#152. Our robotic future: dancing dogs and stochastic parrots

July 6, 2025; 18 min slightly apocalyptic read

Unitree’s B-2W: WOOF

Robots are popping up everywhere — the other day they debuted on America’s Got Talent, dancing to Queen’s “Don’t Stop me Now” — even though one of the five robo-canines went on strike one-third through their pre-choreographed set. Sadly, they weren’t freestyling — they were just repeating a programmed routine, but the crowd seemed to love it anyway, mostly for the novelty of it all. And they really seemed amused by the fact that one of them fell down on the job.

I’ve been trying to keep up with the pace of robotics development, both from the sheer appreciation of their capabilities, as well as the implications for US - China technology competition. Japan used to be a robotics leader (in fact it still leads in industrial robots), but it had the misfortune to develop robots in the age before artificial intelligence, which is the critical reason that modern robotics are improving so impressively. For example, Honda developed the humanoid Asimo and Sony developed the cute canine Aibo, but both programs were cancelled around 2018.

The current state of the art for robo-canines is probably the Unitree B-2W, pictured above, which appears to have almost magical balancing and navigating abilities, and these competencies will only continue to improve. Real dogs (canis familiaris) have evolved somewhat over the millennia, teaching humans to pick up their poop and such, but it is evident from the video that powered wheels give robo-canines a pretty substantial edge in mobility, a genetic modification that good ol’ Rover is unlikely to ever develop.

In this issue, I look at the nascent world of smart robots within the context of rapid AI advancement, examining what I think will have an underappreciated impact in both economic and political spheres. I will also add to some of my thoughts about AI, building on some of my previous thoughts.

From digital to physical leverage

AI (no need to define it anymore), and specifically the Large Language Model (LLM) space, is having yet another zeitgeist moment, where the ultrafast adoption curve is driving enormous shifts in our technological behavior, and investors are paying astronomical numbers for pre-plan (not only do they have no revenue, they’re not sure what the business model is yet) startups.

AI has crept into our lives already: user queries on Google now usually return an AI-generated answer, rather than a set of links — this AI-first paradigm, which actively meets consumer needs better than a ranked list of links, has the happy side effect (for consumers, not for businesses) of avoiding advertising, by going straight to the probable answer. This diversion could materially change the economics of ad-supported websites, because fewer ads viewed means materially less revenue for content creators — potentially requiring a difficult move to a paid subscription model which is usually a high hurdle.

While there is quite a debate about how AI will affect work and workforces involved in the white collar digital economy (partly because the IT revolution of Internet 1.0 improved processes, but despite the near ubiquity of computers, hasn’t resulted in a marked shrinkage of the labor force), the advent of AI-powered robotics will almost surely have a far greater impact, both empirically and psychologically, on the physical labor force (and not just blue collar factory workers). This is partly because humanoid robots appear to replace humans 1-for-1, in a way that disembodied LLMs or software suites do not. It’s not clear how humans will react to being replaced by a canine robot (“bad dog!”), but one can assume the reaction won’t be good.

Currently, robots are primarily used in industry for very specific tasks — one estimate suggests 4 million industrial robots in use globally, with 230,000 newly installed in 2023, although Amazon estimates that they alone account for 1 million. Their wider adoption has been hindered by their ‘dumb’ nature — engineers need to explicitly program their precise behavior because they cannot learn on their own, or make their own adaptive decisions. Compared to the 500-600 million people estimated to work in manufacturing roles, industrial robots have had only marginal penetration. The same way the IT revolution didn’t appreciably impact the white collar workforce, the low penetration of industrial robots so far may mislead people to believe that even AI-powered robots will never replace humans at scale… but that is a dangerous assumption to make.

Currently, estimates for humanoid robots currently deployed in real world settings hover around 10,000 globally (the bulk of which are probably those creepy security robots which look like overweight R2D2s).

The advent of AI-powered robots will likely change that narrative, because the addition of the AI dimension enables them to ‘learn’ new capabilities and enhance functionality. Industrial robots tend to look very different from humanoids (i.e. disembodied metal limbs) because they are optimized for very specific tasks (like high speed surface mount component placement). The benefit of building humanoid robots is that the physical world is built to accommodate humans, and therefore a highly flexible version of a humanoid form factor could conceivably replicate many human tasks. Historically, humanoid robots have lagged in areas like adaptive locomotion, fine motor skills, reflexes, and sensing (vision, touch, hearing more than smell or taste). But the Unitree and Boston Dynamic videos suggests the gaps are narrowing quickly.

In both China and the United States, the physical capabilities of humanoid robots are rapidly approaching the level where they can start to reliably perform many human tasks, and both the physical capabilities and AI-powered learning layers are improving. And once more humanoid robots are deployed and extensive data / feedback is collected, this will further accelerate the learning and upgrade process. Compared to industrial robots, humanoid robots won’t be optimized for one task, but will be able to do many tasks and therefore their flexibility will make them more useful to industry, because they will be able to do a wide variety of tasks quickly and with very few errors (which they will learn from).

Humanoid robots are the physical link to the extreme advances we have been making in the AI / digital world, and as that link is strengthened and their capabilities and flexibility increase, the implications will be quite dramatic.

Thought experiment: the domestic helper

In Hong Kong, it is common to have a foreign domestic helper (FDH), at a cost of about US$850 a month (free accommodation provided). The tasks of an FDH can be generally grouped into four categories: household chores, and then depending on the case: childcare, elderly care, pet care. Within household chores, the large task subset includes general cleaning, laundry, cooking, shopping.

At the moment, the wide variety of cleaning tasks (vacuuming, dusting, sorting / rearrangement, disinfecting, etc) performed by FDHs or Rosie the Robot Maid in the Jetsons seems far away from the capabilities of modern robots, which might be able to dance set routines and navigate obstacle courses, but robots are starting to acquire the necessary subskills that will eventually enable them to perform a wide variety of tasks.

Laundry represents a higher level challenge, with quite a lot of sub-decisions and background knowledge; cooking is higher still, requiring understanding of a myriad of tools, from stoves to blenders to mandolines, and knowledge like knife skills, recipe interpretation, seasoning technique, and food safety; shopping would necessitate neo-Rosie to leave the residence, which is yet another dimension of complication, so the first generation humanoid would probably remain entirely at home where the environment is controlled. I can imagine a first-gen neo-Rosie receiving packages, opening doors, and supervising workmen, but it’s probably a while until you’d trust her with child / parent supervision, or walking your dog outside. While robots have nowhere near these capabilities at the moment, the speed at which they are acquiring the necessary subskills makes me optimistic that they will get there faster than we imagine.

The increasing pull of tractors

I chose the example of domestic helper as an example of a role with a highly complex task list which is far away from current robot capabilities; in the near term, AI-powered humanoid robots will proliferate in settings like factories, where they have a small but growing suite of simple repetitive tasks (sorting / picking / carrying).

A short history of the tractor, the robot analogue in agriculture is perhaps useful: steam powered tractors were introduced in the late 1800s as part of the Industrial Revolution to replace draft animals, focusing on plowing, threshing, and hauling. Three mini-revolutions widely expanded their productivity:

• Gasoline power (early 1900s) made tractors far easier to operate and gave them more horsepower (appropriately)

• Power take-off (PTO) in the 1920s enabled auxiliary devices like mowers and balers to draw power from the main engine

• Hydraulic systems in the 1940s enabled tilling, planting, and spraying

Therefore tractors dramatically expanded their use cases and productivity over a period of forty years. Academic studies estimate about ⅓ of agricultural productivity since the age of draft animal labor has come just from mechanization (tractors), with the other ⅔ coming from crop science, fertilizer and pesticide improvements, irrigation, etc.

Just as the adoption curve of tractors increased when they added functions and became more flexible (and came down in price), so too will humanoid robots follow that rough curve as their utility improves.

Stochastic parrots and their ilk

Emily Bender, a noted AI pessimist recently profiled in the FT Weekend section, coined the term of ‘stochastic parrots’ to describe what LLMs are actually doing: they have consumed a large amount of lexical information and are simply parroting back (although I would argue in a trained rather than purely stochastic manner) a glob of text which they do not fundamentally understand, but which happens to resemble human intelligence.

A few things to note here:

1. Although I agree with the overall view that LLMs don’t fundamentally understand language conceptually, ‘stochastic’ paints the parrot as dumber than it really is — it has analyzed an enormous amount of human output and can probabilistically return a set of ordered words which approximates a human response. Humans can be like this too: one example of a human being a stochastic parrot is the viral girl who uses ‘misogyny’ as a reflexive defense (but can’t define the word)… and then compounds her issue in the full video by saying she doesn’t have to define the word because she “signed a waiver” – she is an example of a human parrot who doesn’t understand the meaning of the language she uses.

2. Newer language models have the ability to overlay context, and will vary answers based on differences in context; they can also show you their internal thought process and production logic. Those higher order response functions may not ‘understand’ the context either, but the responses are increasingly elevated to levels which humans often fail to reach (having access to the internal thought processes of some humans is a positively frightful idea).

3. The framing of the argument is arbitrarily skewed, because it assumes that higher order thinking has to do it the same way humans do, which is not necessarily true. AlphaGo Zero is the strongest Go player in the world, but it does not determine its moves in the same way as humans. Also, AIs and LLMs have essentially perfect memories, unlike humans, whose cognitive processes should account for predictably imperfect memory (but usually don’t).

4. Most importantly, for nearly everything we want robots to do, even the basic idea of a stochastic parrots is perfectly acceptable for non-creative routines. We are not designing AI-powered humanoid robots to become Pablo Picasso, we just want a device which can move hazardous waste or clean our filthy bathroom or carry a bunch of loose items around. If they don’t perform these tasks during edge cases (after a natural disaster or power outage, for instance), then that is an acceptable tradeoff because the net productivity and economic gains are so large.

The China advantage

At the moment, the US and China appear to be at the forefront of humanoid robotic systems, although the advantages of China are multiplying rapidly. The US has slightly more sophisticated AI when it comes to LLMs, but there are more AI researchers in China, and they are filing patents in the AI space faster than the rest of the world combined (even if quantity is not quality). I would note however that five of the eight high level AI researchers recently seduced by Meta from OpenAI were ethnically Chinese. My perspective is that China has a variety of structural advantages which will enable them to lead the way in AI-powered robotics:

• Supply chain: the historical build out of EV and other strategic technologies means China has the strongest supply chain for many critical robotic components (batteries, sensors, motors, LIDAR and vision systems, etc.). This gives China an enormous cost advantage compared to the US, even discounting the existing dependency of US robotics on the China supply chain and the intrusion of politics on free trade for strategic components.

• Development speed: the breadth and integration within the Chinese electronics supply chain is unrivaled, which allows projects to ramp production very quickly with many of the suppliers in the same concentrated location (usually Shenzhen), who are used to quickly prototyping integrated systems, which materially increases the speed of the development cycle.

• Industrial policy: the Chinese government is fully supportive of efforts in the AI/robotics space, allowing for a variety of advantages including financing, subsidies, and contracts, in addition to a regulatory environment which leans more supportive than restrictive.

• Data access: it is often mentioned that data privacy in China is not what it is in the West; since access to data is critical to train AIs and robots, access to useful data is an important factor.

• Engineering mindset: the West tends to operate on a star system for programmers / engineers who prefer higher level conceptual design or entrepreneurial exits, rather than grunt work; China produces more than 2 million STEM graduates annually, compared to the US at around 500,000 (many of whom are of foreign origin). The ground level implementation of robotics systems will require a lot of detailed systems integration work, the kind of non-glamorous tasks which probably gives China an edge.

• Multiple solutions: The sheer size of China’s STEM workforce also allows them to explore a variety of different solutions rather than arguing for just one; the 20,000 engineers at battery leader CATL developed and are producing both LFP and NMC batteries, as well as exploring a wide number of potential technologies for the future, rather than betting on a single type like Toyota, who have gone all in on solid-state as the battery tech for the future.

• US xenophobia towards foreign talent: making it more difficult for foreign talent to study and work in the United States will not help American efforts, although it is clearly a positive sign that so many top people are studying in the US to begin with, especially in AI.

Economic implications

In the long run, robots will probably change the economics of manufacturing the way tractors changed the economics of agriculture, and their high level adaptability will allow them to proliferate more than industrial robots. Morgan Stanley, which has done the most comprehensive work on robotics that I have seen, has estimated that a robot which costs $5/hour fully-loaded to run, will be able to replace the output of two humans who cost $25/hour, so a net gain of 10X productivity per dollar. And during a time when self-interested voters are trying to keep out immigrants and raise minimum wages, the economic incentive to employ robots will continue to rise.

Robots will take over the common tasks, while humans will be necessary for declining numbers of edge cases, but in far smaller numbers (and at higher pay), especially as their job will be to reduce edge cases and indirectly make themselves obsolete. There are also the dimensions of reliability / volatility / adaptability to consider: humans can quit or exhibit unreliable behavior; have sick or low productivity periods (a bad case of the Mondays?), and tend not to upgrade themselves very much, whereas reliable robots would get regular software and capability updates, and will never have PTA meetings or sick pet emergencies; and wouldn’t complain when asked to work overtime or holidays.

Competitively, if robots aren’t adopted (and no doubt some countries will prevent them from replacing human labor by taxing them, or treating them like immigrants), then those Luddite producers will almost certainly have higher production costs than their global competitors who do leverage robotics, and will find it very difficult to compete without tariffs or other protection (regulation).

Historically, the world has focused on labor costs as a primary determinant of competitiveness, but the labor share of income has been declining as the capital share (mechanization and finance) of income has risen. When AI-powered humanoid robots become capable enough, the other way to think about this is that, in increasing numbers of areas, robots will become the most economically efficient source of labor on the planet, and any country / company who can obtain those robots at the lowest price will have a significant advantage. Emerging economies (Vietnam, for instance) who have relied on inexpensive labor as an input to competitively priced exports may no longer be on the efficient frontier, i.e. their workforces may be too expensive relative to the cost and efficiency of robots, which can then be located in the end consumer economies (to minimize transportation and other frictional costs). Of course there will always be service sectors and niches available to local workers where robots can’t compete or haven’t been adapted for, but economies do not grow via a collection of subscale niches. And for low to medium skill jobs which aren’t accessible to robots, there will presumably be quite a lot of human competition, keeping wages at minimal levels. Some people counter that the creative arts or sports competition will be a robot free zones, but even if that is true, the compensation curve will follow a normal distribution, and demand for both the arts and sports will remain focused on only the best – there will be no outsized payday for the 10,000th best basketball player or guitarist, for instance.

Labor impact: debunking the orthodox conclusion

Economists tend to be naive about the net impact for labor as the result of quantum technological changes. This and other articles basically say that the labor force adapts, and the wealth created by the new technology stimulates new jobs and changes in behavior, which offset job losses. Part of this is that economists don’t want to be Negative Nellies, they want to reassure frightened laborers that they shouldn’t freak out and lynch the latest AI gazillionaire. Perhaps on a very high level it works out, but different geographies can have extremely varied outcomes.

The ‘offset’ conclusion is often true while restricted to the specific locus of innovation, the picture is less encouraging when you zoom out to the wider world. Consider:

• In Britain, the impact of the Industrial Revolution in textiles was fine for the home country despite the objections of the original Luddites because the wealth created there circulated into new products and services, but it had devastating consequences for the Indian economy and textile industry, where few offsetting jobs were created, and this dug a sinkhole in the economy: India, once the world’s largest textile exporter, experienced mass unemployment, which exacerbated the Great Bengal Famine of 1770 (the British imposing tariffs on finished Indian textiles did not help matters either).

• In Mexico, the maquiladoras, largely US-owned Mexican factories which served the US market went through a very difficult period (with no employment offset) when China joined the WTO in 2001 and the Chinese cost advantage allowed them to dominate many markets. (Mexican textile exports to the US fell 25% between 2001-2005, while China’s share exploded, for instance).

• In the US, the advent of Japanese automobiles (and the oil crises of 1973 and 1979) accelerated the hollowing out of the Rust Belt, which lost 300,000 workers in the auto makers alone between 1978-1982 (not counting parts / component company unemployment). Try telling the average citizen of Detroit or Cleveland of that era about the employment offset effect.

Net net, while roles like personal trainer or vegan nutritionist to a few new AI billionaires will be created, but the consumption functions of the Uber-wealthy are quite miserly (as a % of income) compared to factory workers (who spend nearly their entire income), so overall consumption and economic growth will suffer on balance.

Final thoughts

The important idea is that, within the next few years, the world’s most economic, constantly upgradable, and highly flexible worker for many fields will be an AI-powered humanoid robot.

This will manifest first in manufacturing and industry, where humanoid and canine robots will work alongside traditional industrial robots, further reducing the required level of human labor per unit of output. As their micro capabilities multiply, humanoid robots will start to invade the more basic service sectors as well.

China will probably lead in the application space because of their advantages in supply chain, development speed, data availability, government and regulatory support, and broad technical workforce. The US will probably lead in theory and large comprehensive models, unless it follows through with preventing foreign talent from studying/working in the US.

This growing unemployment effect will have both economic ramifications: the global competitive landscape may change quickly as the result of materially lower labor costs, so the labor share of income will continue its long decline; as well as political implications: calls for restrictions / regulations / taxes on robotic work forces, neo-Luddite movements, growing populism. Low skill jobs in both advanced and emerging economies will be at risk, and the skill level waterline will keep rising.

Although I might tackle this in a different issue, the implications for robots in the military are enormous, because armed conflict is almost certainly the most hazardous of all jobs. That said, I see the field of war as more suitable for four-legged and no-legged [air based drones] robots, rather than humanoids.

Widespread adoption of AI-powered robots will further accelerate the income and wealth inequalities that are already creating large rifts in our societies, and the challenge for many governments is how to redistribute this wealth in a way that their constituents think is fair, and which also does not incentivize the winners to move their riches elsewhere. The cost of making mistakes here will be revolutionary.