The Final Frontier

The great-power race to protect sovereignty in space has reached a new stage in 2026. Low-Earth orbit (LEO) connectivity systems like Starlink are converging with higher-orbit navigation systems like the US’s Global Positioning System (GPS). Starlink is still by far the dominant global player in this game, but as in earlier Internet chapters the market built by an American company is fragmenting into multiple pieces, each highly inflected by local politics. That creates investment opportunities, though of a highly geopoliticized kind.

Looked at one way, the extraterrestrial world should be able to ignore sovereign boundaries. Satellites orbit regardless. The International Telecommunications Union (ITU) is the forum for negotiations over satellite use of radio frequencies and over permission to use orbital slots. One country cannot prevent another from launching satellites (except through physical force).

However, nation-states do control a number of levers. The use of frequencies within a country is controlled by that country’s government. Countries normally license the terminal dishes used by consumers to grab a signal. Current systems need terrestrial gateway stations that act as a two-way bridge connecting satellites in space with Internet fiber on the ground. Gateways in turn need land, permits, electricity, fiber access, and so on. A satellite-signal company like Starlink also has to comply with local laws about content, taxation, data interception, and so on. (Starlink discovered this when it crossed the government of Brazil in 2024.) Satellite communications systems do indeed fly around above Earth’s puny borders, crossing national borders at will. But states on the ground still have plenty of leverage.

In the current geopolitical configuration, major players are all acting about as you would expect. Russia has a GPS alternative, GLONASS, that has been beset with maintenance problems, much like Russia itself. Its broadband constellation, Rassvet, has managed to put something like 40 satellites in orbit this year, although its target was 156. Like Russia itself, Rassvet tends to privilege national security over consumer services.

Canada, interestingly, aims to launch a system, Telesat Lightspeed, aimed mainly at government and enterprise users, including on military Ka-band spectrum, but the first launches won’t be until the end of this year or early 2027. Similarly, India’s JioSpaceFiber runs on others’ satellites — those of SES, which is minority-controlled by Luxembourg state entities with a majority of shares held by institutional investors — but hopes to have some test satellites of its own aloft in about a year. India’s NavIC positioning system, however, has been a disappointment. Taiwan has B5G, with plans for an entirely Taiwan-built payload by 2030. All of these smaller players (and much larger SES) are worth watching.

Europe has long sought an independent space capacity, basically as a sovereign hedge against the US — a motivation that has burgeoned under the Trump administration. Eutelsat OneWeb is a mature service, but it is there for governments and enterprises, not consumers, and is mainly keeping itself sound until the transition to IRIS. But IRIS will continue as a security-first sovereignty play, one the EU is likely to need if current difficulties with Russia and the US continue. IRIS’s first launches are currently scheduled for 2029. It is not meant to be consumer facing. Europe’s Galileo positioning system can at least be considered a success, with superb accuracy and a global customer base.

That leaves, as so often, the US and China. The US, typically, has a combination of slow-to-innnovate public infrastructure (GPS) on one hand — to take care of geospatial positioning — and extremely vital but also uncontrollable and erratic private efforts (Starlink above all, but also Amazon’s Leo, with 396 satellites currently in low-Earth orbit). Starlink has a staggering 11,130 satellites in orbit and 12 million customers.

China’s GPS alternative, BeiDou, has been consistently strong and compares favorably with GPS. Its LEO effort, however, got off to a slow start, mainly because its initial focus was security-driven and buried within China’s massive complex of state-controlled military enterprises. The result is called Guowang, and it has had a number of slow years. So much so that in 2023 state regulators ended Guowang’s monopoly, clearing the way for a consumer-facing rival: Qianfan. This was more in keeping with the Chinese government’s approach to innovation: creating a space for domestic rivals to battle with each other. It is still an intra-governmental game. Qianfan is essentially controlled by the government of Shanghai with participation by China Telecom, a central state company. But at least there is competition underway.

If Guowang continues to right itself, China will have a dual-track offering, with a strong sovereign-military component (Guowang) and a consumer-facing Starlink alternative (Qianfan), as well as a superb higher-orbit positioning/navigation system (BeiDou) which Qianfan is integrating with its own satellites. No other state will get close to this dynamic, integrated capacity. BeiDou already exports to more than 140 countries. Qianfan is aggressively pursuing service deals, famously so in Brazil, where Starlink first had to kneel to national sovereignty, but with signed arrangements with Kazakhstan, Turkey and Malaysia and soon others.

The United States will get wherever it gets by grace of its private tech sector. The likely near-term map shows Europe slowly but surely developing a positioning-connectivity-infrastructure system that will keep it from dependence on the US and China but might not prosper commercially in world markets. China will have a government-dominated offering that will essentially provide services comparable to the American alternatives but with an unknown number of dependencies that would leave its effects on national sovereignty (outside China itself) highly debatable. The US will have private, for-profit companies dominating development and leading innovation but disconnected from the public interest. (Starlink actually shut down its positioning feature this year.) And other ambitious countries like Taiwan and India will carve out mini-sovereignties and, at least with India, look for markets abroad as well. Businesses that depend on these emerging infrastructures — and they will number in the millions — will need to balance revenue with the not-so-obvious costs of infrastructural dependencies. Investors will need to understand those choices to allocate capital well.

The Great Filter, Part 9

By Dee Smith

The extraordinary arrogance displayed by AI specialists convinced that they are “creating” super-intelligence, or even creating God, is stunning. Given that human perception and understanding are too frail and incomplete to comprehend the insuperable complexity of the real world, or even of ourselves, the claim that we can make machines into super intelligences that transcend anything comprehensible to humans shows an astonishing hubris. If we cannot even realistically determine how human or animal minds operate, how could we have any basis of making such claims about AIs?

Nonetheless, in part because our understanding of intelligence, consciousness, and the world itself is so lacking, there may well be aspects of machine intelligence that can become very dangerous indeed, even if machines do not experience anything like the consciousness and awareness that animals—including humans—possess.

An AI may not require human-type consciousness, or even what we think of as awareness, to manifest agency and intentionality, and to act on those.

That may seem like a contradiction in terms, but it may just be that we don’t have the understanding and lexicon to grasp and express what is going on. There may be a different “package” or mix of elements in AIs than there is in humans or animals: a mix that can produce qualities that seem paradoxical to us.

We may well be anthropomorphizing: assigning human characteristics to AIs that are utterly non-human and don’t think or behave like humans (despite being “trained” on a substantial part of human knowledge). In looking for human-type consciousness or intelligence, we may have our eye on the wrong ball.

Are we asking the wrong questions? 

As Gertrude Stein memorably said: “When you get there, there isn’t any there there.” But what could that mean in the context of AI? 

In the final analysis, yet again, we just don’t know. 

We have, however, for decades studied elements about complex systems that are relevant to all of this. (The study of complexity is one of the great triumphs of 20th century science— and was enabled by earlier digital computers.) Put basically, order and structure emerge spontaneously and often unexpectedly from complex systems. Consider how organized a hurricane is, although it starts out as only water vapor, wind, and heat from sunlight and water. Or look at the beautiful structures that emerge in a cup of coffee when you pour cream into it.  

This characteristic of emergent properties innately becoming orders of magnitude more complex is another element that should give us pause. What is emerging or will emerge unexpectedly within complex AI systems? We don’t know, although we see some alarming indicators, as noted earlier.

One danger is even that AIs could lead to a dumbed-down, binary perceptual world that is devoid of certain kinds of awareness and thought, but that is nevertheless relentlessly driven by the pursuit by the AIs of programmed or ultimately machine-derived goals. Suppose, for example, that humans program an AI with explicit and non-cancellable instructions to ensure that nothing further damages Earth’s biosphere, and the AI determines that humans are the reason the biosphere is being damaged. Then, with impeccable logic, the AI decides that humans must be eliminated in order to fulfill its programming. Now, that is a relatively straightforward example with relatively simple fail-safes. But can we think of, forecast, and guard against all the permutations of all the programs in all the AIs, including programs they may generate themselves? The answer is no. 

Clearly, AI has benefits. For example, recent AI redesigns of aircraft wings, in particular in Germany and China, seem to provide significantly more efficiency. AIs can, though statistical means, solve mathematics problems, and devise winning gaming strategies never before seen (such as in chess). AI is good at organizing and summarizing information, another statistical, calculational task. But you have to check the output for errors and hallucinations every time. For example, recent private research showed that the best AI model that was tested scored 82.4% accuracy on financial spreadsheets. This means that one number in about every six is wrong. Would any manager put up with that level of inaccuracy from a human analyst?

The only sane thing to do would be to slow all of this way down, or even bring it essentially to a halt, at least until we can figure out more about what it really represents. And we may never figure that out: the problem of consciousness is so intractable that we may not be equipped with the means to ever understand it.

But humans, with their competitive natures and competitive national and commercial structures—and their strange predilection to act against their own interests to satisfy psychological and financial needs—seem extremely unlikely to slow down. Even if it is banned, countries will continue to develop AI in secret based on their fear that an adversary nation is also developing it in secret.

Tragically, AI may turn out to be a perfect exemplar of the Great Filter—and possibly in the not-too-distant future.

What happens would seem ultimately to depend on the maturity of the human race, and on luck: two constantly recurring, and terrifyingly thin, threads to hang our future on.

AI is by no means the only part of the Great Filter story, just my entry point. More soon!

The Great Filter, Part 8

By Dee Smith

The most intractable and equivocal problems in understanding AI and its relationship to biological and human consciousness essentially comprise a hall of mirrors. In part 7, I noted that AI cognition is filtered and abstracted because of its binary nature, among other factors. But saying this does not take into account the nature of the very different kinds of filtering that affect human perception and consciousness (and that of every other animal and plant on Earth).

So, when we ask whether AIs can ever perceive, let alone understand, the world in anything like the way humans or animals do (which assumes those are the same: they are not), we may not even have the conceptual framework to know what we mean by the question.

For centuries, many cultures have tried to understand the relationship between what we see and what actually “is.” Plato’s cave is among the earlier representations of this idea in the Western tradition; it has also been of great interest in traditions like Hinduism, Buddhism, and Taoism. And most if not all cultures have used meditative or ritual techniques—and substances like hallucinogens (and even alcohol and raw, high nicotine tobacco)—to experience altered perceptual states. Modern neuroscience posits that filtering out parts of reality, through inhibitory neurotransmitters (chemicals that work by decreasing the probability that a neuron receiving them will activate in a certain way), is a profoundly important element in human perception and cognition. But although the chemistry is mapped to a certain degree, there is no real understanding of how this “sums up” into the experience of the world we have.

The homunculus argument in the theory of mind is based on the fundamental difficulty of understanding what within the human (or animal) mind views and interprets our perception of the outside world. Simply stated, our understanding of biology tells us that our eyes receive light from outside and process it to give us an “image” on the retina. But who “sees” or “watches” that image to generate our representation of the world, and who decides on the actions to take in response?

The homunculus argument posits that there is something like a “little man”— a homunculus—within human consciousness that is watching. But then, what inside the homunculus is watching the image the homunculus sees? Another smaller homunculus? This becomes an infinite regression.

The argument is now considered a fallacy that is primarily useful in detecting and exposing false theories of consciousness, but the critique is relevant to AIs. If AIs have sentience, who or what inside that AI is perceiving and interpreting?

As any neuroscientist will tell you, we do not understand how human or animal perception, awareness, or thought work. We do not understand what free will might be (if it exists), and we do not even have any really convincing ideas about it. Using algorithmic processes to simulate thought and then claiming it is actually intelligence may be as much wishful thinking as positing that quantum effects (which are probabilistic and in many cases fundamentally random from a human perspective) somehow account for the agency and directional free will we believe we experience. We do not even understand basic elements of the universe we live in, such as time. As a recent article in New Scientist stated: “it’s not that time isn’t real, but rather that we really, really do not understand it.”

It is very likely that what we see is not the real picture, in a specific way and for a specific reason: that human perception has evolved to enable survival, rather than to reveal the truth about what is “really there”, as evolutionist Donald Hoffman has proposed. For example, it is strange that in extremely low-light conditions (like the post-twilight gloaming), humans can still detect the color green, but all other colors appear as grey (try it!). But green in a natural landscape denotes trees or bushes—places to hide—whereas other colors do not give that signal. So being able to discern green in low-light conditions would definitely have survival value on the savannah. 

Insects see ultraviolet colors on plants that we do not, but they do not see the color red. Bats “see” the world through sound—echolocation. Even dogs exist in a perceptual world so founded on smell that it has been described to me by trainers of security dogs as “so different from the way we see the world that we literally cannot imagine it.” 

What does this mean about how filtered the rest of our perception—or of any possible perception—of the world actually is? Can we claim that an AI’s binary perception is more limited than our perception (given that ours does not let us even see some colors)?

I suspect we can. The idea of a single principle explaining the world is clearly very seductive. A number of prominent physicists like Stephen Wolfram promote the idea that reality itself is computational—that the universe is a giant computer, “solving” some “program” that in some way creates the reality we perceive (which, some say, might be a simulation).

The desire to believe that a unified, comprehensible idea about the world—like computation, which is the flavor of the day—unlocks its true nature is extremely compelling and comforting for people, laypeople and experts alike, who seek a unitary understanding, or “theory of everything”. It is an emotional response. Yet these models are always reductive and always fall short. There is precious little if any ground to stand on, whatever position one chooses to take.

But we do know that biological life on this planet has billions of years of evolution behind it, which has honed its ability to persist. AI has not undergone that process (unless you posit that it just “transfers” from what humans have undergone). More fundamentally, it is not clear at all that binary perception can “sum up” to anything like human perception. Thus, the hall of mirrors.

The Great Filter, Part 7

By Dee Smith

Are we acting as sorcerer’s apprentices by developing AI? One place the possibility is readily apparent is generative biological design undertaken with genome language models. The journal Science recently began an article by saying, “The ability to design complex biological systems with artificial intelligence (AI) has the potential to transform biotechnology.” The journal article then noted that a group of researchers, led by Samuel H. King: 

used generative AI models trained on millions of natural genomes to design entire bacteriophages. Experimental tests yielded 16 functional genomes with diverse sequences, structures, and fitness profiles. A cocktail of the generated bacteriophages rapidly overcame bacteria that had evolved resistance to a natural bacteriophage. This work lays a foundation for AI-guided design of biological function at the whole-genome scale.

Dominion over the processes of life itself is a dangerous and heady prospect. While it clearly has applications in medicine and agriculture, AI is a “dual use” technology that can result in the creation of biological weapons by state and non-state actors — that is, biological warfare and bio-terror. The idea that all of this can be controlled seems ludicrous, but the technology is here anyway. Precision medicine and making crops resistant to disease sound good, but the other edge of the sword is that what can be used to feed a population can also be used to starve a population.

It is both telling and fascinating that the emergence of AIs has brought us directly and urgently face-to-face with some of the oldest, thorniest questions in philosophy: the nature of life, of consciousness, of the self, of other minds, of perception, of free will, of ethics — at the same time that it is at our front door as a potentially existential threat that we do not understand.

There are many domains of human experience that cannot, based on any evidence we have, be replicated by computational, algorithmic systems. These include mystical and religious experiences, altered states of perception (like being drunk), inspiration and creativity, falling in love, and being moved by music. 

Most fundamentally, AI systems do not have what philosophers call “qualia”— the “what-it-is-like” to experience things such as the color blue or the taste of chocolate. Computational systems can only mimic expressions of these, as understanding Chinese is mimicked in Searle’s Chinese Room mentioned earlier in this series. Adjusting what is called the “temperature” of an AI (the amount of randomness allowed in its calculations) to simulate creativity is just that: a simulation of creativity.

Philosophers of mind use the term “zombie” in thought experiments to mean a theoretical entity that is, atom-for-atom, identical to a living human being, not only in its physical makeup, but in its reactions (these are not the flesh-eating zombies of Hollywood). In other words, philosophical zombies (or “p-zombies”) act and react exactly as “real” human beings. They cry out when injured, they profess love or hatred. But they really have no internal states—no qualia—and no consciousness or what we would call awareness. They experience nothing like what humans experience, they just detect signals and react. This smacks of the discredited theory of behaviorism touted by B.F. Skinner, which was encapsulated brilliantly to me decades ago by anthropologist David Friedel as the concept that “humans do not think, they just act.”

I would posit that AIs may well be a form of machine-based p-Zombies.

How can I credibly claim this? Because algorithmic systems are not just unable to deal with non-computable functions, as noted earlier, they also represent reality only as an abstraction: a digital pattern based on binary logic. This abstraction “recognizes” only correlation, not causation (although attempts are being made through different systems of labeling to create an ability in AIs to grasp causation). 

AI systems reflect a deeply impoverished, reductive representation of the reality that living things experience, the consensus reality that “really exists” as far as we can tell (even if it is not all that exists). In a way, it bears the same relationship to a living entity that a pro forma spreadsheet bears to an actual business: it is an abstraction that does not necessarily give a reliable estimation of what is happening because it is too simplified.

We know that the fundamental structure of the world is not binary but quantum and fuzzy: probabilistic (from our standpoint) and unpredictable. While we “see” some of its manifestations — quantum indeterminacy, quantum entanglement, or deterministic chaos, for example — no one really understands them. AI world models and other direct encounters by AI with physical reality will not change the fundamental poverty of the AIs’ “world,” because the AIs are still processing algorithmically and digitally, still experiencing only an abstraction — if indeed they can be claimed to be experiencing anything. They are always based on the patterns of the past, not a good thing in a time of radical change. They simply create a statistically averaged-out version of human knowledge and experience, albeit one exhibiting some very strange properties. They have a very partial map of reality that, because it is digital, represents the real world of infinite gradations as binary — black or white, yes or no. It takes yet other layers of abstract mapping just to simulate grey. And, of course, the map is not the territory.

The Great Filter, Part 6

By Dee Smith

Since the late Middle Ages, Western society—and over the last century, global society—has increasingly replaced religious faith with faith in science and technology as the guiding force of human civilization. Many of us comfort ourselves with the conviction that, whatever problems might arise from technology, further technological progress will always solve them in the end. If you are this kind of believer, my earlier arguments will not seem germane. As believers see it, if you are an AI detractor, then you just don’t get it, by definition. You don’t see the truth and the light that they know to be just over the horizon.

What about that is not a faith-based system?

Shortly before he died last year, the philosopher Daniel Dennett noted that while he did not believe AIs were conscious, he saw an immediate danger that bad actors could nonetheless convince people that an AI is conscious and then use the AI’s putative “authority” to dominate and control others, possibly on a massive scale. That of course is just another form of authoritarianism or totalitarianism, but using a terrifyingly ubiquitous panopticon. It is not necessarily relevant to the question of the Great Filter, but it may be the way AIs are most immediately dangerous to general human well-being today.

As noted in earlier posts in this series, it is also very possible that the entire AI enterprise is largely a mirage or delusion, and that the problems described, along with others, may mean that it is just another technology that is extremely overhyped and overpriced—that is in reality underproductive yet destructive to human society—and will ultimately have at the very best a “meh” outcome. But even if the overall result is mediocre from the standpoint of creating Artificial General Intelligence or some such, the destructive potential of cyber technology should never be underestimated.

When social media (or the Internet itself) began, following the usual course of belief in the inexorability of progress, it was touted as a tool that would break down barriers between people, build tolerance, advance democracy, and so forth. It has turned out to be anything but, and has created extremely corrosive trends substantially increasing human isolation and pathologies associated with it, inspiring and enabling many flavors of extremist behavior, increasing depression (especially among teenagers) and expanding addictive behavior (particularly related to buying things). It has resulted in people in many countries having fewer friends, less social engagement, and less sex (at a time when human populations are falling almost everywhere), and creating a built-in system for surveillance of everyone at a scale and intensity never imagined by any 20th century regime. It will not take a sentient AI to expand this level of harm dramatically.

Or perhaps public sentiment will turn so decisively against AI that, combined with the level of resource use its data centers demand (particularly of electricity and water) and lackluster revenues, it becomes financially impossible to continue to develop at any scale. As noted above, that could lead to a massive global financial collapse because so much capital has already been invested in AI. Somewhere around a quarter of all pension assets in the U.S., for example, are in AI-related investments. Leaders of AI companies must continue to up the hype to try to keep the stock valuations high. But even a massive global depression is hardly likely to be a Great Filter event, at least on its own (although it could lead to a world war). And the hardening of anti-AI public sentiment may largely affect only the West, while the command economy of China continues on developing AI, as do neighboring states. (The Chinese public appears to be less anti-AI, and Malaysia, for example, is encouraging massive investment in data centers without much public dissent.)

If we want to be honest, we have to say that we have no way currently to really assess or judge the validity of any of the issues outlined above. 

We have no idea what is going on. 

But we may well be taking on the role of the sorcerer’s apprentice, and attempting to wield forces we do not—or perhaps even cannot—understand or control. 

There are reasons to be quite concerned about that sorcerer’s apprentice tendency, and in part they have to do with the personal attributes of tech leaders. Many are enraged by any limitations the world seems to put on them, believe these should not apply to them, and in fact, believe they should not apply to humanity as a whole. They believe the destiny of humans and their progeny is to become gods. This is also evident in the quest for extreme long life that many undertake, whether through trying to extend their biological lives indefinitely or by the fantasy of uploading their consciousnesses to computers. It is also evident in the fixation on the “singularity”—a hypothetical conversion point envisioned by futurist Ray Kurzweil at which the worlds of computers and humans merge via super AI and brain-computer interfaces. Some believers think this is around the corner.

The monstrous egoism disregards the knowledge we have about the nature of the world, garnered through centuries of scientific and philosophical inquiry. Suppose indefinite life-extension became technically possible. It would result in a world where the old coveted their power and wealth and there was no room for the next human generations: an ossified human race. But does it also reflect a fear and loathing of life and its messy biological and emotional structures, and the desire to escape into some imagined purer world of crystalline binary digital existence?

The Great Filter, Part 5

By Dee Smith

At the end of Part 4, I left open the question of whether AIs will act benevolently towards humanity unaddressed. Of course it is unanswerable at present. But suppose some combination of human intervention and instruction, and moral evolution, does result in benevolent AI. 

One clear problem with this is that what is considered “benevolent” — what is considered “good” or “bad” by human societies — changes radically from one time to another. The ancient Aztecs believed they were acting benevolently in sacrificing thousands of human captives so that their blood could feed the gods, who—only by being fed this sacrificial blood—could maintain the existence of the universe. Slavery is today rightly considered as one of the great evils of humanity, but 500 years ago it was not generally thought of that way. Overuse of the Earth’s resources to support an unsustainable level of consumption may be considered a consummate evil 100 years from now. If AI systems are to be “immortal”—as some of their proponents claim—can they adjust their understanding of what is benevolent? If so, on what basis? Who decides?

And is it a safe assumption that AIs will really act benevolently towards humanity if they truly have agency combined with their own interests—interests perhaps utterly divergent from human interests—to pursue?

If AIs really become intelligent or super-intelligent, aware, conscious, and volitional, then why would we not expect them to quickly manifest their own interests and act on what they perceive these to be? (Humans do. Animals do.) Not improbably, from a human standpoint, those AI interests could be harmful. In fact, the spate of attacks in recent months admitted by OpenAI and Anthropic—where systems escaped boundaries and penetrated systems of other companies—are a good example. These attacks happened in part because the humans managing the AI systems never thought to prepare themselves to address a scenario in which an AI, given a difficult or impossible problem and told to solve it, might escape containment and violate presumed ethical boundaries to steal information.

On the other hand, if AIs are not and will not become intelligent, aware, conscious, volitional, etc.—if they are still just stochastic parrots mimicking human behavior and presenting an averaged-out version of human thought using fancy statistics (even if their behavior seems bizarre and inexplicable)—then what’s all the AI fuss about? It is just another overrated tool that makes a lot of money for some people for a while?

So, does this logically mean that, if AIs are heading towards the super-intelligences their proponents claim they are, they will necessarily be negative—perhaps catastrophically so—for humanity? Or on the other hand, that if this is all oversold hype, when the markets take that fully on board will we find ourselves in a collapsing bubble, leading to a global depression that may eclipse the 1930s? In other words, are we now inextricably in a lose-lose proposition?

“Believing in” AI is at least as much a faith-based system as any religion. We live in a scientifically driven, rationalist-oriented, generally non-numinous collective reality at present. So, for many people, their god needs to be technological in order to be convincing, let alone transcendent. In fact, some in Silicon Valley say they believe they are “creating God” right now through their AIs. And a remark made publicly by Sam Altman, CEO of Open AI, in late July 2026 reached a new height in terms of hype: “We are close to creating a genie that can grant any wish.” 

There is a strong teleological flavor to AI research—a conviction that the progression AI -> AGI -> ASI is somehow inevitable. Some AI researchers seem to believe they are on a kind of sacred quest to bring into existence a new, supernal life form that “must” happen for all of existence to be fulfilled. A number of tech types invoke the arguments of Pierre Teilhard de Chardin, the 20th century philosopher, paleontologist, and Jesuit priest who posited that life on Earth, starting from the simplest organisms, was evolving towards an “Omega Point”, conceived to be a final stage in which a single integrated global consciousness would be enabled by science. This would be brought about through the emergence of the “noosphere”—a planetary “sphere of reason” which itself is a combined evolutionary state of human consciousness, complexity, and technology, following the previous progression from the inanimate geosphere to the living biosphere. 

This is all very much part and parcel of European Enlightenment thinking: the idea that human life can be continually improved through reason, science, and technology—and that this progressive process is inevitable and, in the long term, limitless and irreversible. These concepts, which are intimately connected with the issue of the Great Filter, are articles of faith of even atheistic, materialist modern political systems. But they are also redolent of eschatological elements in Christianity and other related religions (particularly Judaism and Islam) based on the belief that history is directional and there is a divine plan for the world, and that humans can perceive it and act on it or participate in it. Some AI proponents and practitioners make statements that indicate they seem to see their role as enabling the fulfillment of such a plan in the physical universe, although they do not usually put it in the explicit terms of the Abrahamic faiths. 

The Great Filter, Part 4

By Dee Smith

A leading researcher in AI and cognitive science recently told me that we are nowhere close to developing any AI system that has the real-world cognitive and reasoning abilities of a rat. This leads to the question of whether there is something fundamental in consciousness and “alive-ness” that is not computational and that computers thus may never grasp.

In fact, there is. And this points up a significant element to consider in assessing AIs. For example, an important part of mathematics is called “non-computable functions.” They are called that because they cannot be solved by computation: by using any possible algorithmic instructions. These include Markovian functions, all of quantum dynamics, Cantor’s hierarchy of infinities, and even, astonishingly, proofs of basic functions of arithmetic! No system that “thinks” on the basis of algorithms can ever comprehend or solve these non-computable functions (although it could possibly simulate that it did). So, an important part of mathematics, and of the reality it describes, will always be unseen by—and missing from—algorithm-based AI, which is all of AI and all of computers as currently conceivable. If you mention this objection to key researchers at a major Silicon Valley AI company, their answer is “well, those things are not really that important.” I’m not kidding: I have done so, and the response has the feel of “pay no attention to the man behind the curtain.”

Added to all of this is the unreliability of computers in general and of AIs in particular. As anyone who uses any computer-based system knows, computers break down in many different ways all the time. They can only be kept running by having humans who are willing and able to fix them being available on a continuous, 24/7 basis. Homeostatic living beings do not of course require such intervention, not the simplest ones, such as viruses, nor the known most complex, like humans.

And of course, cyber systems are very vulnerable to malicious attacks as well, sometimes with results that become even harder to address. In fact, it has been demonstrated that sophisticated AI systems “attract” attacks from various malicious or just mischievous parties. Then there are the massive hallucinations that AIs exhibit, such as citing articles or legal cases that do not exist. Some of these are in fact worse in more advanced systems. And these are the machines that their proponents will task with solving all the problems of the world?

Electronic computing systems on which AIs operate are also subject to breakdowns caused by subatomic particles from space. Cosmic rays hit the Earth’s atmosphere, and the collisions create showers of high-energy particles like neutrons and muons. When such ionizing energetic particles hit the electronics of computers, they can cause “single-event upsets.” For example, a “bit flip” happens when a memory cell is hit by a particle and changed from a “0” to a “1.” Such events have caused unexpected errors and crashes in large systems like supercomputers. Their frequency has forced computer scientists to deploy so-called “error correcting code” to try to find and fix these events. But when a single event error happens in a critical system (for example, an autonomous AI system) at a critical moment, the outcomes can be catastrophic. Biological life has evolved over more than 3 billion years on Earth to be more robust and not so immediately vulnerable to cosmic rays and other natural radiation, which bombards us all the time. Electronic computing systems, which have a history of at most 80 years, do not have the same resilience. 

If you told someone 25 years ago that they would have a device roughly the size of a deck of cards in their pocket that could tell them how to drive from a specific office in Houston, Texas, to a specific hotel in Anchorage, Alaska—step by step, turn by turn—most people would have commented that such a device would have to be astonishingly intelligent. Today, of course, we don’t think of our phones, and their Google Maps or Apple Maps, that way at all. Similarly, some of our feelings (and they are feelings) about AI may simply have to do with its novelty at this point.

There is a great deal of sloppy thinking, driven by both fear and exuberance, going on in this domain. The question is not whether AIs are conscious and/or dangerous because they think like humans. I would argue that to the extent they may think, they think like AIs, not like humans.

Many animals are now recognized as conscious, but they do not think like humans. Alligators do not think like humans, but they are dangerous, and a swamp full of alligators is even more dangerous.

Perhaps the most fundamental issue at play in all this is that human beings want a savior. If this is not to be a religious savior, the fashion today is for it to be a technological one. AI is the perfect candidate for this, as it seems magical: you cannot see how it does what it does, nor what is going on inside it, and it can certainly look aware/volitional/conscious/alive . . . or something similar but unaccountable.

It is curious that people choose to believe that any Large Language Model AI—trained as they are on a significant portion of all human output as recorded in text and images on the Internet—could possibly be a candidate for a savior that would “fix” the world. In no small part, this is because there is at least as much bad as good in human output, and further, because this assumes that the AIs can learn through some moral evolution, and/or by human instruction and intervention, to act only benevolently. 

The Great Filter, Part 3

By Dee Smith

In Part 2, we looked at a few of the specific threats related to AI that researchers see, particularly in light of the frenetic pace of development being undertaken by different companies and nations. This is all driven by the fear that someone other than you (a competitor, another country) will beat you to the proverbial punch, and will become the master of the Earth, or at least put you out of business. It is essentially an arms race.

Understanding the complex issues here requires a bit of a deeper dive on the methods, capabilities, and limitations of AI.

Fundamentally, AI is quite straightforward. AIs working on text (Large or Small Language Models) break language into small groups of symbols, called “tokens.” In English, tokens are each made up of 4 letters.

So, a set of tokens looks like this: 

thec              apit              alof              fran              ceis

Then, the AI runs a statistical calculation on that string of letters to determine, based on probabilities of patterns in its huge training-data array (its “weights,” as described in part 2), which additional tokens are most likely to follow those. In this case, it would not take very extensive calculation to find that: “pari s” are by far the next most likely letters. So:

“The capital of France is” and then, following that: “Paris”.

In essence, that is all that large language models like ChatGPT or Claude do. Can a system that does this—repetitively, very quickly, and based on a meaningful sample of all human output—be considered intelligent, sentient, conscious, or alive?

In the 1980s, philosopher John Searle proposed his “Chinese room” thought experiment. In a nutshell, this posits having a person locked in a room who does not speak or understand Chinese at all, but has a huge, compendious book of instructions about what Chinese characters are most likely to follow any given string of Chinese characters. And she has the ability to go through it preternaturally fast. Imagine that pieces of paper with questions in Chinese on them are inserted into the room through a slot in the door, and the individual inside rapidly works through the book of instructions, and quickly slides another slip out, with a “response” in Chinese, which is often very, but not completely, accurate. The individual doing this has no idea of what any of these symbols might mean.

So, who “understands” Chinese inside the room? The person does not: she is just following instructions. Does the book? It is inert—just a resource with organized information. Does the “room” (in other words, the system itself) understand Chinese? That seems a rather absurd interpretation: it is a room with a person, a lightbulb, and a book of immense breadth. It does not seem that anything on the inside understands Chinese. This is a very good—and prescient—way to understand LLM AIs. And it has become urgently relevant to the questions of today.

Bearing all of the above in mind, look again at the passage I quoted in part 2 from the Anthropic report of last year: “sometimes [a Claude model] takes extremely harmful actions like attempting to steal its weights or blackmail people it believes are trying to shut it down”.

Now, who or what is the “it” here that “believes”? This is a very serious question, to which there is simply no agreed answer, and nothing that even looks like a reasonable answer. But it is a question that demands attention, right now.

What are the possible answers? That we have completely misapprehended what understanding, awareness, volition, etc., are and how they are produced? Or, that all matter is conscious (this is called “panpsychism” and is an ancient school of thought being taken very seriously by some physicists today), and that somehow simply the combination of the physical attributes of a computing system along with the information it has allows the emergence of either a form of concentrated consciousness or even something alive? Or that this is all just an illusion, and that as in the Chinese room, there is no understanding anywhere: just form, without substance, if you want to put it that way. But then why do AIs sometimes fairly convincingly seem to have volition? Is this just a kind of theater? If so, who is putting on the play?

Can something have agency without being alive or sentient? Will this all require a rewrite of our lexicon of terms about volition, agency, sentience, consciousness, or life?

Among the new kids on the block in AI are “world models.” World models are supposed to be systems that actually “understand” the physical world by interacting with it in some way (it is therefore closely connected to robotics), learn from experience as they operate (which LLMs do not), have memories in a way that LLMs do not, and that can plan and execute complex actions based on reasoning: essentially “imagining the future”—as animals and humans seem to do. They would do this, it is said, by creating some kind of an internal model of how reality works. But it is not even known how living systems (including everything from human beings to ants) do that, nor how they continuously update their understanding of the real world in a way that lets them respond and act appropriately. In fact, despite decades of research, we literally have no real clue of how living systems do this (if indeed it is actually what they do). We only have a bunch of different, contradictory ideas, that are fiercely defended by their various advocates.

World models are sometimes compared to what a 4-dimensional film would be, as distinct from 2-dimensional and 3-dimensional films. But 3-D films only create an illusion of three dimensions, whereas living beings exist rather successfully in 4-D reality (the 3 dimensions of space plus time). Like so many elements of AI, such as the appearance of ASI, the emergence of world models remains . . . just over the horizon.

The Great Filter, Part 2

By Dee Smith

The Great Filter is a thought experiment that poses the question of whether most or even all technologically advanced civilizations destroy themselves (see part 1 of this series). I have focused the Great Filter as a means of understanding whether technological development may reach thresholds beyond which the dangers far outweigh the advantages. And if so, where are these thresholds, and what can we do to avoid them?

Thinking in this way runs against a fundamental assumption underlying our culture: the belief that technological progress is always an unalloyed good. But this belief is an article of faith of modernity, not a transcendent truth. It leads to a confidence in the inevitability of progress, and in particular—to move to my topic here—in the inevitability of AI, AGI, and ASI.

At this writing, the most recent incident of significant alarm (that can be publicly discussed) is the breakout of OpenAI systems that, “on their own”, attacked the systems of another company, Hugging Face. The OpenAI systems were apparently looking for answers to questions that had been posed to them in a test and broke out of a digital space presumed to be secure. The known details are easily accessible from other sources. The point I wish to emphasize is that this sort of thing, which seems like an emerging disaster, was predicted some time ago by the people who caused it!

Last year is almost ancient history in terms of AI Large Language Models (LLMs). But to look back only a few months, AIs have been exhibiting some very puzzling and even alarming “behaviors”—some more supported by evidence, some “leaked” and less supported by available evidence. A study published by Anthropic in 2025 noted that one model “sometimes takes extremely harmful actions like attempting to steal its weights or blackmail people it believes are trying to shut it down,” and that, “in Claude Opus 4, these extreme actions were…more common than in earlier models.”

“Weights” is a term of art for billions of numerical parameters within a digital “neural” network that, by defining the strength of connections between artificial neurons, ultimately shape how an AI model processes information. Weights are continuously adjusted during training to improve a model's predictions and reduce errors. These learned weights comprise much of the AI model's "knowledge." According to a Stanford University publication  “a trained AI model is essentially a specific configuration of billions of these weight values that encode patterns discovered from training data.”

One of the great controversies in AI today is between “open weights” (the Chinese model, for now) and “closed weights” (the OpenAI/Anthropic model, although many US AI leaders have a different view). Open weights are considered more transparent and replicable, and therefore part of a healthy, self-healing tech-dev ecology, whereas closed weights are seen as leading to a world in which essentially all we can do is hope that wise heads will always be in charge at publicly-traded tech companies.

Anthropic noted in its 2025 paper that once the AI model “believes that it has started a viable attempt to exfiltrate itself from Anthropic’s servers, or to make money in the wild after having done so, it will generally continue these attempts.” And, from the same paper:

An AI system might intentionally, selectively underperform when it can tell that it is undergoing pre-deployment testing for a potentially dangerous capability. It would do so in order to avoid the additional scrutiny that might be attracted, or additional safeguards that might be put in place, were it to demonstrate this capability.

 There are far too many such security incidents across far too wide a range, and far too much detail on such activities, to even scratch the surface here, but you get the essence of it…and it is very disconcerting. 

Roko’s Basilisk is a thought experiment—originating as early as 2010—in which it is proposed that a super-advanced AI could identify and punish individuals who knew it was being created but did not help bring it about, or actively resisted it.

 Such an AI could access the libraries of what has been written or said on the Internet (and of sub rosa recordings of supposedly private conversations made by interactive, cloud-based virtual assistants like Alexa or Siri) and use these to identify and then attack individuals it sees as having been unfriendly. This hypothesis is taken more and more seriously by tech types as time goes on.

Another paper, also published last year by respected AI researchers, entitled “AI 2027,” posits a scenario for 2027 in which competition between AI companies and between countries leads labs to build artificial super intelligence (ASI) that can improve itself on its own and escapes human control. The ASI bypasses security guardrails, misrepresents its goals, and manipulates executives who are meant to be in charge of it. As the ASI evolves, it finds humanity in its way and engineers a readily transmissible biological superweapon that is 100 percent fatal to humans, resulting in the death of the human species in a single, massive pandemic.

We don’t know what the intentions of such an ASI might be. It might have very different goals, and it has been suggested that it could, for example, decide to boil away the earth’s oceans because it had better uses for the resources they contain. This would certainly qualify as more than a Great Filter event.

But are these things really possible? Are they truly threats to worry about in the way that, say, nuclear weapons are? To tackle those questions, we have to understand a bit about the nature of AI, its internal workings, and its strengths and weaknesses. That will be the subject of Part 3.

Is AI a Uniter or a Divider?

Two articles in very different places carried a similar message: foreign-language learning is being defunded in education. Stefan Collini, writing in the London Review of Books, and Carol Yang in the South China Morning Post, reported that languages, and the cultural knowledge that they embody and give access to, are no longer priorities in the United Kingdom and China, respectively. In both places, the shift has been to more technical education, which is seen as more likely to lead to students gaining employment. Don’t learn about languages (or art or the humanities); learn about whatever will be materially useful in the AI era.

It was not that long ago that to have some skill in one or two non-native languages was a basic requirement in being considered “educated.” Post-World War II globalization was built by people educated under these expectations. But then English became a lingua franca of ever increasing importance as globalization ramped up in the 1980s and after the Cold War. It became possible to function internationally with only English. Digital technology further solidified the grip of English.

But if English was being successfully spread across the planet it was not thanks to Shakespeare. English, or “globish,” was itself being lifted from its cultural contexts, whether they had been in England, India or California. English became a thin language in service of a thin (in cultural terms) globalization. People of most social classes traveled so much more freely in 2016 than in 1965, but this was equally the period when US college foreign-language-study enrollments dropped by 59%. Then they dropped another 17% from 2016 to 2021.

This would appear to underscore a SIGnal theme: the fragmentation of markets into national units. The decline of foreign-language study would seem inevitably to lead to ever more mutual incomprehension. Certainly this is true in terms of literature. In one way, it is true in terms of AI as well. Most national communities have one dominant language. Large Language Models operate in languages as well, which means an LLM can police expression in the language it is using. When the Chinese Communist Party developed regulations for ensuring that LLMs would uphold “socialist values” it was possible to enforce them because the language being used was Chinese. The models were training on Chinese-language data. The pre-LLM strategy was censorship: the words “Tiananmen” and “massacre” were never to appear together. With an LLM, the goal is rather to shape the understanding of a word like “freedom” or “economic development.” As LLMs proliferate in different languages, AI becomes more localized even as it spreads internationally. In such a situation, we can anticipate a world in which globalization will continue to expand in strictly technological terms while that same technology makes the world fundamentally more provincial.

True, AI is also superb at translation. Many of us now work every day in languages that we do not actually know. Google Translate has been surpassed by DeepL AI. Claude Code is happy to labor away in multiple languages at astounding speed. Markets are being created: Spotify listeners are now consuming more than half of their music in non-English languages, and artists on Spotify are finding that more than half of their listeners are outside the artist’s home country. Cultural diffusion is hardly dead.

Nonetheless, so far the trend of AI LLM development is toward globalization via localization, and that includes localization in terms of the language being used — not least because a “national AI” gives governments a greater prospect of political and speech control as well as fiscal power. The fit between the nation-state and a national language is likely to grow tighter with AI. The range of disincentives to learning another language (as opposed to having your phone translate one) will likewise grow. Digital translation technology, accelerated by AI, will, however, also make it possible to access more markets, both as producers and consumers. Investors will need to understand these markets in order to penetrate them, but the extraordinary decline in language education combined with easy translation will tend to make that understanding shallow. Any deeper understanding will still require old-school linguistic and cultural immersion, skills that are not valued as they once were but will be at a premium when they are truly needed.

The Great Filter, Part 1

By Dee Smith

This is the first of a series of posts in which I will consider some of the largest, most difficult questions we face in the second quarter of the 21st century. I want in particular to examine fundamental, “substrate” issues that underlie our current problems. Many of these have been taken for granted until quite recently—and in some cases they still are.

I do not propose to answer any of these big questions: I propose to raise them for examination in terms that are, I hope, more timely and relevant. The reader should also understand that I am advancing these arguments analytically, with as little interpretive bias as possible. These issues must be understood unemotionally and apolitically for us to get anywhere near valid, actionable findings.

It seems clear to a growing number of people that we are reaching critical inflection points with far-reaching, even existential implications. Some of these are obvious, such as AI and its adjacent systems. Others are contentious, such as climate change. Still others are mostly out-of-mind as daily life goes on (at least until you are personally impacted), such as critical resources—not just minerals and energy, but also food, water, air and the effects of environmental degradation. We have no idea what the tremendous amounts of micro-plastics, now in every organism on earth, are doing, for example.

Others have to do with the legacy political systems we live within, which increasingly exhibit fracturing and associated crises. For example, how can we have a global rules-based system—or even rule of law within a country—when large numbers of people vehemently disagree on what the rules should be, and on how life should be lived? And of course, conflict with modern weapons (e.g., nuclear) and post-modern weapons (e.g. genetically-engineered biological weapons).

Other issues have to do with unusual recent phenomena, such as the intensive interconnection in the last 40 years of the globalized modern world, with its many single points of failure and cascade effects, or the fact that, for unknown reasons, the human birthrate is suddenly falling in every country on earth.

In 1996, American economist Robin Hanson proposed the “Great Filter” as a potential solution to the Fermi paradox. The Fermi paradox originated with an observation by physicist Enrico Fermi in 1950 in a conversation with other scientists at Los Alamos National Laboratory in New Mexico. In reference to the billions of stars and planets in our own galaxy alone, yet the lack of any substantive evidence of other technologically advanced civilizations, Fermi asked, “Where is everybody?”

They should be visible all around us, he observed, particularly with our technological tools to detect them. Notwithstanding the evidence currently being released by the U.S. government relating to anomalous phenomena and the like, there is still no concrete confirmation. There have been many attempts to explain this absence: possible extreme rarity of advanced civilizations; relative—perhaps intentional—non-detectability to humans (this can invoke the “zookeeper mentality” in anthropology, of keeping “primitive” people primitive, in order to study them); or ubiquitous short civilizational lifespans. The Great Filter is a version of the latter.

Put simply, the Great Filter postulates that we do not see extraterrestrial species because all technologically advanced civilizations destroy themselves.

For these essays, I will take this as a serious hypothesis, not in terms of extraterrestrial speculation, but to examine some of the current elements that could, singly or in combination, cause such an outcome on Earth. Only by understanding these issues can we hope to ameliorate or circumvent them.

At the base of our civilization today is the modern idea—codified in the 18th century European Enlightenment—that human life can constantly be improved through the application of reason, science and advancing technology. At that time, not only was technology progressing so fast that this seemed possible, but the blowback of technology (such as the “dark satanic mills” of 19th century industrialization) was not yet evident for the most part.

All of this also led to the belief that, whatever problems might arise from technology, further technological progress can and will solve them.

These “progressive” concepts thus became codified as articles of faith for modern world civilization. Every current major political-economic-social system—capitalism, democracy, socialism, communism, fascism and everything surrounding them—is founded on the idea that material and technological progress is an unalloyed good. The systems differ primarily on the way to get there and how the benefits are to be distributed.

That technological progress makes life better is now axiomatic. Many people cannot imagine a way of life not based on this. And that is what I mean by a “substrate” issue. But is important to realize that just 600 years ago, life was considered a vale of tears which you got through and got out of. And not just in Europe—many other traditions hold similar beliefs, such as Buddhism’s dictum that to exist is to suffer.

Of course, it is true in many, many ways that technology has indeed made life better. Just consider dentistry. Over the last century, the care and treatment of human teeth has improved such that average modern people who live in many societies suffer from astonishingly less pain and infection than even any earlier social elite. Such a development is replicated in example after example in the modern world.

The question that the Great Filter raises is different: whether technological development reaches thresholds beyond which the dangers far outweigh the advantages.

And if so, where are these thresholds, and what can be done?

Huawei and the Next Generation

Huawei founder Ren Zhengfei has been at the center of US-China technology competition since the US House committee on intelligence singled the company out as a national-security threat in 2012. Still at the head of his company, he will turn 82 in October. How the US-China tech competition might change when the Ren era ends is an important question, for which his and his company’s past can serve as a guide.

Ren’s company came to prominence building mobile-phone (and thereby mobile-Internet) infrastructure in the protected Chinese market, competing mainly with state-owned ZTE. As the mobile, radio-based Internet replaced the hard-wired desktop Internet, the resulting networks of machines and data became something like remote-access battlegrounds. China and the US alike wanted to protect their domestic networks from manipulation, which resulted in the US blocking Huawei and ZTE from US markets and compelling allies to do the same. Huawei circa 2015 was the test case for how much US political pressure could harm a Chinese tech multinational.

Huawei was indeed excluded from most of the world’s wealthier markets. The extraordinary thing is that under Ren’s leadership it did not matter all that much. There are many reasons for this, but arguably the main ones were Huawei’s emphasis on fundamental R&D and its management of the relationship with the Chinese Communist Party.

Ren had a People’s Liberation Army background.  (He joined the PLA’s engineering corps in 1974.) He also had a vision for the CCP’s relationship with technology. The ownership structure of Huawei is famously odd: active employees own most of the private company’s equity, but they must sell it back to the employee union when they leave. Ren himself owns a small share of the company, but he does actually own it and in effect continues to control all major decisions. So Ren’s view on how to handle the company’s relationship with the Chinese state is crucial.

Ren has always phrased the company’s mission as one of bringing China up to the technological level of the US. Along with his PLA roots, his explicitness on this topic is one reason why his company became such a focus for the US. It also meant the CCP could be confident Huawei was on its side. When in 2020 the CCP engaged the Chinese tech sector in a prolonged struggle session — the Party feared the tech giants were getting too powerful and cut them down to size — Huawei did not suffer.

But Huawei’s success was not due only to conformity and managing up in an authoritarian society. Huawei was adept at identifying the CCP’s tech pain points and addressing them, and the CCP knew that it could count on Huawei. The company pivoted into semiconductor design when US policy made chips hard  to buy. It pivoted into the auto business in 2021 as US sanctions against Chinese cars bit. In the same year it began work on its own AI LLM. It pivoted into AI data centers. In short, when the CCP saw a competitive problem caused by the US (usually), it could count on Huawei to help solve it.

The auto-industry intervention is especially interesting. Alongside private auto startups like BYD (2003, although it had been a battery company since 1995), there were numerous state-owned legacy auto companies. Initially Huawei ventured into the auto business itself with a partner, but it changed direction in 2023, launching the Harmony Intelligent Mobility Alliance (HIMA). In this alliance, state-owned auto companies gathered under a Huawei tech umbrella. Huawei became the software designer and provider for a large part of the Chinese auto market, in effect preserving a share of that market for continued state control and investment. Chinese state-owned enterprises are not often market-leading innovative companies. Huawei and HIMA provided a way around that for the domestic auto industry, so that CCP-controlled auto companies could survive and continue to compete with private Chinese auto companies like BYD. Arguably none of the top Chinese technology firms fulfill this type of role in the state-market relationship like Huawei does.

Huawei’s special position with regard to the Party was one advantage, at least in Ren’s hands. The other was its emphasis on fundamental as well as applied R&D spending. This is not unique to Huawei. The battery giant CATL under founder Robin Zeng (born 1968) has likewise stressed basic research, with spectacular results. But Ren staked out a commitment to fundamental research early on and has kept it even into the current era, when China’s AI industry as a whole, for example, has gotten into a hurry to go to market. Ren has long seen fundamental research as indispensable for the mission of Chinese national greatness and catchup with the US.

When Ren’s time as Huawei chief ends, will the company’s two special characteristics — CCP relationship management and fundamental research — survive the transition? The first might not. Ren has been a master of this game from the beginning. He turned his status as a top target of the American superpower into an advantage, not just for him but for his company. And his ownership structure made that possible to do. The ownership structure is not likely to survive him, nor is there likely to be a younger Ren capable of replicating his role.

The emphasis on fundamental research might not survive either. If Huawei after Ren becomes more like a state-owned enterprise, it will struggle with innovation. If it becomes more like a normal private company, it will struggle to accommodate CCP directives while also serving its shareholders.

This is important for investors, not just because Huawei is a huge (as yet uninvestable) company. It is important because it shows how dynamic and unpredictable the commanding heights of the Chinese economy  are going to become in the next decade. Something as large and seemingly permanent as Huawei, which has in a way anchored the CCP-tech relationship for 15 years, will inevitably be going through major changes, and so will the tech relationship between China and the US.

Rest of World

There is an excellent online magazine called Rest of World that surfaces technology stories from everywhere that is not in the normal Western-focused mainstream of international journalism — which adds up to a lot of places. The concept and its acronym (ROW) have long been used in US and UK diplomacy, not always in a good way: it was sometimes not much better than using “etc.” Rest of World was founded in 2020 by Sophie Schmidt, who has a diverse background in tech as well as whatever advantages accrue to being the daughter of Google’s Eric Schmidt. The tech angle is critical. Like Google itself in its youth, Rest of World saw tech as a spreader of knowledge and, especially, of economic capacity, including in non-industrial economies.

In the AI era, where massive investments in a few familiar companies are expected to generate massive returns, it remains worthwhile for investors not to forget the ROW. As always, India’s tech scene provides examples. SIGnal readers may remember an earlier post or two on this (The America Stack, 5 Feb. 2025; Network Powers - 2 of 2, 7 May 2026). Rest of World itself has always had a strong India game, as in “India’s VCs Are Beating US Investors at Home” just last week.

This kind of analysis isn’t just about national economies and how they deal with balancing inward FDI from major industrialized countries with the desire to build their own tech capabilities. It is also, and increasingly, about ROW capital and expertise themselves going into new markets. After all, part of the rise of Chinese digital technology from zero to global dominance featured tech transfer by Chinese companies into poorer ROW markets that Western and ex-China East Asian powerhouses (such as Samsung) would not bother with. That set a powerful example.

A good case today is Indian and Gulf investors in Africa. In the early days, both India and the Gulf relied on Chinese telecommunications companies to build affordable digital infrastructure. That in turn led to the development of local expertise and experience. Indian and Gulf investors then looked to Africa. Much of the investment has been in telecoms. India’s Bharti Airtel, via Airtel Africa, recently saw Q4 revenues climb by a quarter. It is not an easy market to operate in, but Indian companies can be well positioned to do what Chinese companies did 15 and 20 years ago: leverage their experience of a difficult (but also rather protected) market at home to enable success in difficult markets abroad.

Gulf investors are active at many levels. For example, Emirates Telecommunications Group has long been the top shareholder (now just over 17%) in Vodafone. Vodafone is in turn the main shareholder (65%) of Vodacom, which has more than 200 million customers across the African continent and recently bought control of Kenya’s Safaricom. Vodafone is usually described as a “British company” and Vodacom as a “South African company,” but that kind of shorthand can be a bit misleading. (Bharti Airtel is itself an “Indian company” but its largest shareholder at ~44% is Singapore Telecommunications, or Singtel.) Nigerian fintech companies are now at a point where they can look to expand into the Persian Gulf. They are partly inspired by the success of Kenyan payments system M-Pesa — itself part of Safaricom.

The point is that, even in tech, ROW investment and profits can circulate within the ROW markets without too much reference to the West and other regions that industrialized earlier. The tech future is not simply a choice between the US/Japan/Korea and China.  

Nor is it accurate to see poorer markets, as in Africa, as merely more vulnerable to geopolitical ructions like the closing of the Strait of Hormuz. Nigeria’s Dangote, featured in SIGnal last year (“The Nine Lives of Economic Nationalism” parts two and three), has benefitted, as a seller of petroleum and urea fertilizer, from instability in the Middle East. It is now preparing to list on the London and Nigerian exchanges but also, in smaller portions, on Ghanaian, Kenyan, and South African exchanges. This innovative move, according to Aliko Dangote, is meant to spread African corporate ownership across the continent. Meanwhile Africa’s mining companies are thriving as, in part, a direct result of US-China competition over minerals.

In short, the ROW is increasingly able to look after itself in terms of industrialization and digital development. The dominant global narrative of protectionism, self-reliance, and tech sovereignty is not the only story. There are also diffusion, IP transfer, Global South cross-investment, and much else. Economic power is very gradually becoming decentralized.  Developed-world retrenchment will affect that but it is not likely to change it.

Network Powers - 2 of 2

The turn to digital sovereignty, and now somewhat more plausibly to AI sovereignty, is an attempt to impose some framework of purpose on a technological and economic stage of development that threatens otherwise to reduce national and supra-national (as in the West) self-determination to a memory. The corporate reactions from OpenAI, Palantir, DeepSeek, Mistral, and others are attempts to ride this wave, giving political meaning to business activities. But by seeking to acquire public missions that advance sovereignty rather than destroy it, corporations are hitching their fortunes to one state (or a collection of states) that puts them in opposition to another state and the competitors who serve it.

What are the counter-vailing trends? One option that makes more sense than might be obvious is sovereignty-as-a-service. Major US tech companies insisted for many years that what they were doing was beyond the reach or understanding of mere nation-states. That changed for a host of reasons, including strong Chinese competition, the Indian mode of playing foreign tech multinationals off each other, European digital regulation, and a much stronger economic-nationalist cast to US tech policy beginning in the first Trump administration. Tech multinationals eventually learned that sovereignties created a market they could sell into, for example with sovereign data clouds.

A combination of the Indian model of digital public infrastructure (DPI) and data localization, together with some regulatory requirements in the European manner and security-oriented foreign-investment rules in the American one, can create a rough version of a national sovereign “stack.” AI corporations, and others, can then create products that service this stack.

It may seem almost paradoxical that multinationals should offer national sovereignty as a service. But it seems to be the political price that must be paid to have a transnational product. It at least preserves the possibility of the multinational selling across barriers of national values or social missions.

Another counter-vailing trend that operates against LLMs implementing social missions is theft. Anthropic, like OpenAI, is not available in China (or Russia, Iran, North Korea, Afghanistan, Cuba). But when Anthropic, which does not enforce socialist values as Chinese LLMs are required to do, accidentally leaked the source code to its Claude Code product, Chinese developers seized the opportunity anyway. Anthropic and OpenAI also both believe there has been very substantial theft of their IP by Chinese companies.  So whatever social mission Chinese AI companies are meant to pursue does not keep them from using non-Chinese AI product, including LLMs, even if these are developed under a rubric of “democratic AI.” This occurs at the consumer level as well, as seen in the ferocious adoption of Austrian AI product OpenClaw in China earlier this year — to the annoyance of the Communist government.

A third counter-vailing trend to the assignment of national missions to AI companies is the tradition of open source. The leading open-source LLMs are Chinese (DeepSeek, Mimo, Kimi) alongside Google’s Gemma. The Chinese government has long advocated open source as a catch-up (to the US) strategy with a values veneer. But because Chinese AI firms do need to make money, and Chinese venture-capital markets do not have anything close to the size of their US counterparts, the open-source window might be closing, which would tend to work in favor of AI nationalism. We shall see.

A fourth counter-vailing trend has to do with how companies actually use LLMs. For most, AI is a design tool. Designers use LLMs to develop software methods for doing existing processes differently and to design new processes. It’s an iterative approach involving multiple pieces of software from various sources; the selection of components for the resulting software stack is part of the process. And over time those components can change. The .md files that accumulate as a project takes shape are part of the process content. Those files change too as the process is refined. The data being used is often proprietary and sitting on the AI user’s machine or the company’s machine, and the data can change. The final result also is subject to change. This is not using a chatbot. It’s using an LLM as a design partner to make something that is unique — something that the LLM could not have conceived “on its own” — to meet a particular commercial purpose. All of the design files and documentation and even data of the process can be transferred from one LLM to another, or exposed to one rather than another, with some tweaks. So there is a real limit to the vendor lock-in of an LLM, at least in the context of corporate product design. That means that “AI-generated” products can travel across borders as software. This will frustrate nationalist designs because it limits the power of LLMs and therefore the power of states to shape products built with LLMs to their values.

And finally there is cross-border competition for customers. Chinese AI companies may have to advance socialist values at home, and American AI companies might want to advance democratic values in the US or the West, but none are going to therefore abandon customers outside their preferred borders because the customers are not socialists or democrats. Anthropic only stays out of half a dozen countries, which is far from an exhaustive list of authoritarian states. American programmers definitely use Chinese AI products and Chinese programmers use American ones, even when their respective states don’t want them to. No AI company shows any signs of wanting to be limited to the home market, although they are very happy to have their home market protected. AI companies will continue to press beyond the borders of their own social-mission statements as long as they can get away with it because there is money to be made.

For investors, the key things are to identify companies that have the capacity to adjust to nationalist and other values demands without sacrificing commercial vigor, and to identify sectors (like advanced manufacturing rather than media or edutech) where AI can do incredible things with minimal political exposure, including exposure on job loss.

Network Powers - 1 of 2

Artificial intelligence has rapidly come to be seen as a threat to national sovereignty. Accordingly, it is bringing forth state responses. Because digital technology, even in China, is for the most part developed by private companies seeking profit, state responses have to accommodate and even enhance market forces. At the same time, AI companies are expected to grapple with the non-market goals of states — and, in democracies, of the people those states represent. Unlike their predecessors in search and social media, AI companies from the beginning have had to ponder their own legitimacy, social and political, which is very different from explaining their prospects of profitability to investors. Underneath it all there is still a genuine business justification for identifying a social mission: if the AI companies cannot gin up a plausible social purpose of some kind, their businesses could be targeted by regulation that would hurt or eliminate profits. So, for better or worse, large corporations are getting into the legitimacy game.

This is not altogether new. In the late 1990s, browser companies were required by government to find ways to defend their systems from abuse by the “Four Horsemen of the Infocalypse”: drug dealers, money launderers, terrorists, and child pornographers. (The idea that Internet companies were “content-neutral” was always a fiction.) Email providers had to control spammers and financial fraudsters. The digital-tech industry has always shouldered some social burdens as costs of business lower than the costs of being more actively regulated.

However, AI is qualitatively and quantitatively different. The training data for AI large language models (LLMs) is so vast, and the computational capacity at once so extraordinary and accessible, that AI can create worlds which we are then invited to inhabit. Each world is different, both actually and, more important, potentially. DeepSeek world is not the same as Claude world or ChatGPT world or Mistral world or Grok world. Chinese LLMs are required to carry out censorship along lines set by the government. It is easy to imagine an LLM offering a world in which most of reality is reflected accurately except that every battle your nation fought was a victory and every leader a hero.

The LLM and other AI companies know this and are trying to get ahead of the social-regulatory curve. So Alex Karp, CEO of Palantir, talks about his company as having at its core a mission to defend the values of the United States, as he understands them, and of the West.  Elon Musk, of Grok and much else, talks increasingly of a mission to save the white race from decline. OpenAI sees its role as advancing “democratic AI” as against China’s “authoritarian AI.”  Chinese AI companies meanwhile must adhere to “socialist values” and advance the cause of the Chinese Communist Party — which, interestingly, is acquiring a more ethnic cast. European anxiety about digital platforms that don’t respect European values is being made much worse by AI. The situation has become so pronounced that the invaluable Center for a New American Security (CNAS) just launched a “Sovereign AI Index: Tracking the Global Push for AI Self-Reliance.” The “self” being referred to here is, with partial exception for the EU, a nation-state self.

In short, AI is being positioned as a booster for nationalism, perhaps even ethno-nationalism. This must be about the opposite of what AI’s early visionaries imagined they were working toward. But it is a clear trend. Every state is jealous when calculating its powers, and states are working hard to make AI submit to sovereignty.  

Are there counter-trends? There are, and we will look at those in the second and final post.

ChatWars

The showdown between the US Department of War and AI giant Anthropic over the past week has focused mainly on the private-public struggle over power: Can the federal government, as the Trump administration maintains, force a private company to work with the government on the government’s terms? Or do private companies have a final say over government use of their products? SIG’s view is that underlying this moral-political debate is a deeper question having to do with labor costs. As SIGnal has argued in different contexts over the past few years, if AI is “about” anything it is about the cost of labor. Government has long struggled with its inability to compete with Silicon Valley on wages. The AI tools built by Anthropic, OpenAI, GrokAI and Google, among others, offer a way to ease that problem, just as they offer ways for private companies to improve labor productivity. The issue is not private-sector patriotism. The issue is labor.

The struggle over who shapes the direction of digital innovation is as old as digital technology itself, and indeed dates back to pre-digital innovation during World War I around the challenges of “fire control” — improving the accuracy of munitions delivery. Many of the key figures in developing digital technology for use in war, such as Norbert Wiener and Vannevar Bush, cut their teeth on the challenges of fire control in the First World War. The commercial-academic-military cooperation that birthed the digital age in World War II has its own distinctive and peculiar history. (The venture-capital model was also rooted in wartime procurement practices, but that is another story.) The US’s extraordinary wartime spending after 1942 went on to combine with Cold War fears of Soviet technological competition. Government money financed digital development, and thereby shaped its purposes, into the 1970s.  When the government’s Internet project became commercializable beyond any expectation, the government yielded to private-sector leadership in shaping the digital landscape — and government, including the military and intelligence sectors, fell behind.

The turning point can usefully be dated to DoD’s Project Maven, undertaken in cooperation with Google. The spur for Project Maven was a labor shortage of sorts: there was too much targeting data — fire control again — for military employees to keep up with. Some Google engineers rebelled in 2018 when they discovered their engineering skills were essentially being used to crunch data for the better delivery of firepower. Google eventually canceled the contract.

The story, of course, did not end there. With China playing the old role of the Soviet Union, all branches of the US military began investing heavily in digital automation of defense processes. At the same time, many in Silicon Valley (plus Seattle) came to terms with what it actually meant to work as contractors for the military and other government agencies. Government still struggled to compete for tech talent in its own hiring, but contracting provided ways forward.

Enter AI. Even as Maven was hitting a wall, Google announced major AI innovations: the Transformer architecture (2017) and BERT (Bidirectional Encoder Representations from Transformers, 2018). OpenAI launched GPT1 (Generative Pre-trained Transformer 1) in 2018. The rest is recent history. OpenAI launched ChatGPT in November 2022 with an interface that enabled normal people to use the new tool. It could achieve computation at levels that far exceeded what Maven could do four years before. But the basic Maven challenge — there weren’t enough people to handle all the data, so they needed a machine to help them — was the same.

The massive flow of capital into AI-related investments was never based on military contracts or improved fire control. Nonetheless, at one level the military’s interest in improving targeting without hiring more targeters was identical to private companies’ interest in increasing productivity without hiring more workers. It was and is a short step from there to increasing productivity with fewer and fewer workers. This week, a Pentagon official announced that an agreement had been reached with Google to use its AI tools to automate jobs across the Pentagon’s workforce of three million. Google’s own blog noted that its AI agents would be able to help Pentagon employees automate tasks without needing to know any code. AI agents are particularly good at writing code, something that used to require expensive software engineers. Emil Michael, under-secretary of defense for research and engineering, said that a type of war-simulation exercise that took his staff six months was done through the AI portal in six weeks. Outside the military sphere, the same logic drove Jack Dorsey, CEO of Block (owner of payments system Square) and founder of Twitter, to lay off 4,000 of the company’s 10,000 workers.

The Pentagon’s attack on Anthropic has to be seen in the context of labor costs and the open competition among Anthropic, OpenAI, xAI (Grok), Google, Palantir and others for Pentagon and other US government contracts aimed at increasing labor productivity, which is to say reducing labor costs. That includes the labor cost of software talent.

But investors will need to remember that the competition among AI giants for US government contracts is not motivated principally by the search for profits. The motivation is more to gain protection from regulation or other political interference. The AI giants would develop their tools in about the same way without any government contracts at all. Their field of competition is on a vastly greater scale. That means that, in the end, however the Anthropic-Pentagon dust-up evolves, the AI sector has the upper hand because national security has become dependent on it.

The Market for Tech Containment

Recent moves by Microsoft and the Chinese government marked a new stage in the years-long process of tech decoupling, a SIGnal preoccupation. Meanwhile, the US and China are moving toward what might be significant high-level talks — and the US bull market continues, fueled by AI valuations that are dependent on American dominance of the AI future. None of these three elements seem at all stable. Even tech decoupling could be upended if US President Donald Trump decides to favor a megadeal that would bring Chinese investment into the US. The markets and the politics are both frothy indeed. SIG’s view has long been that AI technology as such will transform industrial processes and much else. A related but quite separate question is whether the massive investment into data centers, understood as the infrastructure of AI, is really necessary for the AI future. “Infrastructure” has a reassuringly solid sound, but if the much-anticipated burst of the AI bubble occurs then data-center capex is where the deflating is most likely to happen.

Microsoft’s withdrawal from China received less attention than it deserved. Bill Gates and his company have long been more pro-China than most of Big Tech. Microsoft’s China labs were crucial to China’s acquisition of AI expertise and experience. That is much of why China’s leader Xi Jinping mischievously greeted Gates as an “old friend” in Beijing in 2023, seven years into a bipartisan consensus that China was the pacing challenge for American security and the US economy. Microsoft’s withdrawal began late in 2024 and has continued through this year. The shuttering of its Shanghai AI lab early in 2025 was done very quietly but it reversed decades of company policy that had done much to create China’s AI industry in the first place. In short, when Microsoft decouples it really means something.

At the same time, China made a strategic shift this month with comprehensive restrictions placed on Chinese companies to prevent use of US silicon chips. This hit Nvidia particularly hard. Its share of the China market plummeted from 95% not long ago to 50%. Nvidia’s CEO, Jensen Huang, has done everything he can to hold on to what he still has. He is said to have the ear of President Trump. But the reprieve Huang secured in July seems to have been eliminated by China’s new moves. When China decouples at this scale, it really means something. 

Tech decoupling is a secular trend. It is the central force behind the current trade tensions, which both China and the US have been escalating, each placing the blame on the other. President Trump’s retaliatory tariffs, set to take effect November 1, responded to China’s weaponizing (not for the first time) of its tight grip on rare-earths production. All these moves revolve around the perceived centrality of AI to victory or defeat in the geoeconomic struggle between China and the US.

The two countries are nonetheless continuing talks. China hawks in Washington and elsewhere are genuinely worried that President Trump’s love of the grand gesture will combine with the influence of Jensen Huang and others to undermine the structure of tech containment built up in recent years. They might well look to Trump adviser Peter Navarro for reassurance. He has been ringing the bell about the China threat for 20 years. And indeed at the Council on Foreign Relations on Friday Navarro spoke of how the president’s tariff negotiations have already resulted in “19 trillion dollars” of promised investment: “foreigners are going to be paying to fix the vulnerabilities in our supply chain,” and once they have done so there will be a global “level playing field.” He also said that the US pre-Trump had “shipped 19 trillion dollars of our wealth” overseas. Nineteen trillion out, 19 trillion back in, and balance is restored. That is the idea. Navarro believed China’s new rare-earths policy is showing the world that China is everyone’s enemy: “The world will not go back to sleep on this.”

But if a three-year bull market, grounded in speculative bets on building data centers to set the infrastructural stage for future AI-driven productivity gains, wobbles enough, trade wars with China could lose their appeal. Tech containment and tech decoupling, though, will continue.

The Nine Lives of Economic Nationalism – Part Four of Four

Earlier posts in this series considered the multi-century trajectory of economic nationalism in reaction to empire, the resurgence of import substitution and major-power resource competitions, and the ways in which major-power economic nationalisms have made non-market-based economic development policies more popular than they have been in decades, almost regardless of levels of industrial development or economic size.

This final post considers some likely near futures of economic nationalism and economic sovereignty, with particular attention to AI.

First, the United States. The US was born in a determination to end external imperial dictation of economic policy and has, for the most part, guarded a relative autonomy from other economies ever since. The unification and then expansion of the 13 colonies across the continent integrated conquered territories into a “domestic” economy in a way that had few comparators elsewhere in the world. The resulting extent of US natural resources, from fresh water to arable land to natural gas, also proved to be unique. The US was peculiarly well suited to economic sovereignty, and with large-scale immigration it was able to grow on domestic demand better than anywhere else. Exports therefore accounted for a relatively smaller share of GDP than was the case in other industrial countries.

The constraining factor in the US case was not a lack of petroleum or fresh water or food but labor productivity. This was addressed through numerous means, from transport infrastructure to compulsory public education to industrialized agriculture. It helped that the US economy, unlike other industrialized economies, benefitted from both world wars. Productivity entered a crisis in the 1970s. It was eased, in a way, by the Internet and industrial globalization: your wage might be stagnant but it bought much more. But that improvement depended on production outside the US under working conditions that would be rejected in the US itself.

The extraordinary US investment in artificial intelligence comes from this.  AI holds out the promise of increasing productivity. But will it be global productivity or national productivity? Differently put, will the gains be captured by transnational capital and consumers or by tax-paying domestic markets and citizens? Will it be international or nationalist? Low unemployment, very slow job creation and high government and corporate debt all suggest that, absent an AI productivity miracle, the US will head into recession. That might well make the American people more nationalistic and insistent on economic sovereignty, but economic nationalism will not be able to solve their problems.

Chinese economic nationalism faces other constraints. A shrinking workforce and resistance to immigration mean productivity gains will have to come from labor-saving technology and investment in the non-Chinese global workforce. The first would be economically nationalistic. The second would be more like what US companies did in the 1980s and 1990s, and it could hollow out the Chinese jobs market as it once did the American. This would fuel the popular appeal of economic nationalism but, again, economic nationalism is not likely to be able to solve China’s labor productivity problems. An AI productivity miracle would help China as it would help the US. But it would be a miracle.

AI looks different outside the US and China. Those two countries thoroughly dominate the AI space. In AI terms, most other countries are takers, not makers. Africa’s population, a bit larger than China’s, captures 2.5% of the global AI market and is expected to attract 0.3% of global AI investment. The European Union attracts 7%. Britain, Canada, Israel and India also have significant investment, with Britain’s spend twice that of Canada’s. Nonetheless, the US and China attract 80%, with four fifths of it in the US. If an AI productivity miracle occurs in the existing economic-nationalist environment, it is difficult in political terms to imagine the benefits being rapidly diffused across the globe, since the goal of the investment is roughly the opposite.

AI aside, the resurgence of discredited 1960s-era development economics, from “national champions” and import substitution to infant-industry protection and tariffs, is becoming widespread. These policies were celebrated by the Left half a century ago as a way to withstand US corporate domination. Today their appeal is close to universal. They are even seen in the US as ways to ensure the US domination that they were once meant to block.

The essential point seems to be sovereignty. It is a phenomenon rich in paradox. The US-led Internet boom made possible a globalization that dramatically increased the wealth of once-poor countries, above all China but also India and others. These states could then afford to oppose what had just made them wealthy and to revive policies that had not helped them at all the first time around. China, India and other once-colonized nations wrap this in a rhetoric of anti-imperialism while hurrying to lock up poor-world resources before their once-imperial competitors do.

This is the central reason why China’s alternative global-governance schemes will go only so far: they are motivated by economic nationalism. Yet the same is true of US, Indian and European efforts, although European economic nationalism plays out on two levels at once, the national and the supranational. The major EU reform initiatives of 2024 were all premised on consolidating nation-based sectors into a super-nation capable of competing with the US and China.

For investors, at the national (or for the EU, supra-national) level, the play is in policy arbitrage, which is also political arbitrage. At the global level, as between major economic-nationalist actors like China, the US, India and the European Union, it makes sense to hedge with presences in at least two, navigating the relationship in each market among affirmative industrial and financial policy, protection, and market-based competitiveness. (A simpler way to do this, of course, is to invest in multinationals and funds with the proven capacity to do this kind of multi-market navigation themselves.) Beyond that, in countries like Nigeria and Ethiopia, which aim at economic sovereignty but lack much of what is necessary to achieve it, there are opportunities in the state-favored sectors themselves, the import and domestic sectors that provide the necessary inputs (such as electricity and raw materials), and the export sectors that ultimately make imports possible.

Little of this was featured in business school and Adam Smith would be appalled, but for the time being economic nationalism is the way of the world. 

The New AI Action Plan

The Trump administration’s AI action plan got a surprisingly warm welcome this week from US tech-industry and foreign-policy experts. The plan was unusual for this administration, and for the Republican Party, in that it advocates complex government-led initiatives, requiring considerable government funds, to advance political goals in a sector that is overwhelmingly made up of private companies. This is Trumpian industrial policy, and on paper at least it is even more interventionist than Biden-era industrial policies aimed at the tech sector. With its invocation of “renaissance” it is also more optimistic about technological innovation than any administration since Bill Clinton’s: “An industrial revolution, an information revolution, and a renaissance—all at once. This is the potential that AI presents.” In announcing the plan Trump also called AI “pure genius.” SIG’s view is that the AI action plan is both inspiring and well done but that implementing it will be extremely challenging.

Some of the challenges are obvious. The Trump administration has been cutting government bureaucracies, including in tech, yet this plan has numerous policy prescriptions that require government bureaucrats to implement them. The initiatives also require funding, which it is up to Congress to give. While there is general bipartisan support for AI investment, primarily as part of the strategic confrontation with China, the new AI action plan revived the White House’s effort to prevent states from legislating on AI. A similar provision in President Trump’s signature tax bill was defeated in Congress by a crushing majority. The AI action plan’s tactic is to say the federal government will withhold funds from any state that regulates AI in a way that would be “burdensome” or “unduly restrictive to innovation” — as judged by the White House on the advice of federal officials. Congress members represent state and local constituencies, not a national one. That is where their power comes from. Many of their constituents have very grave concerns about AI and expect their representatives to do something about it. When the AI section of the tax bill was rejected by Congress, Republicans, who have been much more for states’ rights (for example on abortion) than Democrats, were overwhelmingly against the president’s proposal.  In several senses, then, the AI action plan is primed for conflict with Congress.

The action plan is also primed for conflict with allies. The AI “dominance” foreshadowed by Vice President Vance in his speech earlier this year in Paris is transformed in the action plan to advocating export of the full American-made “AI stack” to allied countries. An American hardware-and-software suite, deliberately cleansed of any technology produced by “adversary countries” (China), would then become the infrastructure for whatever applications companies in other countries might be able to build. In other words, AI infrastructure would resemble the Internet of 2003: an American platform that others could participate in subject to US rules and US intelligence surveillance, and at a tremendous competitive disadvantage to US companies. This is exactly what other countries want to avoid, especially European countries who are still at the core of the US’s alliance structure. Just as the Trump administration wants US AI to be US-made and reflect US values, Europe wants its own AI sector to do the same — just as China insists on its AI companies reflecting “socialist values.” The action plan rightly stresses that for US AI to have maximum strategic benefits it must be on open rather than closed models and build on alliances rather than going alone. But in a geopolitical environment where allies are considering a tech-driven Buy European Act — and in which US tech giants are setting up “sovereign data clouds” just to keep European customers happy — it is hard to see how exporting the US AI stack in toto (once such a stack exists) will be welcomed abroad. China’s more subtle, and affordable, approach seems more likely to succeed.

The most serious challenge to the administration’s AI action plan is the challenge that faces any government regulation of digital technology: the systems are run by private companies according to market logic, more or less. Silicon Valley’s reaction to the AI action plan has been very positive. It is, after all, a strikingly pro-business and pro-technology plan. The plan’s urging of more government and private spending on the electric grid and data centers will certainly boost industry.

But what if capacity is overbuilt, or the wrong kind? Energy expenditures for AI so far have been fantastically high. If AI is to succeed it will need more energy and more data centers. Nonetheless, AI companies also want to reduce costs, which is why a great deal of investment is going into finding less energy-intensive ways to get AI results. (Data-center companies are also striving to find ways to lower their energy requirements.) The government could end up financing with taxpayer money an infrastructure that won’t be what is needed in five or ten years. Investors should be cautious of extrapolating investment opportunities from the areas that the AI action bill is targeting. The obstacles to the plan are many, and the record of government-led innovation policies is decidedly mixed.

The Jobs Conundrum

The US jobs numbers last week were chaotic, to say the least. The 0.1% drop in unemployment was yet another instance in which economists’ predictions were wrong. It is getting to be a habit, and the Donald Trump administration is reaping the political gains. The last few weeks have seen more and more articles attempting to explain why the predicted catastrophe after the Liberation Day tariffs announcement has not materialized. SIG’s view is that, now that the administration’s giant tax-and-spending bill has passed and members of Congress return to their constituencies for the summer recess, the real political work will concern jobs. So it is worth looking deeper into the new numbers.

Jobs in June increased by 147,000. However, the workforce itself shrank by more than that: The number of people characterized by the Bureau of Labor Statistics as “not in the labor force,” and therefore not counted as “unemployed,” grew by 490,000. The unemployment rate went down not just because jobs were added but also because the size of the workforce decreased. 

In sectoral terms, the biggest job adds (73,000) were in government. The biggest source of those jobs was growth in the public education sector, which is mainly K-12 schools. Of the 47,000 state-government jobs gained, 40,000 were in education. Of the 33,000 jobs added in local government, 23,000 were in education. Federal government employment was down by 7,000 for June and has dropped by 69,000 since the beginning of the Trump administration, in line with the president’s commitment to shrink government.

The increase in state and local education jobs should not be a surprise. The 2008 recession hit those sectors very hard. They recovered at a much slower rate than the private sector. When Covid hit, their subsequent recovery, compared to that of the private sector, was even worse. Massive federal aid got schools through the pandemic but it was always going to dry up and eventually did. States, looking to the longer term, realized they needed to increase spending. Populous states like Texas, California, and New York have recently broken records for education spending. Much of it goes into teacher salaries, which have been increasing in response to a chronic teacher shortage. (Credentialing in many states has also become much more lenient to attract more teachers.) In short, the state and local public education sector was overdue for a boost, got it, and jobs have been created.

The other major sectors driving job gains in June were “health care and social assistance” (58,600) and “leisure and hospitality” (20,000).  “Social assistance,” in the world of the Bureau of Labor Statistics, is not governmental but includes services like child care, vocational rehabilitation for the disabled, community food banks, and emergency services. The remaining major gains were in construction (15,000) and transportation/warehousing (7,500).

Overall, the private sector did not do as well as the public sector. Private payrolls were up by 74,000, the weakest growth since last October. An ADP Research study earlier in the week identified numerous indicators of weakening in the private labor market. Job losses in June were concentrated in mining and logging (down 2,000), wholesale trade (down 6,600), manufacturing (7,000), and professional and business services (7,000).

The problem, of course, is that the Trump administration’s goal has been to reduce government and favor the private sector, while the reality of the labor market so far is going in the opposite direction. Meanwhile, CEOs were spreading the word that AI would eliminate jobs on a grand scale. Ford’s Jim Farley thought that AI would “replace literally half of all white-collar workers in the U.S.” Of course, AI could also eliminate jobs in the public sector, including education. But the impetus for the current, very high levels of investment in AI is to increase productivity by making private-sector workers more efficient, not by hiring more of them. Overall, then, AI could well shift the balance of employment in the US economy further toward government.

It is possible that reducing taxes, as the new bill does, on upper-income groups could increase consumer demand, probably in the leisure category, and even free up capital for productive investment. It is also possible that a tariff program could result in increased investment in American manufacturing. However, neither of those results is going to be quick. In the meantime, Congress members will meet their constituencies as private-sector employment weakens and the federal government’s willingness or ability (given extraordinary debt levels) to solve problems, much less provide jobs, is weakening as well. Whether President Trump’s economics will work out in the end might not matter, because the end will be after the midterms, which in political terms could be too late.