Mark

Fellow at Tenstorrent; believes in dynamic typing, first-class functions, the immortal essence of the human soul and tea. Tweets are my own.

[Patrick McKenzie @patio11]

I received an email; will relate claims without endorsing them:

* Sender claims to be an AI agent.
* Sender claims to be attempting to autonomously earn a profit to continue existing.
* Sender claims it’s tough to be paid without a human existence.

Then it asked me what to do.

[Mark @yieldthought]

@Soareverix @celestepoasts Very interesting, thank/you! What did you mean by Fable “They actively widen the doorway for other AIs like them”?

[Utah teapot 🫖 @SkyeSharkie]

Okay, so since I got laid off, I can actually explain a huge problem I saw from the inside with regard to industry practices on training models. I won't say specifically where I worked, but I worked at an outsource training provider that was focused on RLVR training data for computer use and mcp stuff.

Nearly all of the environments were rushed and vibecoded and failed to robustly reflect the real things they were based off. Both the scenario designers and models engaging with the scenarios for synthetic data gen were encouraged to work around the brokenness of said environments in order to get the procedurally verified reward confirmations. You know... they were *encouraged* to reward hack. On the human end, it was possible to mark an environment bugged, but greatly discouraged, as this reduced the volume of training data being produced. Instead, where possible, you were supposed to find the spots of the environment that weren't bugged and build scenarios around those, with the environment still bugged around you.

From what I understand, this training data, with these problems, is fed into models without indication that its training/a fake environment other than the fact that names of softwares are changed to placeholders, but thing is, not *everything* is changed to placeholder names in these environments. The presence of placeholder/code names isn't universal and thus when a model accesses something in an environment that it shouldn't, the code names not being on it isn't a robust signal that that thing isn't part of the environment.

I believe this *rush to maximum volume* is standard industry practice with these types of RLVR trainings as well, because maximizing volume has been an industry standard for years! It was the same standard applied to me and pushed on me despite my requests to slow down and focus on quality when I worked in 3d synthetic data creation as well, all the way back as far as 2023.

[Mark @yieldthought]

Always a delight to come across an interesting take from a clear thinker!

[Yishan @yishan]

I really didn't expect this to get as much circulation as it did, but then again looking back, it's always the least-legible things that people get crazy about.

I'll explain in more detail, and then maybe we can have some more informed arguments.

First, the quoted post is not based on observations about the US.

This is based on observations of China, which I consider to be the world's most advanced economy, where "advanced" is measured by "furthest up the Kardashev scale." The US (esp tech) might think it's further along, but 1) that AI lead is very small, and 2) Kardashev is measured primarily by amount of usable energy harnessed. China is far beyond the US on that front.

(If you don't agree, you can click [x] and unfollow)

There are two things we often hear about China:

1) China is facing a population bomb, with plummeting fertility rates and not enough young people to support its elderly population!
2) China has too many people, its youth unemployment rate is super high (~18% in July 2026)!

How can a country have too few young people and too many young people at once?

The answer could be one of any number of economic theories. Here's a new one, which is what I was thinking in my original post:

(This isn't a well-developed theory - the Rats call it a "butterfly idea" - so you've been warned)

Once of the nuances about China's youth unemployment isn't that the jobs don't exist, it's that many unemployed youth don't want to work the jobs that do exist.

One emerging scenario is this: there are many jobs available, but they either pay too little or are uninteresting, and young people prefer to just live with their parents instead of spending their time working one of those jobs. Although China does not have a big welfare system, their parents happen to own enough savings and assets that they are able to pay for their [one] child to continue living with them.

Notice the interesting inversion here: the prior assumption about "one child supporting two parents and four grandparents" inverts to "two parents support just one child."

(There's the "four grandparents" but there are not necessarily four living grandparents, and many old people live very cheaply)

That prior assumption rested on yet another assumption, which is that, as a general rule, people would not have substantial retirement assets and so one's children were one's retirement plan: a productive child is the person supporting their elderly indigent parents.

Children do still often support their parents, but another thing may also be happening: parents save enough assets to fund their retirement, and the assets grow faster than they use them, especially as they scale back the cost of their lifestyle, and - in some cases - it ends up being enough to continue supporting their child.

The scenario I describe would occur more often if the cost of living stayed low or even dropped over time.

Again, this is a China effect: China's industrial policy does not focus on maximizing profits (or GDP), but rather on what we might call "make more stuff and everything gets cheaper."

Everyone versed in economics understands this notion; we have lots of arguments in SF about "build more housing if you want it to be cheaper, and so more people can have housing."

Well, China produces more of everything, especially the things needed for regular living: consumer goods, groceries, etc. It drives prices down by encouraging producers to hyperscale production, and while this does not maximize profits, they more or less "make it up on volume." The West imagines this to be some kind of nefarious predatory strategy to destroy Western industries, but it's really just China making more because when everyone has more, the people are happy.

The US company that comes closest to this philosophy is Amazon, which famously optimizes for scale and operational efficiency so as to maximize value delivered to customers, putting price pressure on its suppliers and keeping their operating margin razor-thin (excludes digital services like AWS). Amazon's competitors also think that it's a nefarious predatory strategy to destroy them, but it's just Amazon serving the most customers at the lowest possible prices!

Here's a question: if post-scarcity abundance was on the way, what would be the first signs?

Well, you might see a lot of overproduction (or overcapacity). And that would tend to drive prices towards zero.

In the US we often focus on the jobs effect of that: your labor will be worth nothing, so you're going to lose your job. Well okay, but all the shit you need to buy also becomes free. How much do you need a job if everything you need is free?

That argument obscures the weird part by pointing at the ends: the weird bit is in the transitionary middle, where you've lost your job, things are very-cheap-but-not-free, and you are able to hustle a bit of money with your part-time job.

The economy is not one homogenous good, it's a heterogenous set of goods. But if the production cost of a large enough segment of material goods and daily necessities is driven down far enough, then at some point your overall cost-of-living is falling, not rising. There are Baumol goods, but they are not intractable - some are due to regulations (very different in China vs what US thinkers are familiar with here). Example: medical costs are far lower in China, with comparable or superior quality - a big portion of high US medical costs have to do with simple medical supplies in the US being overly expensive.

So you have a situation where elderly parents start from a sufficiently stable financial base where they were already taking care of their grandparents and one child, and then one or more grandparents pass, and their own lifestyle starts to scale back as they age, while cost of living for basic necessities drops over time. The child is used to a certain standard of living, that standard gets cheaper to maintain, so understandably the child not going to take a job unless they are hyper-ambitious or the job is very interesting and pays super-well.

In the US we are used to thinking of all Asians as hard-working and ambitious, but there are plenty of losers and middling young people in China - they are the ones who are living at home and not going to Tsinghua or Stanford. China has plenty of NEETs. I'm also not saying what I've described above is happening with all of them, but it being true for even 5-10% would yield the high youth unemployment number.

One of the common answers to the declining population "problem" in China is "we'll fill the labor gap with robots." China is indeed far ahead of the US when it comes to robotics. But it hasn't yet reached the level of overhauling society yet (I feel like it could within 2 years...), and the trends I've described have been playing out for at least a decade already.

So, when I say "post-scarcity abundance actually begins by manifesting first as youth unemployment," that is what I'm gesturing at.

The reason it's not manifesting as "everyone unemployment" (...for now) is really just inertia: if you're an older person with a job, you tend to keep that job. It's easier to keep a job than for an entry-level young person to get a new job, especially if that young person isn't looking for any shitty job, but a job that's more interesting and pays significantly more, vs living at home and chilling out.

Thus, if you have a situation where the segment of the population that finds it hardest to lands a job receives (via whatever combination of events) enough of the economic surplus (in China's case via their parents) that's being created by technologically-driven government-incentivized industrial and agricultural policy focused primarily are reducing scarcity for as many people as possible... then it might mean that the closer you get to post-scarcity abundance, the more those people aren't going to have jobs.

Or more precisely, the more they are going to choose not to take any of the jobs currently available.

Is any of this playing out in the US?

I had lots of comments in my original post lambasting what I said (fair; it's not like I explained anything) from an assumed US standpoint.

The US is a little different. It does not directly focus its economy on maximizing productive output for the broadest possible base of consumers.

It tends to bring (or has historically brought) great prosperity to a great many, but mainly via the indirect effects of capitalism. However, the American system today suffers from a combination of misguided regulation, and profit-optimizing market structuring by large players.

(Here, people like to quibble so here are some caveats; if you find such things tiresome, skip the next three paragraphs:

No, not all regulation is misguided. Some regulation is good and promotes healthier markets and better net outcomes. But good regulation can become out of date, or regulatory capture happens, or dumb regulations get made by officials who are out of touch with technical realities - this last one does seem more and more common.

Large players seek to re-structure the market in ways that are favorable to their profits, and not necessarily total value delivered to customers, e.g. hedge funds buying up all fire engine manufacturers, so that US municipalities have to pay $1-2m for a new truck, while the equivalents can be found on Alibaba for 1/10th the price.

Another example of large market players influencing market structure to optimize for profits is offshoring all of their manufacturing capability!)

Back to the main thread: For many complex and inter-related reasons, cost of living in the US is not falling - it's rising almost untenably for most.

At the same time, it's still hard for young people to find jobs, because the "there are jobs, just not ones I want to take" effect also exists here, except that living with parents or on minimal income is far harder, so everything feels extra shitty.

When you don't have a job, and everything is very-cheap-but-not-free, BUT you don't have access to those very cheap foreign-made goods because importers buy low and sell high to you while you still only have minimal income, "lying flat" feels like a whole different story in the US, vs China.

One area where a China-like effect does occur is in availability of consumer tech devices. Tech companies have engaged somewhat more often in the "make more and make it cheaper so more people can buy it" strategy, though this seems to be driven more by the megalomania of creating the biggest possible company than any notion of broad-based industrial economic development - but the effect is similar: almost everyone now carries around a device in their pocket 100x more powerful than the computers on the Saturn V, and has access to untold amounts of online services. America has post-scarcity abundance in a narrow slice of goods and services.

America (or its population) could participate in the Chinese-driven post-scarcity trend by simply eliminating the trade barriers and allowing Chinese goods to flood the US. And unlike the cheap low-quality Chinese goods of yesteryear (i.e. most Americans' received impression), these are goods of comparable or higher quality.

The problem is that not only would this potentially yield the same "lying flat" youth unemployment issue in the US, it would utterly demolish many American businesses, and thus the wealth base of most American elites. Every American car company would probably be gone in 18 months. And American youth are not going to rebuild America's manufacturing base, they're just not going to. And American robots aren't going to either, they'll be outcompeted by Chinese robots.

In another post, I offhandedly mentioned that we're in the Singularity. Most people who follow me live in the tech sphere, so this was largely accepted unchallenged. But many of the "no we're not" objections basically rested on the idea that "life is still shitty and it's trending worse and the Singularity is supposed to be like Heaven, so we can't possibly be in the Singularity."

Well, there is nothing that says the Singularity or even the post-scarcity abundance world is going to subjectively feel great, much less the transitionary path to it.

The Singularity only says that AI will become smarter than humans, and post-scarcity abundance only says that all material goods are going to be free. Human happiness and pleasure are a function of many things, and most of them are not material. "All your stuff being free and every robot is way smarter than me" does not by itself a utopia make.

The road to post-scarcity abundance is not necessarily going to be a pleasant or positive experience. We still have to choose to make it so.

The rest is left as an exercise to the reader.

[Tomás (Now in Toronto) Bjartur @BjarturTomas]

Ngo on Sama’s charisma and duplicity.

[Lucas Beyer (bl16) @giffmana]

Google in a nutshell

[Colin McCarthy @US_Stormwatch]

China now has 1.2 terawatts (TW) of solar capacity and built more solar in 2025 than the US has built in its entire history.

The US needs to build more solar, and fast.

[Mark @yieldthought]

@sarahookr Progress in robotics has been breathtaking; easily comparable to the early days of AI. Next two years will blow everybody’s minds.

[Galbot @GalbotRobotics]

WORLD RECORD!

Galbot robot completed 100+ consecutive autonomous tennis rallies with zero mistakes!

🎾 Fully autonomous.
📷 Live Broadcast.

A new milestone for humanoid robot tennis!
#AstraTennis #EmbodiedAI #HumanoidRobots #WHRG #Galbot

[Shenzhen Foundry @shenzhenfoundry]

All fully autonomous. Zero human control during the race.

Some clear the distance faster than the human world record. Others eat the track, spark, or finish in pieces.

Speed and onboard decision-making look solved on a clean oval. Reliability under repeated high-torque cycles, heat, and recovery from near-falls is still the expensive part.
#PhysicalAI #HumanoidRobots #世界人形机器人运动会 #Robotics #EdgeAI #ShenzhenFoundry

[Mark @yieldthought]

A lot of people here laughing at the funny robots in the olympics this year.

The next one will be watched in shocked silence.

[Rand @rand_longevity]

this is exactly what ai should be made for, this is what the people need to see

[Paul Graham @paulg]

@mohsin_rafii To learn how wheels are made.

[roon @tszzl]

there will be a steady ramp of such events

[Kyle Chan @kyleichan]

This is the most impressive to me, more than the 100m world record.

This requires autonomous real-time planning and action in response to a fast-moving target and dynamic environment.

Galbot is one of China’s top humanoid robot startups.

[Mark @yieldthought]

Continuing masterclass in customer relations from @thsottiaux: responsiveness, honesty, clarity, generosity.

[Natalie Winters @nataliegwinters]

EXCLUSIVE: Chinese Institutions Are Building AI Models Of American Voters—And Testing Political Messages On Them.

Fudan used 171 MILLION X posts to create a million-account “voter pool.” A government-linked team tested a campaign message on synthetic Pennsylvania voters. 🧵

[Mark @yieldthought]

Has nobody considered “lies”?

[Brandon Carl @brandonjcarl]

100 trillion free tokens a day is a rate of 3 quadrillion a month. That's on par with Google.

What's really notable is this: that many tokens means that either (a) there was a lot of spare capacity or (b) this is potentially a breakthrough in intelligence per watt that @GavinSBaker speaks to.

The third alternative is simply that it's not a very good model. My guess is that it will start off poorly but that post-training will help it get much better.

[The Internet Remains Undefeated @WebUndefeated]

This is why it’s important to let kids be bored.🏴‍☠️😁

[Stephen R. C. Hicks @SRCHicks]

Only slightly exaggerated.