can flexibly substitute for human labor. Does that change the brutal economics of the situation? No. The problem of having no moat is a supply-side problem, but LLMs being incredible just means demand will be high. The price war dynamic we analyzed above can happen with something that has incredibly high demand. In fact, that is sort of the usual case in economics. Think about basic goods like water, food, chairs, pencils, etc. There is incredible demand for these goods but no company makes a killing off them because prices are competed down to roughly the cost to provide the good.

Just because human labor is paid $60-70 trillion in wages a year does not mean that AGI labs will capture even a tiny percentage of that amount from automating labor if supply is unrestricted. In fact, in the simple economic model we considered above with price undercutting, the equilibrium is AI firms capturing exactly 0% of the surplus!

We can even see this phenomenon now – the SWE-Lancer benchmark evaluates over 1,400 freelance software engineering tasks from Upwork, collectively costing $1 million to hire freelancing to complete. The top-performing model, Claude 3.5 Sonnet, completes tasks worth approximately $400,000. But it only costs a few dollars for consumers to pay for Claude to complete these tasks. Even though consumer willingness to pay for AGI, does not mean that almost any surplus goes back to the companies building AI.

Leave a comment

Implications for the Path to AGI

1. Novelty and Speed Premium are Very High.

When an AI lab makes a breakthrough, they have an incredibly short amount of time where they are a monopolist on this more advanced AI capability. This allows them to charge very high margins and make up their R&D costs and more. Therefore, maintaining a lead and maintaining it for as long as possible is perhaps the core variable in how much profit AI labs can make. Therefore, we should expect firms to put a high premium on going as fast as possible. From an AI safety perspective perhaps this is quite undesirable.

A related point is that AI labs taking very different paths have the chance to extract monopoly profits for a much longer time. For example, SSI, the new AI lab, run by Ilya Sutskever is supposedly “climbing a different mountain.” than other AI labs. If this strategy works and they are able to keep trade secrets, then they would be able to extract monopoly profits for far longer than other AI labs who are all roughly pursuing the same strategy (scaling pre-training, RL post-training).

Subscribe now

2. Higher Economic Incentives to Automate AI R&D

Right now, there is a big debate on whether AGI will first self-improve by automating AI research or AGI will first substitute for labor in the general economy. Those arguing for general substitution of labor often appeal to the very large economic benefits of automating labor in terms of wages saved.

However, taking into account the supply side features of AI development flips the sign on this argument. Just because the demand is very high for automating labor does not mean the AI labs will get a high return from it. As we mentioned before, because of unrestricted supply, labs will get prices competed down and almost all the surplus will go to the demand side, strongly dampening the incentive for broad automation.

In fact, the economics with AI labs will, if anything, push towards automation of AI R&D. AI R&D is unique because AI labs will be both the supply side (they sell AI) and the demand side (they buy researchers who work on AI). Therefore, even if the surplus goes to the demand side because of price wars, AI labs still capture the surplus since they are the demand side for AI R&D. Therefore, AI labs actually have a unique incentive to automate AI R&D before economy-wide automation.[3]

Subscribe now

3. We might See Major Bankruptcies of AI labs, even after AGI

Right now AI firms are bringing in billions of dollars in investment. Because investors are putting them in startups in high growth mode, they can tolerate large losses in the short term. But as time goes on, and if AI labs struggle to turn a profit in 3-4 years, I think we may see major bankruptcies of AI labs such as OpenAI or Anthropic. Again, this might happen even if Anthropic or OpenAI have something that looks like AGI. After that, you could imagine someone like Google or Meta buying these labs and incorporating them, which would obviously be a major shakeup in the race for AGI .

Share

4. There may be major changes in pricing structure.

Right now, there are a wide variety of models but all sell with a similar pricing structure – some constant price per token. This pricing structure leads to the brutal economics discussed above, where price wars eventually destroy everyone’s margins. Perhaps, after a while, AI labs will try alternative pricing structures like building in advertising into model responses.[2] I’m not sure this will work – if users dislike ads enough, then firms will just get into the same brutal economics but instead of competing on price, they compete on how little ads they show you. But it might, and I think those taking AGI seriously should think if these ad-focused models are encouraged, discouraged etc. [5] If ads are undesirable, now might be a particular high-leverage time to pass laws banning ads since you might face little resistance right now.

Conclusion:

AGI is strikingly plausible in the next few years, but will AI labs be primary beneficiaries of that? I’m not so sure, and I’ve tried to argue why the brutal economics of AI labs won’t change even after AGI.

Subscribe now

Footnotes

[1] It seems the people who are bearish about the economics tend to be bearish about the technology (c.f Gary Marcus). I want to stake out the position that is bearish about the economics but bullish about the technology.

[2] There could be some pricing of fixed costs in as well, but this makes the situation even worse for the leading lab because the fast follower has lower fixed costs since they don’t need to spend as much on R&D.

[3] At a technical level, these results show that the economics of AI labs is not well approximated by a social planner because of the externalities involved in AI R&D. These problems cannot be fixed by including a wedge parameter because the competitive dynamics introduce differential incentives for different sectors of automation, rather than a uniform slowdown. We need more microeconomic models of AI labs to complement the macro models!

[4] Or alternatively, model providers will try to capture other parts of the stack.

[5] For what it’s worth, my intuition is that having AI agents whose minds are modified to like certain products would be pretty dystopian, but I need to think about this more.

While we agree that market forces might somewhat under-provide AI Safety, we argue that existing incentives faced by AI labs significantly mitigate the extent of this market failure. In this article, we will review two major arguments for AI Safety being a market failure and our arguments for why these arguments are overstated.

Subscribe now

1. Externalities and AI Safety

The first major argument for AI Safety being a market failure is that AI risk is an externality and thus basic economic theory tells us that AI Safety will be under-provided. While this argument has some truth in it, this argument overlooks that AI labs internalize much of the potential harm associated with AI risk, reducing the severity of the externality.

Imagine a catastrophic AI failure causing significant human casualties. Such an incident would severely damage the reputation of the responsible AI lab, trigger massive financial liabilities, prompt public outrage, and could lead to extreme regulatory responses, potentially even a complete ban on advanced AI research.If these outcomes seem crazy, note that right now 50% of people support banning AI more powerful than GPT-4. After a disaster, drastic measures—such as permanent bans on advanced AI—could rapidly become politically acceptable and even demanded by the public

AI Safety research also directly benefits product quality. The main technical safety contribution to date—Reinforcement Learning from Human Feedback (RLHF)—was essential in developing useful AI assistants like GPT-3.5. Aligning AI systems with human preferences isn’t merely altruistic; it directly enhances product marketability and profitability.

That’s why AI labs are paying $300k-1M+ total compensation for scientists to work on alignment and AI safety. Such high compensation contrasts sharply with other severe market failures, such as factory farming, where similarly high salaries for harm reduction roles are rarely observed. We also don’t see high salaries for AI Safety in the government, indicating there is more optimization for AI Safety in the market, hardly something you normally see in cases of market failure.

Subscribe now

2. Arms Races Are Compatible with Safety

A more advanced argument for AI Safety being a market failure is that the competitive dynamics with multiple firms lead to over-investment in capabilities and under-investment in safety. The argument goes that if e.g. Firm A decides to not add some useful but dangerous capability to their AI, then Firm B will just speed ahead while investing nothing in safety. Often this argument will explicitly appeal to the Prisoner’s Dilemma, a classic game theory example where everyone would be made better off by cooperating, but the equilibrium involves everyone defecting. The argument claims AI labs are in the same situation, where they all want to invest more in safety, but the only equilibrium is going as fast as possible with as little safety as possible.

We think this argument is substantially oversimplified for two main reasons.

a. Defecting May Not be a Dominant Strategy in an AI Arms Race: In the prisoner’s dilemma, defecting is a dominant strategy. No matter what your opponent does, defecting leaves you better off. See the graph below for an illustration.

However, an AI arms race may be crucially different. Unlike the classic Prisoner’s Dilemma, defecting—pursuing more rapid capability advancement at the expense of safety—is not always beneficial. AI labs face substantial risks (reputation, regulation, liability) from unsafe behavior, meaning the potential cost of defection can outweigh short-term competitive advantages. Overall, defecting may or may not be in your interest depending on how much return there is to having slightly more capabilities and how much harm there is by having a slightly less safe model. Therefore, we could model the AI Arms race in the following way, where the variable x controls the benefit of defecting. If x <= 1, i.e. the returns to being faster is less than the cost of less safety, then both players playing the safe outcome is an equilibrium.

Moreover, in this simplified model there are only two options, safe and unsafe. In a more realistic model, there would be a continuum of actions and whatever margin you are currently at controls the benefit of defecting. Even if x starts being > 1, it will fall as you get close to guaranteeing a disaster, so even if cooperation on the fully safe outcome is not possible, partial cooperation on safety is.

b. AI Arms Races are not One Shot Games. The canonical model of the prisoner’s dilemma involves a single turn where players make a decision and never interact again. But AI Arms Races occur over time with many repeated decisions on how much to invest in capabilities and safety. In game theory parlance, an AI Arms Race is a repeated game. A very famous game theory result, the Folk Theorem, shows that you can get cooperation in the prisoner’s dilemma in repeated games under a wide variety of settings! Repeated interactions allow firms to build trust and punish defection through subsequent actions, significantly increasing the possibility of stable cooperation over time.

Of course, repeated games tend to have many equilibria, so perhaps the current market is not in this good cooperative equilibrium. But, this result shows that cooperation is not something that the market forecloses. Indeed, current examples exist where AI labs have responsible scaling policies and alignment research—indicating partial cooperation already occurs.

Share

Conclusion:

We have argued that the claim of AI Safety as a severe market failure is overstated, as substantial incentives already exist for labs to prioritize safety [3]. Nevertheless, the importance of AI Safety remains paramount. Problems do not require a perfect market failure to warrant serious public attention and proactive policy solutions. The market may demand a solution, but humanity still needs to provide the talent to solve it! [4]

Leave a comment

Footnotes:

[1] Throughout the article, we will use AI Safety to refer to investments that prevent direct danger from AI systems by loss of control or misuse. This is distinct from research that aims to prevent a concentration of power, optimal governance, etc.

[2] The cost of someone defecting against you is controlled by y. The value y should take in an AI arms race is somewhat unclear, but it doesn’t matter much for our analysis.

[3] We are also sympathetic to more advanced Econ-style arguments for underinvestment that focus on AI firms with heterogeneous beliefs about how much safety research is needed. We are working out a formal model of this.

[4] We are also somewhat sympathetic to the implication that other aspects of AI may be more neglected than Safety. For example, perhaps preventing a concentration of power after AGI is more neglected because no private actor has any incentive to prevent that.

Subscribe now

1. AGI is “Achieved” on Short Tasks

The best argument for AGI soon is that we have something pretty close to AGI in capabilities right now on simple QA questions. o3 is getting 25% on Frontier Math problems, 36% of Humanity’s Last Exam, and near human expert levels of 90% on Google-Proof Graduate Questions. All of these tests measure brutally hard PhD and beyond level performance in a wide variety of academic fields. We are quite confident no human on earth exceeds all three of these scores. Therefore, it is fair to say that AGI has pretty much been achieved on simple QA benchmarks.

What is left to achieve AGI? Well, it’s mostly long-term planning, executive function, agency, etc., to turn that raw intelligence into useful labor. That is not much left to achieve! It wouldn’t be surprising if one small tweak—scaling, RL, or something—adds agency, and then AGI is here.

2. Agency Seems On Trend

Of course, how close we are to AGI depends on how hard agency is to add to LLMs. Maybe agency is very difficult to achieve. For example, Epoch argues that agency may be difficult because it is a very fundamental trait in animals, baked in by evolution millions of years ago. Since it is so fundamental in humans, there is not a lot of text data about how to learn it, nor is there a clear way we know to ‘teach’ agency in the same way we teach math and code.

Nevertheless, LLMs seem to be getting better at agency quite quickly. For example, METR estimates that the task horizon at which models can effectively operate on has doubled roughly every 200 days for the last few years. Moreover, Claude plays Pokemon has seen substantial improvements from Claude 3 to Claude 3.7. While both benchmarks show there is still quite a ways to go on long-term planning/agency, the trend on both shows that we are heading in the right direction with general post-training methods. If we naively extrapolate the trends, models achieving human-level performance on 24-hour tasks is about 4 years out. That is well within the frame of short AI timelines!

Share

3. There Seems to be a lot of Low-Hanging Fruit Left in ML

Since the release of GPT-4, we have seen a very large increase in various benchmarks. Most of the returns have come from simple algorithmic improvements like better RLHF, better character training, RL on verifiable rewards, etc. None of these are ‘genius’ ideas. ML researchers are mostly just doing obvious things, and this seems to be working. We’ve only picked this low-hanging fruit for about 2-4 years, so it would be pretty surprising if there is not a lot left to pick up. Even just normal ML progress should see a lot of progress by picking up these simple ideas.

4. The Death of Pre-Training May Be Overstated

GPT 4.5 on release was seen by many as a disappointment and a sign that the era of pre-training scaling was ending. But we think this conclusion, while it has some truth to it, is a bit overstated. As noted by Epoch, GPT 4.5’s performance on GPQA was above trend for pre-training scaling laws, and many people, like Tyler Cowen, are noting that GPT 4.5 is funnier, smarter, a better writer, and has more world knowledge. These general gains, while not that useful to many end-users, do still indicate that scaling generally boosts intelligence. These general gains will likely lead to downstream boosts when adding reasoning, agency RL, etc. Of course, this is not a sure thing, but I do think pre-training on top will continue to be important for years to come.

Subscribe now

5. Unusually High Investment Will Occur In the Next 5-10 Years

OpenAI hit 4 billion in revenue this year and has seen its valuation balloon to $350 billion. Moreover, demand for AI is increasing, particularly from enterprise where usage increased from $600 million to $4.6 billion. Additionally, AI applications are turning up from left to right from coding assistants like Cursor to agents like Manus. Accordingly, demand for AI is high and rising. We expect another large jump in revenue for OpenAI in 2026 as well.

This money flowing into AI will be aggressively reinvested into R&D, creating a one-time, massive acceleration in AI R&D over the next 5-10 years. Note that this investment growth cannot go on forever. Once you start investing a significant share of GDP in AI R&D, it becomes hard to increase that investment. This creates a narrow time window where exponential investment growth can drive exponential capability gains, followed by a natural ceiling once AI becomes a mature economic sector with more constrained growth potential.

Leave a comment

Conclusion

We have now reviewed the various arguments for and against short timelines. Our main takeaway is that AGI in the next 5-10 years is plausible! It’s certainly not a sure thing for the reasons we mentioned in the previous post, but even a 30% chance of AGI in the next few years is striking. One interesting result of our analysis of the arguments is that AGI is likely to be soon (<5 years) or in decades, with little probability mass on the intervening years. If LLMs, with some tweaks, take us to AGI, then we could see huge progress even in the next 3 years. If LLMs don’t take us to AGI by 2035, then we will be back in the wilderness.

In this article, we are going to present the strongest arguments we have for why artificial intelligence progress might take longer than some expect, e.g. 10+ years. These reflections are our “inside-view” about why AI timelines might take a long time rather than appealing to the market, forecasters, etc. In a few weeks, we’ll follow up with a companion piece that discusses the best reasons for anticipating shorter AI timelines.

1. Pre-training May Be Running Out

Caption: Figure of pre-training scaling laws.

The strongest argument for the near-term arrival of AGI was from the predictive power of pre-training scaling laws. The substantial performance gains from GPT-3 to GPT-3.5 and subsequently to GPT-4—achieved largely by increasing computational resources—were striking. This progression serves as strong evidence that scaling laws warrant serious consideration, as it demonstrates that expanding compute can indeed lead to greater intelligence.

And yet, it looks like pre-training scaling is starting to run out. GPT-4.5 is seen by many (but not all) as a disappointment. It certainly didn’t crush benchmarks as GPT-4 did on release, and improvements seem more marginal. Yes, you can decrease next-token prediction loss by scaling up your system, but the mapping from next-token prediction to raw smarts seems to be falling (although this is debatable, a point we will discuss more in a few weeks). This failure will significantly weaken companies’ appetite to train GPT-6 or 7-sized models.

The failure of pre-training has led many to state that there are other methods we can get to AGI including reinforcement learning (RL) on top, test time compute, and automating AI R&D. We will discuss all those possible routes in the rest of the article, but please notice that these are not the pre-registered predictions that were made a year ago (e.g. here) which focused on scaling pre-training. Moreover, they are far more speculative. We had a lot of empirical evidence on the returns to scaling pre-training, but we do not have that for these other categories. While these other paths could work, we do not have much evidence on these other paths to AGI other than vibes.

Subscribe now

2. RL May not Work on Long-Time Horizons

Since scaling may start to fail, many of those with short timelines point to o1 reasoning styling models as a ‘new dimension to scale’. While, like many, we find o1/o3 extremely impressive, whether this will ‘scale to AGI’ easily on domains we actually care about (e.g. running a business, being a remote worker, etc) is still in question.

Reasoning models are trained via reinforcement learning (RL) on math and code. Because these settings have a simple verifier, models can learn quickly to produce more and more accurate answers. The mention of RL makes people think we can easily apply the same setup to any verifiable setting like business, trading, robots, etc. But, in fact, these cases could be much, much harder than math and code because they involve long-term actions and planning. The deep learning revolution has taught us that these systems work really well with a large amount of data. However, in these long-term tasks like running a business, you do not get a large amount of feedback data. You get these weak intermediate signals before the business either takes off or fails. Somehow, humans learn from these very weak signals to make decent decisions, but it is not a setting that is very amenable to RL systems.

To make this point, we will need to get into a bit of the weeds about what RL is because the RL training in recent reasoning models is barely RL at all. In a reasoning task, models simply take one action (outputting their answer) and then get a known/verifiable score. This contrasts with a true RL environment, where models take many actions and get intermediate feedback before getting some final score. The later environment is much harder because the model needs to balance exploration and exploitation, credit assignment, learning a reward model for the intermediate states, etc.

Simply speaking, if you take 1 action and then get told if it is good or bad, it’s very easy to update quickly to start taking good actions. But if you take 1000 actions before you are told if your end state is good or bad, it is very hard to tell if your first action was good or bad. To get AGI at the level of drop-in remote workers, we will have to get the true RL environment working, and so far that has remained elusive.

Share

3. Not all Aspects of Test-Time Scaling Can be Scaled

When reasoning models like OpenAI’s o1 were first released, one of the major talking points was that it unlocked a new dimension we could scale —test-time compute. However, conversations on this are usually ambiguous. There are really two dimensions on which we can scale models in the reasoning model style. First, we can scale the number of reasoning examples. Second, we can scale how long models respond for. We are bullish on the first method of scaling but not the second.

The first dimension we could scale is just the number of training examples. We are bullish on this because you have more reasoning tricks you can learn. The original o1 let models learn basic reasoning patterns like “I think I made a mistake, let me go back” which led to downstream performance increases. We expect that more math examples will let you learn more of these tricks. People have already seen what these improvements look like in the progress of o1 to o3.

However, while we are bullish on this direction, we are somewhat skeptical that this will let us scale to AGI directly. First, we have quite a limited understanding of how far we can push out reasoning RL. We really have just two data points in o1 and o3. It’s plausible that we can scale it out further, but it’s not like we have good data to predict the trends accurately. Therefore, there exists a high degree of uncertainty about how far one can go in this paradigm. It might become superhuman very quickly, but it also might hit a wall. Second, we have seen fairly limited generalization from math/code to other tasks, so it is clear that reasoning RL is not alone sufficient to get us to AGI.

The second dimension is scaling up the length and attempts. However, in both categories, these are probably not the items we can scale directly. Take this now famous graph.

One common interpretation of this graph is that we can scale test-time compute (AI thinking for a longer time) to get higher accuracy in the same way we scaled train-time compute (training for a longer time). However, we think this interpretation is quite misleading as it confuses correlation for causation. In the train-time case, we had direct control over train-time compute and could scale it up by adding parameters, data, etc. The increase in this compute caused accuracy to go up. But the test-time case is different; test-time compute increasing did not cause accuracy to increase. Rather, RL on reasoning traces caused both test-time compute and accuracy to increase. RL increased both of these by teaching the model to use reasoning patterns like “let me check my work”, “I think I’ve made a mistake, let me go back”, etc. These reasoning patterns took more tokens, which increased compute usage at test-time but were also useful to get correct answers. But notice, the increased compute did not cause the accuracy to go up; it’s a byproduct of using more tokens for reasoning. This fact matters when we think about extrapolating these graphs out. If you have already learned all the reasoning patterns, there is no way to “scale” test-time compute up further that will get you more accuracy.

Leave a comment

4. Automating AI R&D Does Not Necessary Cause a Self-Improvement Loop to AGI

Finally, the last frequently mentioned point for AGI soon is that AI systems are close to automating AI R&D, and this will cause systems to self-improve to AGI. Our core view is that automated AI R&D with coding agents, etc., will likely accelerate “normal” ML science where you run experiments, check benchmarks, etc. So, insofar as the path to AGI looks like normal ML science, we expect automated AI R&D to get us there a bit faster. However, the case for longer timelines we sketched out above relies on normal ML not getting us directly to AGI. If RL, scaling up pre-training, etc., do not get us to AGI, then we need a paradigmatic breakthrough on the order of neural networks, transformers, etc., to get there. Automated AI R&D may help a bit here as you can experiment faster, but that’s not really the binding constraint for paradigmatic breakthroughs. Therefore, automated AI R&D will be more of an acceleration on the already short timeline worlds where we know how to get to AGI via the above methods, but it will not have much effect in the worlds where timelines are long because LLMs are a dead-end to AGI.

Subscribe now

Conclusion:

The progress of language models over the last 2-3 years has been shocking, and yet the effect has also been shockingly narrow. The raw intellect of the models seems extremely high at first glance, but even basic computer use is taking longer than expected. The path to AGI could feel similar. We might get blazing smart models on Q&A formats, but when it comes to automating significant shares of human labor, it might take a surprisingly long time.

Leave a comment

There are two main arguments for such a tariff:

• The U.S. should protect its own semiconductor manufacturing business

• TSMC could avoid this tariff by building more plants in the U.S. So, instead of a subsidy that incentivizes TSMC to build plants in the U.S., President Trump will use the tariff as a punishment for NOT building plants in the U.S., hopefully reaching the same effect — more chip factories in the U.S.

Here, we examine each of those arguments one by one and point out what to take caution of.

Argument 1: The U.S. should protect its own semiconductor manufacturing business

President Trump claims that major American tech companies have shifted chip production almost entirely to Taiwan, which he says accounts for about 98% of the global semiconductor market. He believes imposing tariffs would spur domestic manufacturing and lessen the need for government subsidies like those in the Chips and Science Act.

Part of this claim is definitely true, the U.S. tech companies did shift chip production to Taiwan. The chip industry in Taiwan began with the Technical Advisory Committee at the Industrial Technology Research Institute (ITRI) made a deal with RCA for a 10-year technology transfer agreement that stipulates a $2.5 million technology transfer fee and a $1 million technology licensing fee in 1976. Then, later in 1987, the ITRI established TSMC. TSMC’s founder, Morris Chang, spent 25 years at Texas Instruments—and thus brought extensive semiconductor manufacturing and management experience from TI. The experience at TI gave Chang deep insight into chip manufacturing processes, supply chains, and management. (A side story: apparently Morris Chang suggested doing a foundry when he was at TI but his idea was not accepted. Otherwise, it would have been Texas SMC instead). In 2021, Taiwan’s chip manufacturing industry occupied 64% of the global market and for advanced process technology under 7nm, TSMC has almost the whole market.

Now, is it practical to protect the U.S. semiconductor manufacturing business? Sometimes people think “country A gave country B the technology. Now A wants it back. Country A just needs to take it. That’s all.” However, the reality is that chip manufacturing requires long-term dedication. The semiconductor industry has a relatively low turnover rate, meaning it takes a long time for new product generations or innovations. It not only requires a huge amount of money but also a persistent effort at fostering talent. Yes, the U.S. once led this industry, but after outsourcing it to Taiwan for over three decades, it is not super easy to take back.

One piece of evidence for this claim is the important personnel in TSMC. The founder of TSMC, Morris Chang, studied at Harvard and MIT and then joined TI in 1958. The second, third, and fourth generations of high-ups in TSMC are listed below:

These figures show that leadership at TSMC is now being nurtured in Taiwan instead of the U.S. Wang Ying-lang, Liao Yung-hao, and Chang Zong-sheng all graduated from Taiwan’s universities, and pretty much immediately joined TSMC upon graduation. This means that the chip manufacturing experts (professors, engineers, etc.) are now concentrated in the little island of Taiwan and they are nurturing the next generation of experts. These network effects are hard to break - almost every talented 20-year-old engineer in Taiwan wants to work at TSMC and has access to many mentors. If the U.S. is serious about re-shoring semiconductors, they have to be prepared for a long-term investment in a similar network.

I believe in the U.S.’s potential to achieve anything. However, it is not going to be a quick fix. If the US wants to re-shore, it needs to tackle the problem seriously for decades and do more than put a quick tariff on TSMC. U.S. firms will have to learn from TSMC and potentially collaborate just like TSMC did in the 1970s and 80s.

Subscribe now

Argument 2: the U.S. could force TSMC to build more factories in the U.S. by punishing it for not doing so.

“The only way you’ll get out of this is to build your plant —if you want to stop paying the taxes or the tariffs— you’ll have to build your plant right here in America,” Trump added. “That’s what’s going to happen at record levels.”

This argument is more valid. Having such an important sector almost entirely concentrated in Taiwan is a major geopolitical risk.

However, the core problem is that TSMC’s supply of factories is not very elastic in response to tarrifs. Why not?

• First of all, the cost of making chips is way higher in the U.S. than in Taiwan. Right now, TSMC makes plants in the U.S. and Japan. However, engineers earn roughly 3 times as much in the U.S. than in Taiwan on average. On the contrary, Japanese engineers earn roughly 1.5 times as much as Taiwanese engineers. Not to mention lower-skill workers who earn even less in Taiwan.

  With this high cost, it is questionable whether it is cheaper to just pay the tariff or build more factories in the U.S. From the beginning, it is more likely that TSMC chose to build plants in the U.S. to circumvent geopolitical risks than simply to avoid shipping costs.

• Second, the U.S. is not giving a high subsidy compared to Japan, the other destination of TSMC’s overseas expansion. According to Taipei Times, “Japanese Minister of Economy, Trade and Industry Ken Saito said that the Japanese government would grant ￥732 billion (US$4.98 billion) to JASM to build the second fab in Kumamoto. Along with a subsidy of ￥476 billion ($3 billion) for the first fab in Kumamoto, the Japanese government’s total grants to TSMC would top ￥1.2 trillion ($8 billion).” On the other hand “TSMC is set to secure more than US$5 billion in federal grants to support its investments in Arizona, much lower than the subsidies planned by Japan.” Japan gave TSMC a higher subsidy with a lower cost of hiring labor. It is more difficult to build a TSMC factory in the U.S. as a result.

• Third, the U.S. working culture is very different from Taiwan’s. TSMC engineers do not have a work-life balance at all. An anonymous comment on Glassdoor says: “The working hours could be very long sometimes for an engineer. Average working hours is 8:30 till 20:00~21:00. When abnormal things happen in the Fab, you might need to work till 23:00 for 2 or more weeks.” Former Executive Vice President & Co-Chief Operating Officer Chiang Shang-yi also said that TSMC’s employees, whether day or night, are always willing to stay at the company, put in overtime, and work through the night without a word of complaint. He once even joked with foreign friends, “Because Taiwanese men all serve in the military, they’re used to being woken up at two or three in the morning for guard duty. Working the night shift at TSMC is just like standing guard—compared to Westerners, they can endure more hardship and show greater willingness to sacrifice.”

  As a result, it is a challenge for TSMC managers to organize the American engineers. Naturally, one way to solve it is to import Taiwanese engineers. I myself have noticed how many more Taiwanese restaurants have opened up in Phoenix when I went back there. Indeed, of around 2200 employees at the TSMC plant in Arizona, half are from Taiwan. Of course, TSMC claims that this percentage will decrease in the future as mass production is more on track.

The above are all the reasons why it is difficult for TSMC to build plants in the U.S. That is also why mass production is delayed from 2024 to 2025, given that TSMC rarely breaks its promises on production. But then why is TSMC making factories in Arizona at all now? Their primary motivation is

1. Geopolitical: They understand the risk from China and do not want to put all their eggs on the Taiwanese island.

2. Energy use Constraints: TSMC already uses about 8% of energy in Taiwan. Moreover, energy use is expanding by about 20-30% at each new generation. Starting factories in new countries is particularly appealing from an energy perspective, especially the U.S. which has cheap energy in relative abundance.

Overall, tariffs give some incentive to TSMC to build more factories in the US, but the incentives are fairly weak compared to the larger forces at play. Unless President Trump is planning extremely large tariffs - they likely won't move the needle for TSMC on factory location (but this might cause low motivation for Taiwanese engineers). Other levers, like subsidies and geopolitical promises, may get more return on investment.

Conclusion

To conclude, while we understand some of the motives behind the U.S. government's adding tariffs on TSMC, we think there are considerable cautions it should take. One policy that the U.S. could take is a combination of the carrot and the stick — while continuing to subsidize TSMC and promoting the positive image of TSMC in the U.S., also threatening to add tariffs if TSMC does not build enough plants in the U.S.

Taiwan has tried to build plants in the U.S. WaferTech and Foxconn were precedents. But their results were not ideal. If TSMC in Arizona also fails, it will be a harsh blow to Taiwan, as TSMC is the most successful company in Taiwanese history. So if the U.S. really wants to have more semiconductors in the U.S., perhaps the highest return thing is to make sure the Arizona plant performs well. We certainly hope TSMC will succeed in the U.S., making a win-win for both countries.

Subscribe now

Other references that are not already linked in the article

Lin, H. (2023). The Light on the Chip Island. Zao An Caijing. https://www.amazon.com/dp/B0CPHV67H9

This article takes you on a journey through time, uncovering how Japan's tea-serving beauties of the Edo era paved the way for café waitresses, hostesses, and finally, the maids of today’s Akihabara. Let’s start our exploration by stepping back to the streets of Tokyo during the Edo period.

Edo Period (1603–1868): The Mizuchaya (水茶屋)

During the Edo period, many mizuchaya, or teahouses, opened along the roads. In the beginning, they were very simple shops with just a reed screen and served tea and some small traditional desserts. When it came to the middle Edo period (say, ~1750 in Edo and possibly earlier in Kyoto), to be competitive, many teahouses had beauties serving tea. Many men would be attracted to the teahouses because of a beautiful tea maid and would pay around twice the price for tea. Some were even willing to pay 20 times the price at normal teahouses!

Kasamori Osen (1751-1827) was one of those tea maids. She started to work for her family business in the Kagiya teashop near Kasamori Inari Shrine in Yanaka when she was 14. Due to her extreme beauty, many customers came to the teahouse just to see her. As her popularity grew, writers and many Ukiyo-e artists started to depict her in their works. She gained quite some fame and then disappeared when she was 19 years old, leaving only her father in the teahouse. The male customers were so disappointed that they said “A tea kettle has become a medicine pot.” It is said that she married a Shogun guard. Today, there is still a monument for her in Yanaka.

Here are some artworks depicting Kasamori Osen. For example, the famous Suzuki Harunobu has this artwork below, where Osen is located in front of the torii entrance and turning from her admirer.

Of course, many teahouses also evolved into intermediaries between customers and prostitutes. But this is another story that I might cover in future articles.

Early 1900s: Café Waitresses (女給)

To begin with this period’s story, I want to first mention this interesting transition from teahouses to Cafés. During the Meiji period (1868-1912), coffee was popularized in Japan. Previously, Japanese people were aware of this bitter and dark thing that Dutch people drink, but they were not big fans of it. They had to advertise it as if it was some herbal medicine for health. But when it came to the Meiji period, coffee became popular. The first coffee shop was opened by Nishimura Tsurukichi, called Kahi Chakan (可否茶館). By the way, fun fact—Nishimura also had a Chinese name, Tei Ei-kei, or 鄭永慶. He was related to the famous Koxinga, or 鄭成功 in Chinese. Even today, you can still go to Ueno and find a little site saying this was the site of Kahi Chakan. For people who are interested in the history of coffee in Japan, I recommend doing this little expedition when you visit Tokyo. See Map. I also recommend the book Coffee Life in Japan by Merry White, if you are interested in Japan’s culture of coffee (I will refer to her work for this section as well).

Let’s refocus on café waitresses. As more and more cafés were open for socializing venues around the turn of the century, competition again led to the hiring of pretty girls who could pour coffee and talk to the customers. Beyond serving food and beverages, those café waitresses were expected to engage in polite conversation with the right amount of flirtation, enhancing the café's welcoming atmosphere. These flirtations became bolder and bolder as time went by (especially after the opening of Café Tiger). Those cafés also began to serve alcohol and play jazz music. In Hayashi Fumiko’s autobiographical novel Diary of a Vagabond, she depicts her life working in a café where alcohol is also served. She mentions that a customer tells her: "Try drinking ten glasses of King of Kings. I bet you 10 yen!" Since those cafés were so different from the traditional cafés that just serve coffee, the traditional cafés had to rename themselves as “pure café” or “純喫茶.”

As for the salary, it was not much and the waitresses had to count on the tips for their living. Usually, the mistresses were tipped 1 yen per time of service by the male customers (One gallon of milk is ~0.8 yen).

Despite the fact that the waitresses still catered to a majority of male customers, these positions also gave women opportunities to be in touch with the fast-evolving public realm. As cafés were indeed hubs for interesting conversations and frequented by customers of various careers, café waitresses sometimes served as models for books and movies. As White collects in her book, Tanizaki Junichiro’s novel Naomi, Kazuo Hirotsu’s Jokyuu, and Takeda Rintaro’s Ginza Hacchome are all examples. Moreover, many women who grew into famous public figures worked as café waitresses before. The aforementioned Hayashi Fumiko, who wrote about her experience working at a café in Diary of a Vagabond, later joined Pen Corps and became a quite controversial reporter (as her writings were pro-militarism). Another famous female writer, Sata Ineko, published "Restaurant Rakuyo," which described her experiences as a waitress at a café and was highly praised by Kawabata Yasunari. She was a prominent figure in the proletarian literature movement and focused on women's issues. Especially in the aforementioned "Restaurant Rakuyo," she vividly depicts the struggles of working-class women, particularly those in low-paying and often exploitative jobs. Her experience working as a café waitress very likely gave her some inspiration and helped with her realistic writing.

In the 1930s, Police Departments started to crack down on the café-bars. Regulations for Controlling Special Restaurants and Bars were enacted, where special restaurants and bars were defined as "a restaurant or restaurant with Western-style facilities where women serve and entertain customers." Obviously, it referred to those café-bars. Then, during WWII, due to import restrictions and cultural shifts against Western goods, coffee itself became less popular.

Post-War Era: Hostess Clubs

After WWII, Hostess clubs evolved from pre-war café-bars, cabarets, and dance halls. Entertainment districts such as Ginza and Akasaka became hubs for hostess clubs. They catered primarily to the influx of American GIs and the emerging Japanese business class, as Japan’s economy grew fast. Its direct origin was the "Oasis of the Ginza", opened in November 1945 in the basement of the Matsuya Ginza department store by the Special Comfort Facilities Association's cabaret division. It was established for the occupying forces.

Hostesses’ roles were largely similar to Café waitresses’ roles, with several caveats. First, using “café” as a camouflage was no longer necessary, those new establishments were more like Western nightclubs. According to the Act on the Control and Proper Management of Entertainment and Amusement Businesses in 1948, those venues that entertain customers and provide food and drink, cabarets, waiting rooms, restaurants, cafés, and other facilities were now permitted. Second, girls were often dressed in Western clothes instead of kimonos. And in terms of the work content, Allison described in her book her work experience in a club called “Bijo” in the 80s. her work included “light customers' cigarettes and pour their drinks” and there were three prohibitions “don't smoke, don't rest your elbows on the table, don't put your hands in your pockets in front of customers.”

Interestingly, as the economy thrived in the 60s and 70s, salarymen often went to clubs to socialize with their workplace friends. The environment of those clubs seemed to help solidify group identity and hierarchy among coworkers. Drinking and socializing in a liminal, semi-private space fostered camaraderie and trust among coworkers. This mirrors the environment of the cafés in the 1920s — it was not purely about the sexual relation between the men and the women, but sometimes more about the males bonding with each other.

By the way, during this time, host clubs, where men serve women customers, also started to emerge. It began in the 1950s in a small scale, but today, host clubs are everywhere.

Hostess clubs are still in place today (though they underwent quite a lot of reform and evolution. Big nightclubs were replaced by small-scale bars). Nowadays, the hourly pay for the hostesses is around 5000 yen (33 USD) per hour. If you are interested in the nightlife in hostess clubs and how salarymen socialize by going to hostess clubs with their workplace friends, I recommend the book Nightwork: Sexuality, Pleasure, and Corporate Masculinity in a Tokyo Hostess Club by Anne Allison.

2000s until now: Maid Cafés

A new form of entertainment appeared as Japan entered the 2000s (or is it that new?). Maid cafés became popular in Japan in the early 2000s, with the first recognized maid café, Cure Maid Café, opening in Akihabara, Tokyo, in 2001. These cafés gained momentum due to their association with Akihabara's otaku culture, attracting fans of anime, manga, and video games.

Cure Maid café is known for its elegant and subdued Victorian-inspired atmosphere. Staff dress as maids provide polite, graceful service, adhering to a nostalgic and sophisticated concept reminiscent of European aristocratic households. Cure Maid Café serves a menu of teas, light meals, and desserts. It also collaborates with various animes (you can see this on its website), catering to both otaku culture enthusiasts and casual visitors seeking a unique and relaxing experience.

Just as how café waitresses changed during the 1920s, maid cafés also evolved into something more playful. As maid cafés gained popularity, they leaned into fantasy-based interactions. Maids began calling customers "master" and offering playful services like drawing on omelets with ketchup, playing card or video games, and performing dance routines. The variety also expanded. Variants such as school-themed cafes, tsundere cafés (maids acting cold at first and warming up later), and more, emerged. Many maids have become so popular that they even joined idol groups. For example, at-home café has made the first maid idol group.

Moreover, nowadays, not only men frequent those cafés but women customers also visit them. For example, there are quite a lot of girls online posting their experiences in the at-home cafe. Instead of calling girls master, maids would call them my lady (お嬢様). Of course, butler cafes with only male butlers (instead of maids) also showed up. For example, Swallowtail Butler Café is one of them.

As maid cafés became popular, maids also started to appear in other venues. For example, a maid taxi has been introduced in Kanazawa City. Customers can talk, play video games, or watch DVDs in the car with a maid. According to the website, most customers are men in their 30s.

The hourly pay in maid cafés is around 2000 yen (13 USD) per hour in Tokyo. The maids could work part-time in many of those cafés, so students might take those jobs as well.

Some Concluding Thoughts:

Many people think that maid cafés are purely products of the anime culture after the 1990s. However, it could be traced back to as early as the Edo period in the 1700s.

First, the slow evolution from a purely traditional Japanese style to a more Western style shows Japan went from a closed country to a country that integrates various global elements. What people find entertaining reflects the cultural changes in Japan.

Second, the changes and constants of women’s role in public spaces like teahouses, cafés, bars, and nightclubs also reflect changing gender norms, expectations, and agency in Japanese society. Although often exploited, women sometimes got rare opportunities to interact with intellectuals and professionals, thereby helping them improve their social mobility and participate in public spaces. Emerging host clubs and butler cafes for women also symbolize a divergence from the traditional gender role.

Lastly, the history of adult entertainment also illuminates the history of economic trends in Japan. The commercialization of adult services often goes hand-in-hand with economic booms, as people have extra money to entertain themselves. As demand for coffee and alcohol went up, cafés had to compete by hiring pretty waitresses. The purchase of positional goods like opening an expensive bottle of champagne with a club hostess also reflects the economic boom, as people seek to distinguish themselves in a wealthier society. Furthermore, the consumption of high-quality coffee and alcohol implicitly stemmed from international trade as well.

Today's maid cafés, while distinctly modern, represent the latest chapter in centuries of Japanese hospitality traditions. When you visit those maid cafés and enjoy their elaborate parfaits and omurice, don’t forget that you are participating in a tradition that has continuously adapted to reflect Japan's changing relationship with the world, gender roles, and economic development - a living testament to the dynamic nature of Japanese culture.

The Chinese Gaokao is often compared to the SAT. However, I have not seen a serious description of the Gaokao. In this article, I would like to give a concrete idea of how difficult the Chinese Gaokao is. This article will first introduce the Gaokao system and then analyze several test questions from the Gaokao in recent years.

The History of the Gaokao

The Gaokao, also known as the Chinese college entrance exam, stems from the Chinese civil service exam, which dates back to the 6th century during the Sui Dynasty. Before that, they used the nine-rank system which relied on a central Controller’s grading and recommendation of potential government officials. The nine-rank system had several defects such as corruption and low social mobility. Thus, the Sui Dynasty improved on the old system and was considered the start of the Chinese civil service exam.

Different dynasties had different rules on the civil service exam, but in general, there were several levels of exams to take. Successfully passing each level would grant the examinee a title. For example, in the Qing Dynasty, when the examinee passed 院试 or a prefectural exam, they gained the title “生员” and they officially entered the gentry scholars class, the highest-ranked class in ancient China.

The subjects in the Chinese civil service exam varied across dynasties as well. However, after the Ming Dynasty, it became more and more limited to Zhu Xi’s Four Books. Many students studied only the Four Books and Five Classics. It became more of a “teaching to the test” practice.

The legacy of the civil service exam remained after the Qing Dynasty. In 1938, the Nationalist government started the Gaokao, which was the first unified Gaokao of China. It only existed for three years, then it was disrupted by WWII. The first Gaokao after the CCP took power was in 1952 and then it was disrupted by political turmoil. It restarted in late 1977. The Ministry of Education standardized the test. Many legacies of the civil service exam can still be seen in today’s Gaokao. For example, today, if a student gets first place in the Gaokao in his province, he will be called xx province’s Gaokao Zhuangyuan. In the civil service examination, Zhuangyuan refers to the first place in the palace examination. For another example, today's Gaokao still balances regional differences, similar to how the civil service exam allocated quotas to each region.

The current Gaokao

Let’s get some facts about the Gaokao straight:

1. What subjects are there?

   The compulsory subjects are Language and Literature, Mathematics, and Foreign Language. Then each student can choose electives. Usually, they choose 3 electives from Political Science, History, Geography, Physics, Chemistry, and Biology.

2. When/where do students take the test?

   Unlike the American SAT, which happens multiple times a year, the Gaokao happens once per year, around June 7 - June 10. The Gaokao generally takes place at the location of the student's hukou registration or school registration. The municipal education bureau organizes the test centers, usually set up in local middle or high schools.

3. Is it a standardized test?

   No, it’s not a standardized test. There are several versions of the exam and different provinces have different rules for the Gaokao as well. For example, some regions use 3+3 mode, which means that in addition to Chinese, Mathematics, and Foreign Language, students must independently choose three subjects from Political Science, History, Geography, Physics, Chemistry, and Biology as their elective subjects for the Gaokao; some regions use the 3+1+2 mode, which means in addition to Chinese, Mathematics, and Foreign Language, students must choose one subject from Physics or History, and independently select two subjects from Political Science, Geography, Chemistry, and Biology as their elective subjects for the Gaokao. Some provinces share the same exam questions on certain subjects.

   The difficulty of the Gaokao varies across provinces. Here is a figure showing the difficulty of the Gaokao that has been circulating on the Chinese internet (I don’t know who made this, but if you did, please let me know and I’ll quote you properly). The easiest is red and as you go down the legend it gets harder. While I can’t verify the methodology that created this figure, I do intuitively think this figure makes sense. Firstly, impoverished areas usually have an easier version of the Gaokao — thus the western part of China is colored red and yellow. On the contrary, the eastern and southern parts of China have a harder version of the Gaokao because those areas are richer. Secondly, there are several regions that are known to have an advantage. For example, Beijing’s Gaokao is known to be easier, per online discussion.

   

   Lastly, minority groups also receive some advantages. According to the 2009 bonus point policy for the Gaokao introduced by the Ministry of Education, minority groups get 10 bonus points. But the bonus policy also varies across provinces and is gradually decreasing. The bonus points for minority groups also caused fervid online discussions as some people see it as unfair (this sparks similar debates to affirmative action in the United States, with the exception that AA is more legible in China).

4. How many points do I need to get into PKU/THU?

   This also depends on which province you are from, and what subjects you chose (humanities vs. science). Here is PKU’s 2023 admission cutoff score. Taking the province I grew up in, Zhejiang, as an example, it uses 3+3 mode. Each compulsory subject is worth 150 points, and each elective is worth 100 points, summing up to a total of 750 points. PKU’s admission cutoff for Zhejiang is a little above 700. So one can lose only less than 50 points across the tests. To give you an intuitive sense of what losing 50 points off is, the last math question shown below is worth about 20 points. So you can only afford to get 2-3 questions like that wrong to get in!

Gaokao Questions

Math

Now, Imagine you are a Chinese high school kid. You just learned the basic rules of the Gaokao, and now you are ready to study by looking at Gaokao Questions. You open up past test questions, and the first question on the math test is:

Not too bad, right?

Answer:

If you get this question wrong on the exam, your teacher will probably give you a side-eye.

You think this is easy, so you decide to jump to the last question that is known for being difficult. Here is the last question:

Hmm, this is… very difficult. The answer for Q1 is:

You should try to solve Q2 and Q3. I will not post the answers here for now, considering the length of this article.

And just a reminder, if your family is willing to send you abroad for college, here is the last (thus the hardest) question on SAT math:

Language and Literature

The language and literature exam starts with readings like this one:

The whole article is too long, I’m not going to post it here. Then, for the questions:

So it kind of looks like the SAT’s reading part. But the SAT does not have prompts like “President Biden has visited xxx and done xxx, please analyze what is the importance.”

You will also be asked to analyze ancient Chinese texts:

I always find this part quite hard (but actually very useful, now that I work on economic history!)

The last question is usually an essay:

This part looks like SAT writing in my opinion. Of course, SAT writing usually doesn’t ask this kind of essay questions that glorify America:

English

Finally, you have to study a foreign language as well. The English exam contains a listening test and a writing test. Here is an example:

SAT does not require a foreign language, but I have taken Chinese and Japanese AP tests during high school. My impression is that they are at similar levels.

Conclusion

I presented here a picture of the Gaokao system and the difficulty of the Gaokao. On the one hand, I think the Gaokao is really competitive and it’s a pity to spend one’s youth preparing for an exam. Of course, this is partly due to China being a big country and the legacy of the civil service exam (it was always thought of as a fair way to change one’s social class).

On the other hand, I think there are many things from the Gaokao that the American SAT should learn from. For one, I do think the SAT’s math part is too easy. Among reasonably smart people, it’s not a test of your math ability but whether you are good at not making silly arithmetic mistakes. I got a full score on both SAT and GRE on the first try without spending time practicing the math part and so did many of my friends. I would appreciate it if the test could be more selective on pure math skills than arithmetic.

Overall, I think the Gaokao as a system has tried its best to maintain a relatively fair competition in a large country like China. The bonus points for minorities are also well-thought-through and provide a concrete idea of the affirmative action they get (in contrast to American affirmative action which is more to the school’s own discretion).

Lastly, to all the seniors everywhere who just finished high school, have a fun summer!

Recently, Chinese people are passionate about an old case in 1994 — the Thallium poisoning of Zhu Ling:

Zhu Ling^[a] (Chinese: 朱令; pinyin: Zhū Lìng, 24 April 1973 – 22 December 2023) was best known as the victim of an unsolved 1995 thallium poisoning case in Beijing, China. Her symptoms were posted to the Internet via a Usenet newsgroup by her friend from Peking University, Bei Zhicheng, and were subsequently proven to be caused by thallium poisoning. Her case was then reviewed by physicians in many different countries who examined her symptoms and made suggestions as to diagnoses and treatment. This effort was recognized as the first large-scale tele-medicine trial.^[6] Her life was ultimately saved, but she suffered serious neurological damage along with permanent physical impairment, and died in December of 2023.

In 2005, Bei Zhicheng’s internet article pointed out that the suspect, who’s also Zhu Ling’s roommate, Sun Wei, is very likely guilty. His narrative won over Chinese netizens’ trust, especially given the fact that this hotly discussed case has become a high-profile cold case that the police stopped investigating for seemingly no reason. People believe that the police stopped investigating because the suspect’s family is powerful—Sun’s grandfather held a high position in the government.

Upon Zhu Ling’s recent death, Chinese netizens’ anger culminated. Almost 20 years later, today, people have been voluntarily investigating, reposting, and signing petitions for the case. Chinese netizens tried hard so that the Australian government and media would also pay attention to this and banish the suspect, who is now living in Australia. In 2005, this case marked one of the first times that Chinese netizens’ sought justice online against an unfair judgment in their mind, and almost 20 years later, people are still using this powerful tool to voice their thoughts about this case. This article describes some well-known cases of Chinese netizens’ seeking justice online and then analyzes why does the CCP allow such behavior.

Background

With many newer and easier-to-use social media such as Weibo and Xiaohongshu rising in popularity, Chinese netizens have begun to seek justice online. It has become one of the easiest ways to expose and attract attention to corruption, collusion, and unfair treatment.

A famous example is Zhong Guangwei’s case. Zhong Guangwei was a miner from Shanxi. The mining work caused him to develop pneumoconiosis in 2008. Although he won the case against the mine owner, the mine owner refused to pay the promised compensation even 4 months after the judgment became effective in September, 2010. He eventually resorted to Weibo, and his post attracted the attention of Deng Fei, a journalist for "Phoenix Weekly." After Deng’s interview with Zhong, many Chinese netizens helped him — people visited him in Shanxi and some netizens even sent him to a hospital with better treatments in Nanjing.

Zhong passed away in March 2022. Before that, in February 2022, he reposted Deng Fei’s post about the “Xuzhou chained woman incident” which was another famous case of justice sought online. Zhong replied “Thank you, Teacher Deng, for speaking out for the chained woman! If it weren't for your initial support in my case, the consequences would be unimaginable.” This post is like a torch, conveying the warmth and enthusiasm that netizens have extended to him to the next victim of unfair treatment.

I have been describing netizens who more or less described the truth and exposed the injustice. This presents the readers with the bright side of seeking justice online. Nevertheless, as one would expect, many people could distort the truth and incite netizens to achieve their goals (e.g. the case of Liang Ying and Luo Guanzhong).

Why does the Chinese Government Allow this?

You may wonder why is such behavior allowed on the Chinese Internet by CCP. That is a fair point—exposing the cases online usually damages the reputation of the party. The fast-spreading posts on Weibo make sure more people will see the unfair treatment, resulting in criticism against the government across different cities and even across countries. For example, Zhong Guangwei’s case described above makes people reprimand the court for inaction; Zhu Ling’s case even more so destroyed netizen’s trust in the legal system — many even used VPN to send posts criticizing the CCP for the injustice in this case on X. In fact, in 2013, “Chinese Americans have launched a petition on the White House website, demanding that U.S. authorities investigate and deport Sun Wei, who is suspected of causing Zhu Ling's thallium poisoning.”

The reason that the CCP allows occasional online justice-seeking is that it is a decent way to monitor local governments. In a vast land like China, people fighting against local corruption might have difficulty conveying the matter to a higher-level government. As a result, many local problem go untold, like a timed bomb that could explode into some terrible scandal at some point. The central government would like to get more information to spot the local problems and potentially solve them early before the matter grows into something uncontrollable. In short, the central government has to balance between the potential of instability and collective action and knowing too little about the local governments’ performances (there are many economics/political science papers on this trade-off; to list a few: Egorov, Guriev, and Sonin (2009), Lorentzen (2014), Qin, Stromberg, and Wu (2017))

What are the alternatives to getting information from the Internet? One way is through the petitioning system. Since ancient times, petitioners who encountered unjust treatment by local powers would travel to Beijing and tell their grievances. Nowadays, there are online services that petitioners can utilize, but every year, thousands of people still choose to go to Beijing for their cases. However, such methods are usually stringently monitored and brutally hindered by the local government. For example, in 2013, “A Beijing court has issued a verdict against 10 defendants who acted as "black security guards" in Beijing, capturing and detaining petitioners from Henan.” Sometimes, when petitioners return to their hometowns, the local government punishes them by sending them to re-education camps. And with those hindrances of the local government, it is hard for the voice of petitioners to actually go through to the central government. Most of the time, the petitioners are just so discouraged that they give up the idea of petitioning in the capital.

Thus, seeking justice online becomes an acceptable strategy for both the petitioners and the central government. For the petitioners, they circumvent the corrupt local government that wronged them by posting directly on social media that the whole nation’s netizens including central government officials can see. For the central government, it gets to monitor the local governments’ behavior closely, preventing large-scale collective action later. The Internet essentially serves as a surveillance tool in this case. For example, in June of 2023, a video showing a man holding hands with a young woman prevailed on the internet. Netizens uncovered the identities of the individuals featured in the video and found out that the man is a high-up in the SOE, PetroChina. This was scandalous because this man was married to someone else. Consequently, "according to the Disciplinary Regulations and the Law on Public Service Sanctions, Hu Jiyong [the man in the video] is given the punishment of expulsion from the Party and dismissal from public office."

To sum up, I present a short history of Chinese people seeking justice online. Chinese petitioners often resort to the Internet to vent their dissatisfaction with certain judgments or treatments. The central government allows portions of such posts to exist because it utilizes them to surveil local issues, even though the posts do create a negative image of the CCP.

"At 4 p.m. on November 26th, at the emergency triage desk of the Children's Hospital affiliated with the Capital Institute of Pediatrics, the handwritten call number list had reached over five hundred. The crowd queuing at the triage desk blocked the entrance and exit of the emergency department. Family members, unable to find a spot, were seen navigating through the crowd with one hand holding up an IV bottle and the other leading a child, searching for a place to settle. Almost every corner of the outpatient hall was occupied."

I will argue that this social media narrative is mostly false, China’s number of pediatricians is much better than it was 5 years ago. Instead, there is a temporary shortage because of the huge amount of children sick with respiratory diseases.

Background

This winter, many Chinese children and families are affected by respiratory infections. There have been posts on Chinese social media indicating the absurdly long lines at the hospital this winter. One mother posted on Xiaohongshu that her child’s “9 p.m. appointment was not seen until 7 a.m. the next morning.” There are also posts claiming that pediatricians are quitting their jobs due to the long hours. Consequently, people on Chinese social media are pessimistic about nurturing a child — they are worried that there aren’t enough resources, including basic medical resources such as pediatricians.

Pediatrics in China

Firstly, I want to point out that the situation of Chinese pediatrics is not as pessimistic as people on Chinese social media thought and it’s certainly not as bad as several years before. In general, the children (under 14)-to-pediatrician ratio is decreasing. Chinese government is well-aware that they need better medical systems for children so that parents are at ease for giving birth. This is a necessary step for encouraging people to have children and thus remedy the decreasing population. For one, in 2016, National Health and Family Planning Commission published “Opinions on Strengthening the Reform and Development of Children's Medical and Health Services,” stating several policies that would help increase the number of pediatricians. For one, "Starting from 2016, integrated pediatrician training programs have been implemented in 39 universities conducting '5+3' integrated medical education."

(Data from China Statistical Yearbook and World Bank)

Still, 1212 children to 1 pediatrician certainly is not a satisfactory number. Developed countries like the U.S. usually have a number < 1000. I just want to point out that the panic of not having enough pediatricians is resulted mostly from the sudden demand due to the respiratory disease targeting children than a drop in the supply of pediatricians.

Another qualitative evidence is from my friend, who is a doctor in China. He made several points:

• There IS a shortage of pediatricians the past fall, due to the respiratory disease.

• The situation is somewhat alleviated at hospitals with a standardized training system because trainees could be used when there is a shortage of clinical doctors.

• There always has been a shortage of pediatricians. In fact, the situation has been better these recent few years.

To sum up, I think that China has a shortage of pediatricians, but it’s not as dire as Chinese social media states. But I understand why people are freaking out about the shortage at this point—probably it has to do with the polarized opinions toward child-bearing and nurturing in China now.

Subscribe now