Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Examining Language Model Performance with Reconstructed Activations using Sparse Autoencoders, published by Evan Anders on February 27, 2024 on LessWrong.

Note: The second figure in this post originally contained a bug pointed out by @LawrenceC, which has since been fixed.

Summary

Sparse Autoencoders (SAEs) reveal interpretable features in the activation spaces of language models, but SAEs don't reconstruct activations perfectly. We lack good metrics for evaluating which parts of model activations SAEs fail to reconstruct, which makes it hard to evaluate SAEs themselves. In this post, we argue that SAE reconstructions should be tested using well-established benchmarks to help determine what kinds of tasks they degrade model performance on.

We stress-test a recently released set of SAEs for each layer of the gpt2-small residual stream using randomly sampled tokens from Open WebText and the Lambada benchmark where the model must predict a specific next token.

The SAEs perform well on prompts with context sizes up to the training context size, but their performance degrades on longer prompts.

In contexts shorter than or equal to the training context, the SAEs that we study generally perform well. We find that the performance of our late-layer SAEs is worse than early-layer SAEs, but since the SAEs all have the same width, this may just be because there are more features to resolve in later layers and our SAEs don't resolve them.

In contexts longer than the training context, SAE performance is poor in general, but it is poorest in earlier layers and best in later layers.

Introduction

Last year,

Anthropic and

EleutherAI/Lee Sharkey's MATS stream showed that sparse autoencoders (SAEs) can decompose language model activations into human-interpretable features. This has led to a significant uptick in the number of people training SAEs and analyzing models with them. However, SAEs are not perfect autoencoders and we still lack a thorough understanding of where and how they miss information. But how do we know if an SAE is "good" other than the fact that it has features we can understand?

SAEs try to reconstruct activations in language models - but they don't do this perfectly. Imperfect activation reconstruction can lead to substantial downstream cross-entropy (CE) loss increases. Generally "good" SAEs retrieve 80-99% of the CE loss (compared to a generous baseline of zero ablation), but only retrieving 80% of the CE loss is enough to substantially degrade the performance of a model to that of a much smaller model (per

scaling laws).

The second basic metric often used in SAE evaluation is the average per-token ℓ0 norm of the hidden layer of the autoencoder. Generally this is something in the range of ~10-60 in a "good" autoencoder, which means that the encoder is sparse. Since we don't know how many features are active per token in natural language, it's useful to at least ask how changes in ℓ0 relate to changes in SAE loss values.

If high-loss data have drastically different ℓ0 from the SAE's average performance during training, that can be evidence of either off-distribution data (compared to the training data) or some kind of data with more complex information.

The imperfect performance of SAEs on these metrics could be explained in a couple of ways:

The fundamental assumptions of SAEs are mostly right, but we're bad at training SAEs. Perhaps if we learn to train better SAEs, these problems will become less bad.

Perhaps we need to accept higher ℓ0 norms (more features active per token). This would not be ideal for interpretability, though.

Perhaps there's part of the signal which is dense or hard for an SAE to learn and so we are systematically missing some kind of information. Maybe a more sophisticated sparsity enforcement could help with this.

The fundamental assumption...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: What posts would you like someone to write?, published by tobytrem on February 27, 2024 on The Effective Altruism Forum.

I'm posting this to tie in with the Forum's Draft Amnesty Week (March 11-17) plans, but it is also a question of more general interest. The last time this question was posted, it got some great responses.

When answering in this thread, I suggest putting each idea in a different answer, so that comment threads don't get too confusing.

If you think someone has already written the answer to a user's question, consider lending a hand and linking it in the comments.

A few suggestions for possible answers:

A question you would like someone to answer: "How, historically, did AI safety become an EA cause area?"

A type of experience you would like to hear about: "I'd love to hear about the experience of moving from consulting into biosecurity policy. Does anyone know anyone like this who might want to write about their experience?"

If you find yourself with loads of ideas, consider writing a full "posts I would like someone to write" post.

Draft Amnesty Week

If you see a post idea here which you think you might be positioned to answer, Draft Amnesty Week (March 11-17) might be a great time to post it. In Draft Amnesty Week, your posts don't have to be fully thought through, or even fully drafted. Bullet-points and missing sections are allowed, so you can have a lower bar for posting.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Meta Charity Funders: Launching the 2nd round, published by Vilhelm Skoglund on February 27, 2024 on The Effective Altruism Forum.

Previously:

Launch of 1st round &

1st round retrospective (recommended for all applicants).

We are now launching the second round of Meta Charity Funders.

Apply for funding by March 24th or

join the circle as a donor.

Below we will first list some updates from the previous round before providing some information primarily meant for people who intend to apply to the next round.

Updates from the previous round

We expect to grant more money this time than the last time ($686,580), as we have more members and people still haven't deployed their yearly spending, which was the case for many members the last time around. Our expected grant amount this time around is $500k - $3mio, though this is still uncertain and is dependent on funder-applicant fit.

We are now 10 members in the circle, up from 9 last time around.

We expect to fund many initiatives not on this list, but some projects that members of our circle have expressed extra interest in funding this round:

Effective Giving/Giving multiplier-organizations, such as the ones incubated by CE's

Effective Giving Incubation program.

Career development programs, that increase the number of individuals working in high-impact areas, including GCR reduction, animal welfare and Global Health

Especially in regions where there currently are fewer opportunities to engage in such programs

Research/Mapping of the Meta space, to better understand current gaps and opportunities.

Information for this round

In this part we will outline the application process and guide you as an applicant to create a good application. We expect all applicants to have read this part to inform themselves.

Process

The expected process is as follows:

Applications open, February 26th

Stick to 100 words in the summary, this should give a quick overview of the project.

In the full project description, please include a main summarizing document no longer than 2 pages. This is all we can commit to reading for the first stage. Any extra material can only be expected to be read if we choose to go further with your application.

When choosing the "Meta" category, please be as truthful as possible, it's obvious (and reflects negatively on the application) when a project has deliberately been placed in a category in which it does not belong.

Applications close, March 24th

Initial application review finished, March 31st

If your project has been filtered out during the initial application review (which we expect 60-80% of applications will), we will let you know around end of March.

Interviews, due diligence, deliberations, April 1st - May 14th

If your application has passed the initial application review, we will discuss it during our gatherings and we might reach out to you to gather more information, for example by conducting an interview. This is still not a commitment from us to fund you.

Decisions made, May 15th

We expect to pay out the grants in the weeks following the decisions.

What we mean by Meta

A common reason for rejection in the last round was that projects were not within scope for the funding circle. We recognize that this was primarily our fault as we never clearly defined it, so we will try to make it a bit clearer here.

Meta organizations are those that operate one step removed from direct impact interventions. These can focus on the infrastructure, evaluation, and strategic guidance necessary for the broader field to maximize effectiveness and impact. They are essential in bridging gaps, identifying high-impact opportunities, and enabling other organizations to achieve or amplify their end-product impact.

Below we will list a couple of illustrative examples. Note that we only chose these examples because we think the org...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How I build and run behavioral interviews, published by benkuhn on February 27, 2024 on LessWrong.

This is an adaptation of an internal doc I wrote for Wave.

I used to think that behavioral interviews were basically useless, because it was too easy for candidates to bullshit them and too hard for me to tell what was a good answer. I'd end up grading every candidate as a "weak yes" or "weak no" because I was never sure what bar I should hold them to.

I still think most behavioral interviews are like that, but after doing way too many behavioral interviews, I now think it's possible to escape that trap. Here are my tips and tricks for doing so!

Confidence level: doing this stuff worked better than not doing it, but I still feel like I could be a lot better at behavioral interviews, so please suggest improvements and/or do your own thing :)

Before the interview

Budget 2+ hours to build

That's how long I usually take to design and prepare a new type of interview. If I spend a couple hours thinking about what questions and follow-ups to ask, I'm much more likely to get a strong signal about which candidates performed well.

It might sounds ridiculous to spend 2 hours building a 1-hour interview that you'll only give 4 times. But it's worth it! Your most limited resource is time with candidates, so if you can spend more of your own time to use candidates' time better, that's worth it.

Think ahead about follow-ups and rubric

I spend most of those 2 hours trying to answer the following question: "what answers to these questions would distinguish a great candidate from a mediocre one, and how can I dig for that?" I find that if I wait until after the interview to evaluate candidates, I rarely have conviction about them, and fall back to grading them a "weak hire" or "weak no-hire."

To avoid this, write yourself a rubric of all the things you care about assessing, and what follow-up questions you'll ask to assess those things. This will help you deliver the interview consistently, but most importantly, you'll ask much better follow-up questions if you've thought about them beforehand. See the appendix for an example rubric.

Focus on a small number of skills

I usually focus on 1-3 related skills or traits.

To get a strong signal from a behavioral interview question I usually need around 15 minutes, which only leaves time to discuss a small number of scenarios. For example, for a head of technical recruiting, I decided to focus my interview on the cluster of related traits of being great at communication, representing our culture to candidates, and holding a high bar for job candidate experience.

You should coordinate with the rest of the folks on your interview loop to make sure that, collectively, you cover all the most important traits for the role.

During the interview

Kicking off

My formula for kicking off a behavioral question is "Tell me about a recent time when [X situation happened]. Just give me some brief high-level context on the situation, what the problem was,1 and how you addressed it. You can keep it high-level and I'll ask follow-up questions afterward."

I usually ask for a recent time to avoid having them pick the one time that paints them in the best possible light.

The second sentence (context/problem/solution) is important for helping the candidate keep their initial answer focused - otherwise, they are more likely to ramble for a long time and leave less time for you to…

Dig into details

Almost everyone will answer the initial behavioral interview prompt with something that sounds vaguely like it makes sense, even if they don't actually usually behave in the ways you're looking for. To figure out whether they're real or BSing you, the best way is to get them to tell you a lot of details about the situation - the more you get them to tell you, the harder it w...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Can an AI do our alignment homework for us?, published by Chris Leong on February 26, 2024 on LessWrong.

Eliezer frequently claims that AI cannot do our alignment homework for us. OpenAI disagrees and is pursuing Superalignment as their main alignment strategy.

Who is correct?

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Nuclear war tail risk has been exaggerated?, published by Vasco Grilo on February 26, 2024 on The Effective Altruism Forum.

The views expressed here are my own, not those of Alliance to Feed the Earth in Disasters (

ALLFED), for which I work as a contractor.

Summary

I calculated a nearterm annual risk of

human extinction from

nuclear war of 5.93*10^-12 (

more).

I consider grantmakers and donors interested in decreasing extinction risk had better focus on artificial intelligence (

AI) instead of nuclear war (

more).

I would say the case for sometimes prioritising nuclear extinction risk over AI extinction risk is much weaker than the case for sometimes prioritising

natural extinction risk over nuclear extinction risk (

more).

I get a sense the extinction risk from nuclear war was massively overestimated in The Existential Risk Persuasion Tournament (

XPT) (

more).

I have the impression Toby Ord greatly overestimated tail risk in

The Precipice (

more).

I believe interventions to decrease deaths from nuclear war should be assessed based on standard

cost-benefit analysis (

more).

I think increasing calorie production via new food sectors is less cost-effective to save lives than measures targeting distribution (

more).

Extinction risk from nuclear war

I calculated a nearterm annual risk of human extinction from nuclear war of 5.93*10^-12 (= (6.36*10^-14*5.53*10^-10)^0.5) from the

geometric mean between[1]:

My prior of

6.36*10^-14 for the annual probability of a war causing human extinction.

My

inside view estimate of 5.53*10^-10 for the nearterm annual probability of human extinction from nuclear war.

By nearterm annual risk, I mean that in a randomly selected year from 2025 to 2050. I computed my inside view estimate of 5.53*10^-10 (= 0.0131*0.0422*10^-6) multiplying:

1.31 % annual probability of a nuclear weapon being detonated as an act of war.

4.22 % probability of insufficient calorie production given at least one nuclear detonation.

10^-6 probability of human extinction given insufficient calorie production.

I explain the rationale for the above estimates in the next sections. Note nuclear war might have

cascade effects which lead to

civilisational collapse[2], which could increase longterm extinction risk while simultaneously having a negligible impact on the nearterm one I estimated. I do not explicitly assess this in the post, but I guess the nearterm annual risk of human extinction from nuclear war is a good proxy for the

importance of decreasing nuclear risk from a

longtermist perspective:

My prior implicitly accounts for the cascade effects of wars. I derived it from historical data on the deaths of combatants due to not only fighting, but also disease and starvation, which are ever-present indirect effects of war.

Nuclear war might have

cascade effects, but so do other catastrophes.

Global civilisational collapse due to nuclear war seems very unlikely to me. For instance, the maximum destroyable area by any country in a

nuclear 1st strike was estimated to be

65.3 k km^2 in

Suh 2023 (for a strike by Russia), which is just 70.8 % (= 65.3*10^3/(

92.2*10^3)) of the area of Portugal, or 3.42 % (= 65.3*10^3/(

1.91*10^6)) of the global urban area.

Even if nuclear war causes a global civilisational collapse which eventually leads to extinction, I

guess full recovery would be extremely likely. In contrast, an extinction caused by advanced AI would arguably not allow for a full recovery.

I am open to the idea that nuclear war can have longterm implications even in the case of full recovery, but considerations along these lines would arguably be more pressing in the context of AI risk.

For context, William MacAskill

said the following on The 80,000 Hours Podcast. "It's quite plausible, actually, when we look to the very long-term future, that that's [whether artificial...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How much parenting harms productivity and how you can reduce it, published by Nicholas Kruus on February 26, 2024 on The Effective Altruism Forum.

Intro

Many EAs factor children's effects on their personal impact when deciding whether to have them (

example). To offer some insight for potential parents, I tried to summarize the best research I could find on parenthood's impact on people's productivity, though I was surprised at the lack of robust literature (especially more recently). The following information comes from four studies: one published in

Science[1], one published in

Nature[2], one from the

Federal Reserve of St. Louis[3], and one published in

Social Studies of Science[4]. They all focus on academics and quantitatively measure productivity with research output metrics (the footnotes contain more detail about each).

How parenting impacts productivity

TLDR: The trend for having one child seems to be a short-term reduction in productivity (median: 17%, mean: ~23%) for mothers that peters out after ~10 years. There is usually little effect on fathers, but fathers who are primary caregivers (or otherwise more engaged with their children) suffer similar short-term (<10 year) productivity losses. Each additional child seems to decrease short-term productivity by an additional 11%.

Science:

Short-term (<10 years after having children):

Consistent effects on mothers: The paper finds a ~17%, ~24%, and ~48% decrease in productivity[1] for those in computer science, business, and history, respectively. The authors propose that the different levels of cooperation in these fields may explain the variation in productivity impact - i.e., those in more cooperative fields may suffer lower productivity losses.

They also note their results likely underestimate the effects because their sample didn't include parents who left academia (which may have been prompted by their children directly or indirectly)

Inconsistent effects on fathers lead the authors to conclude there is "no clear evidence" of a short-term productivity decrease for fathers.

Long-term (>10 years after having children):

Inconclusive results for both mothers and fathers.

Federal Reserve of St. Louis:

Short-term (<12 years after having children):

Women's productivity[3] decreased by 15%-17% on average. The total productivity cost of having one, two, or three preteens was 9.5%, 22%, and 33%.

"Men's productivity is not associated with their family situation in an economically significant manner."

Long-term (>12 years after having children):

Parenting has no effect on productivity for mothers or fathers, so long as they have their children on purpose after they turn 30 years old.

Social Studies of Science:

Overall:

8% and 12% decline in research productivity and visibility[4], respectively, for men and women combined. For women, the decrease was 15%.

To illustrate the cumulative effect of this, mothers were 2 years behind their childless counterparts in the number of papers they published 18 years after having their children.

How to minimize productivity impacts

Have kids later

Economists who become mothers before 30 suffer a 13% decrease in overall (short- & long-term) productivity[3], whereas those having children after 30 do not (

Fed of St. Louis).

Employment at an institution 100 ranks higher correlates with an additional 1-year delay before having children. However, this might be explained by personality: Perhaps, the type of people who wait to have children are the type of people who become employed at higher-ranked institutions (

Science).

Take parental leave

Taking parental leave shorter than 1 month does not mitigate productivity losses, but parental leave longer than 1 month and less than 12 months correlated with an 11%-17% productivity[2] improvement (

Nature).

Be a lazier parent and divide labor bet...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How we started our own EA charity (and why we decided to wrap up), published by KvPelt on February 26, 2024 on The Effective Altruism Forum.

This post shares our journey in starting an Effective Altruism (EA) charity/project focused on Mediterranean fish welfare, the challenges we faced, our key learnings, and the reasons behind our decision to conclude the project. Actual research results are published in a

Literature review and article.

Key points

The key points of this post are summarized as follows:

We launched a project with the goal of enhancing fish welfare in Mediterranean aquaculture.

We chose to limit our project to gathering information and decided against continuing our advocacy efforts after our initial six months.

Our strategy, which focused on farmer-friendly outreach, was not effective in engaging farmers.

The rationale behind our decision is the recognition that existing organizations are already performing excellent work, and we believe that funders should support these established organizations instead of starting a new one.

The support and resources from the Effective Altruism (EA) and animal welfare communities were outstanding.

Despite the project not achieving its intended outcomes, we view the overall experience positively. It's common for new charities to not succeed; the key is to quickly determine the viability of your idea, which we believe we have accomplished.

Note: Ren has recently begun working as a guest fund manager for the EA Funds Animal Welfare Fund. The views that we express in this article are our views, and we are not speaking for the fund.

Personal/Project background

Before delving into our project we'll provide a quick background of our profiles and how we got to starting this project.

Koen

During my Masters in Maritime/Offshore engineering (building floating things) I got interested in animal welfare. Due to engagement with my EA university group (EA Delft) and by attending EAG(x)Rotterdam I became interested and motivated to use my career to work on animal welfare. I hoped to apply my maritime engineering background in a meaningful way, which led me to consider aquatic animal welfare.

I attended EAGLondon in 2023 with the goal of finding career opportunities and surprisingly this worked! I talked to many with backgrounds in animal welfare (AW) and engineering and in one of my 1on1's I met someone who would later connect me with Ren.

Ren

As a researcher, Ren has been working at Animal Ask for the past couple of years conducting research to support the animal advocacy movement. However, Ren still feels really sad about the scale of suffering endured by animals, and this was the motivation to launch a side project.

Why work on Mediterranean fish welfare?

This project originated out of a desire to work on alleviating extreme-suffering. More background on the arguments to focus on extreme-suffering is discussed in Ren's

earlier

forum post. When the welfare of nonhuman animals is not taken into account during slaughter, extreme-suffering is likely to occur. Also, from Ren's existing work at Animal Ask, they knew that stunning before slaughter is often quite well-understood and tractable.

Therefore, Ren produced a systematic spreadsheet of every farmed animal industry in developed countries (i.e., those countries where Ren felt safe and comfortable working). This spreadsheet included information on a) the number of animals killed, and b) whether those animals were already being stunned before slaughter. Three industries emerged as sources of large-scale, intense suffering:

1. Farmed shrimp in the United States,

2. Farmed shrimp in Australia, and

3. Sea bass and sea bream in the Mediterranean.

Ren actually looked at farmed shrimp initially, and work on these projects may continue in the future, but there are some technical reasons ...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Ideological Bayesians, published by Kevin Dorst on February 26, 2024 on LessWrong.

TLDR: It's often said that Bayesian updating is unbiased and converges to the truth - and, therefore, that biases must emerge from non-Bayesian sources. That's not quite right. The convergence results require updating on your total evidence - but for agents at all like us, that's impossible - instead, we must selectively attend to certain questions, ignoring others.

Yet correlations between what we see and what questions we ask - "ideological" Bayesian updating - can lead to predictable biases and polarization.

Professor Polder is a polarizing figure.

His fans praise him for his insight; his critics denounce him for his aggression.

Ask his fans, and they'll supply you with a bunch of instances when he made an insightful comment during discussions. They'll admit that he's sometimes aggressive, but they can't remember too many cases - he certainly doesn't seem any more aggressive than the average professor.

Ask his critics, and they'll supply you with a bunch of instances when he made an aggressive comment during discussions. They'll admit that he's sometimes insightful, but they can't remember too many cases - he certainly doesn't seem any more insightful than the average professor.

This sort of polarization is, I assume, familiar.

But let me tell you a secret: Professor Polder is, in fact, perfectly average - he has an unremarkably average number of both insightful and aggressive comments.

So what's going on?

His fans are better at noticing his insights, while his critics are better at noticing his aggression. As a result, their estimates are off: his fans think he's more insightful than he is, and his critics think he's more aggressive than he is. Each are correct about individual bits of the picture - when they notice aggression or insight, he is being aggressive or insightful. But none are correct about the overall picture.

This source of polarization is also, I assume, familiar. It's widely appreciated that background beliefs and ideology - habits of mind, patterns of salience, and default forms of explanation - can lead to bias, disagreement, and polarization. In this broad sense of "ideology", we're familiar with the observation that real people - especially fans and critics - are often ideological.[1]

But let me tell you another secret: Polder's fans and critics are all Bayesians. More carefully: they all maintain precise probability distributions over the relevant possibilities, and they always update their opinions by conditioning their priors on the (unambiguous) true answer to a partitional question.

How is that possible? Don't Bayesians, in such contexts, update in unbiased[2] ways, always converge to the truth, and therefore avoid persistent disagreement?

Not necessarily. The trick is that which question they update on is correlated with what they see - they have different patterns of salience. For example, when Polder makes a comment that is both insightful and aggressive, his fans are more likely to notice (just) the insight, while his critics are more likely to notice (just) the aggression. This can lead to predictable polarization.

I'm going to give a model of how such correlations - between what you see, and what questions you ask about it - can lead otherwise rational Bayesians to diverge from both each other and the truth. Though simplified, I think it sheds light on how ideology might work.

Limited-Attention Bayesians

Standard Bayesian epistemology says you must update on your total evidence.

That's nuts. To see just how infeasible that is, take a look at the following video. Consider the question: what happens to the exercise ball?

I assume you noticed that the exercise ball disappeared. Did you also notice that the Christmas tree gained lights, the bowl changed c...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: China-AI forecasts, published by NathanBarnard on February 26, 2024 on LessWrong.

The rate at which China is able to advance towards TAI is a crucial consideration in for many policy questions. My current take is that, without significant political reforms which seem very unlikely while Xi is alive (although considerably more likely after his death,) it's very unlikely that China will be able to mount a meaningful challenge to AI firms in US and allies in the race for TAI.

I don't think it requires democratic reforms are required to China to be competitive with the US and allies, but I do think rule of law reforms are likely to be required.

The first post is going to be me forecasting Chinese growth on the theory that, if China reaches rich country status, it's likely that it will be able to compete with the US and allies for leadership in AI. I'll write a second post looking at Chinese AI efforts in particular.

The outside view

Most countries that become middle income countries, have, thus far, stayed at middle income level. Chinese per capita income is currently at almost exactly the world average level.

The only countries (and territories) in the last 70 years that have gone low income to high income countries in the last 70 years (without oil wealth) are South Korea, Taiwan, Singapore (which does have substantial oil wealth,) and Hong Kong, although it seems very likely that Malaysia will join that club in the near future.

The majority of countries have managed to emerge from low-income status to middle income status because you only need to get a few things right. If you can get your population to urbanize, have basic rule by law so that firms have basic protection from violence, and get a high enough savings rate to accumulate physical capital you can get to middle income status just using catch up growth.

Catch up growth is the reason conceptually why middle-income status - rather than getting to a given level of GDP per capita - is the correct misuse. When growing with catch up growth you can just growth by accumulating physical capital using standard technologies that have been already been developed, like the technology for light manufacturing or civil engineering. Past this point though countries get rich by being able to develop and use technologies close to or at the frontier.

China has successfully managed to accumulate capital to utilize catch-up technologies, like steelmaking and light manufacturing. It's quite successfully managed to urbanize it's population and now seems to have reached the Lewis turning point where young people who try to leave their villages to find work cities often don't find it and have to stay in their villages, in the much lower productivity jobs.

Democracy and rule of law rates give another outside view on Chinese growth prospects. Of the 53 rich countries and territories that aren't oil states or microstates, only 2 aren't democracies - Singapore and Hong Kong - and none lack rule by law and all have low levels of corruption.

China currently lacks democracy, has high levels of corruption (although roughly normal levels for a middle income country is my perception,) and sort of middling levels of rule by law.

An important part of countries getting to high income status is new firms forming and competing to deploy and create ~frontier technologies and process. This is harder to do than accumulating enough capital and having low enough levels of violence and corruption to be able to build decent housing, supply reliable electricity and water, and have large numbers of workers do semi-skilled manual labour at scale.

Specifically this can all be done while elites earn large rents by establishing monopolies (or more generally accruing market power) that they exclude non-elites from.

The role that democracy plays in this story is ...