The river: across, up, and back

The Chinese are optimistic about AI, no Doomers

Vivian Wang, Where are China’s A.I. Doomers? NYTimes, Mar. 4, 2026.

People in China are among the most excited in the world about A.I., according to a KPMG survey of 47 countries last year. While 69 percent of people in China said the technology’s benefits outweighed its risks, only 35 percent of Americans agreed. Other polls have shown similar disparities.

The question is, why?

The answer may be related to how the technology has been deployed in each country, as well as how the government and industry leaders have talked about it.

I don’t find this at all surprising. AI Doom is a projective fantasy, as I argued a couple of years ago in 3 Quarks Daily. Continuing on:

... Chinese tech companies have focused intensely on real-world applications for A.I. By contrast, many leading American tech companies have been focused on more abstract goals, like developing the most cutting-edge model, or achieving artificial general intelligence.

In addition, most of China’s leading A.I. models are free to use, unlike in the United States, where users have to pay for chatbots like ChatGPT to access all their features. (In fact, Chinese companies have been giving away money and luxury cars to entice people to download their apps.)

As a result, Chinese consumers are feeling the benefits of A.I., said Bai Guo, a professor who studies the digital economy at China Europe International Business School in Shanghai.

“A lot of things can already be helped by A.I., and people find that interesting, that’s useful, and so there are quite a lot of positive and active feelings toward it,” Professor Bai said. Potential dangers, such as unemployment or increased inequality, still feel remote.

The Chinese government has emphasized practical use: “Officials say that A.I. could help solve China’s thorniest problems, such as inequalities in health care, or an aging work force.” And so:

In August, the government laid out a plan, called A.I.+, for A.I. to penetrate more than 70 percent of Chinese society by 2027, and 90 percent by 2030. The plan said A.I. will “promote a revolutionary leap in productive ability” and “create higher-quality, beautiful lives.”

Because Chinese officials are promoting A.I. as an economic engine, they may also be silencing those who are more pessimistic about it. Crashes involving autonomous driving have attracted widespread attention online, only for posts to be censored. State media outlets have compared concerns about job loss for taxi drivers to the Luddite movement.

However:

Users have also raised concerns about how easily the government’s restrictions can be bypassed. A Chinese feminist group recently highlighted tutorials for making sexually explicit deepfakes that circulate openly on Chinese social media. Attempts to report the images were unsuccessful, the group said.

The Chinese government has also begun more directly addressing the technology’s potential for disrupting jobs, mental health or the Communist Party’s grip on power. [...]

For all of its potential, China must not let A.I. “spiral out of control,” Mr. Xi warned during a recent meeting of the leaders of the Communist Party.

There’s more at the link.

OJ and waffle

The Pentagon's position on Anthropic is legally hopeless

Michael Endrias and Alan Z. Rozenshtein have a substantial article about the Anthropic mess: Pentagon’s Anthropic Designation Won’t Survive First Contact with Legal System, Lawfare, 1,2,26.

From their introduction:

From the government's perspective, Claude does pose some concerning vendor reliability issues. But the specific actions Hegseth and Trump took have serious legal problems. The designation exceeds what the statute authorizes. The required findings don't hold up. And Hegseth's own public statements may have doomed the government's litigation posture before it even begins.

After considerable reasoning:  

Step back and consider what these positions amount to together. The government is arguing that Claude is so vital to military operations that it cannot tolerate any contractual restrictions on it—while simultaneously claiming that Claude poses such a grave supply chain risk that the entire federal government must stop using it, every defense contractor must sever commercial ties with its maker, and the company should be cut off from the cloud infrastructure it needs to survive. It’s like the joke from “Annie Hall”: The food is terrible and the portions are too small.

That might be funny as a bit of Borscht Belt humor. It is less amusing as a description of the United States government's strategy toward one of the companies leading America's effort to develop what may be the most important technology of the century. What Hegseth is actually describing is not a supply chain risk determination but something closer to the beginning of a partial nationalization of the AI industry: Seize the technology and, if you can’t, destroy the company to ensure that no future AI developer dares negotiate terms the Pentagon dislikes.

Arbitrary and capricious review requires, at minimum, logical coherence. The government cannot credibly maintain that a vendor is indispensable, that its continued integration poses no immediate danger, that its technology is reliable enough for active combat operations in Iran, and that it is nonetheless so dangerous it must be severed from the entire federal procurement ecosystem—all in the same week. Even a court inclined to defer on national security matters will notice that these propositions cannot all be true at once. [...]

The most obvious: if the Pentagon finds Anthropic's usage restrictions unacceptable, it can simply decline to renew the contract and move to a competitor. That is a routine procurement decision, available to any buyer who dislikes a vendor's terms. It requires no supply chain designation, no secondary boycott, and no government-wide ban. The fact that the government reached past this straightforward option for the most extreme tool in the procurement arsenal—one designed for foreign adversaries infiltrating the supply chain—is itself evidence that the designation is doing something other than managing supply chain risk. [...]

The legal problems are so glaring, in fact, that a cynical possibility suggests itself: The administration knows this won't survive judicial review and is doing it anyway, so that when they inevitably lose, they can still claim to have gone hard against Anthropic. This is designation as political theater: a show of force that was never meant to stick.

But there is another possibility. The administration may genuinely believe that a Truth Social post and a procurement statute designed for state-influenced Russian and Chinese tech companies can destroy an American AI lab over a contract dispute. If so, they are in for a rude awakening. The statute wasn't built for this, the facts don't support it, and the courts will say so.

H/t Tyler Cowen.

The Shipyard

David Gallagher, The Exact Shade of Gray; THE SHIPYARD. By Juan Carlos Onetti. The New York Times, June 16, 1968.

An unexpected consequence of Doomer propaganda: The Pentagon wants to control the Doomsday Device

Casey Mock, Pete Hegseth Got His Happy Meal, Tomorrow's Mess, March 2, 2026.

Concerning the current dust-up between the Pentagon and Anthropic:

Something like this was always going to happen. Not because of Hegseth specifically, not because of this administration, but because of the narrative the AI safety community — the world that produced Anthropic, and whose language Anthropic still speaks even while disavowing its label — has been pushing for at least the last three years.

 [Oh, much longer than that, much longer. – BB]

Imagine a six-year-old whose entire media diet includes a steady stream of McDonald’s commercials, a Happy Meal ad at every break, focused on whatever toy is the latest to be included along with the McNuggets. Now put that child in a car that drives past a McDonald’s. What happens?

The Rationalist and Effective Altruist communities — the intellectual cultures that gave us Anthropic, influence many of their employees, and which still shape how Dario Amodei talks about his company and his technology — have spent the better part of a decade insisting, with increasing urgency, that artificial intelligence is the most consequential technology in human history. Maybe it’s civilization-ending; maybe it’s civilization-saving. Either way, it’s the hinge on which everything henceforth turns.

With policymakers and the media largely having accepted the premise, thus surrendered was the argument for treating AI like a normal technology subject to normal governance. Policies being pushed by Effective Altruist groups, like 2024’s SB1047 in California — deprioritize harms happening today for theoretical existential ones in the future; despite the fact that today’s harms that could be existential for the folks experiencing them. These groups incessantly made the case that whoever controls this technology controls the future, and so the hypothetical future needs to be prioritized now. In a Washington now run by people who tend to impulsiveness and contemptuousness of institutional constraint — well, it’s easy to see where this was headed. Hegseth saw the ads for the toy, and so now he wanted his Happy Meal. [...]

Yet the prognostications of the doomer community have been, nearly without exception, wrong — not in small ways, but in the foundational sense that the imagined trajectory keeps failing to materialize. [...]

Thus, this news reveals the rationalists’ under-examined blind spot: they cannot model the messy Pete Hegseths of the world, even as their claims whet Hegseth’s appetite. The rationalist view of the world assumes, at some level, that the relevant actors are optimizing for well-understood, predictable variables and a clear understanding of what best serves their self-interest. What it cannot account for is bad faith, impulsiveness, ideological motivation untethered from evidence, random instances of force majeure, and personal whims and petty rivalries. And so while the doomer community spent years warning about uncontrollable AI systems that do things their creators didn’t intend, they apparently did not consider what would happen when the humans currently running the United States government got access to technology they’d been told was the hinge of history.

H/t Gary Marcus.

I've published an article about Doomers in 3 Quarks Daily: On the Cult of AI Doom, September 12, 2026.

Tree with pink blossoms [Japan]

Reiji Hiramatsu
Such a beautiful piece of work
Japanese artist inspired by Monet pic.twitter.com/s6zQW9nb4l

— helen warlow (@HWarlow) March 3, 2026

Fun in the snow

Pulling things together about creativity & LLMs

Consider these recent posts about AI:

Two Ways to Use AI: Homo Economicus vs. Homo Ludens (2.27.26) 

Chatbots have increased my sense of intellectual agency such that being an intellectual “outsider” becomes a superpower. (2.17.26) 

Three mathematicians are not impressed with the ability of AI to do professional math (2.16.26)

What do they have to do with one another? They are about creativity, interdisciplinary work and, ultimately, about the division of labor between humans and AI, currently represented by LLM-based chatbots. The NYTimes article about mathematicians makes the point that the creative work in mathematics involves creating frameworks in which to present and solve problems. AIs cannot do that currently. They’re better suited to working on well-defined problems.

That’s certainly consistent with observations I’ve been making for a while. All those benchmarks involve well-defined problems. The real problem, in contrast, is to frame such a problem in the first place. I’ve got a working paper where I present three case studies from my own work, cases where I started out with nothing in particular in mind and ended up doing a bit of focused research: Serendipity in the Wild: Three Cases, With remarks on what computers can't do.

The post about my “superpower” is based on the fact that the LLMs are trained on materials published to the web and so “gravitates” toward those ideas. But what if you take up a stance outside the current intellectual ecosystem, but nonetheless are conversant with it and can ground your work in it, at least partially? That’s been my situation since my early work on “Kubla Khan.” In 1978 I filed a doctoral dissertation entitled, “Cognitive Science and Literary Theory.” As far as I know, I was pretty much alone in working that territory. Here’s how I characterized cognitive science (somewhat idiosyncratically):

The basic problem of cognitive science is establishing a five way correspondence between the following:

1. Brain Geometry: Neuroanatomy is the study of the geometry of the brain. Comparative neuroanatomy is the study of the correspondence between brains of different species.
2. Computation: Different types of computers can perform different classes of computations and the nature of the computations depends on the geometry of the computer.
3. Behavior: Much of psychology is the study of the behavior of organisms. The behavior an organism exhibits is determined by the class of computations which its brain can perform which in turn depends on the geometry of the brain.
4. Phylogeny: Animals at different evolutionary grades have different brain geometries. The brain geometry must be capable of performing the class of computations necessary for survival in the animal’s particular ecological niche. But what is the relationship between moving to a new niche and the emergence of a new brain geometry?
5. Ontogeny: As a child matures different brain structures develop and permit new classes of computation sustaining new types of behavior. But how does phylogeny exploit differential maturation rates to create a new class of computer?

Those are five distinct intellectual domains. My dissertation was strongest on behavior, literary texts, and computation, cognitive network semantics, but I touched on neuroanatomy and phylogeny (there’s a section on mirror recognition in humans and apes). Offhand I don’t recall anything about childhood development.

A decade later, however, David Hays and I published “The Principles and Development of Natural Intelligence” (1988). The principles themselves were computational. We identified neural structures associated with each, gave examples of behaviors enabled by each principle, and placed them in both phylogenetic and ontogenetic contexts. That’s a large part of the framework I’ve been working with my entire career – there’s also culture and cultural evolution. Am I an expert across that entire domain? Of course not. But I do have a high degree of expertise in some areas, particularly literary and textual analysis and semantic structures, and I’ve read in the technical literatures across that whole range. I’ve got a “feel” for the material, enough so that I can prompt chatbots (Claude and ChatGPT) across the whole range and follow it when it fills in details that I don’t myself command.

That most recent post (2.27.26) contrasts the way I use chatbots (Homo Ludens) versus more conventional usage (Homo Economicus). You might also look at the still more recent post, Why Gemini 3.1 is so good [long chains of reasoning, across disciplinary boundaries]. That title tells half the story. Somewhere in the video Jones makes the point that, while Gemini 3.1 Pro can construct long trains of reasoning that humans cannot, often crossing disciplinary boundaries, humans can very those chains. That’s my situation. I can check the reasoning for ontological consistency. That’s not the same as truth, but it’s a pre-requisite for it.

Ellie Pavlick, (How) Does AI Think?

c. 40:16 “At various points I’ve like argued really what we’re seeing here is a neural implementation of what is latently a symbolic system like our symbolic AI systems of yore.”

A bit later Pavlick will back off from that statement. However, her first example is arithmetic. Concerning arithmetic note that it is NOT “native” the human mind. Preliterate cultures may not even have open-ended counting systems, it any, and don’t do numerical calculations. Moreover, while children pick up language readily without specific instruction, arithmetic requires focus instruction and fluency requires hours of drill over several years. Careful reasoning is like that as well. Much of formal education is about learning how to reason in various domains.

Keep in mind, “symbolic AI systems” covers a LOT of ground. The expert systems, built on production rules, are perhaps the most visible type of symbolic system. But I think that cognitive nets are a better bet for the latent structure of neural nets. That’s what I argued in ChatGPT: Exploring the Digital Wilderness, Findings and Prospects.

Trump Ban Sends Claude to #1

AI Secret, Mar. 2, 2026.

👀 What’s happening: After Anthropic’s talks with the US Department of Defense collapsed over military AI limits, the White House moved to ban Claude from federal use and labeled it a supply chain threat. Within 24 hours of being publicly targeted, Claude shot from outside the top 100 to number one on the US and Canada App Store free charts, overtaking ChatGPT and Gemini.

🌍 How this hits reality: A federal ban was supposed to isolate a vendor. Instead, it converted policy punishment into consumer demand. SensorTower data shows a direct ranking spike tied to the announcement. Social feeds filled with subscription cancellations and data export tutorials. Billions in defense linked compute shifted toward OpenAI, but retail distribution shifted the other way. Politics instantly rewired both infrastructure allocation and user flows.

🛎️ Key takeaway: State pressure can redirect contracts overnight, but it can also manufacture market momentum. In AI, regulatory confrontation now doubles as distribution strategy, whether intentional or not.

Something interesting to look at on Sunday morning

Words, code, guardrails & weasels: OpenAI, Anthropic, and the Pentagon

I work in government affairs at OpenAI.

My job is federal partnerships. When an agency wants our models, I make sure the paperwork is beautiful. Paperwork is my love language. On my desk I have a framed quote that says "Policy Is Just Code That Runs on People." I bought the…

— Peter Girnus 🦅 (@gothburz) February 28, 2026

I've copied the entire “tweet” below in case you don't want to click. But you might want to glance through the thread. This is the “tweet” where Gimus says his badge stopped working.

* * * * *

I work in government affairs at OpenAI.

My job is federal partnerships. When an agency wants our models, I make sure the paperwork is beautiful. Paperwork is my love language. On my desk I have a framed quote that says "Policy Is Just Code That Runs on People." I bought the frame at Target. It was in the Live Laugh Love section. I did not see the irony at the time. I still don't.

We had a good week.

On Monday, we closed a $110 billion funding round. One hundred and ten billion dollars. Amazon put in fifty. Nvidia put in thirty. Valuation: $730 billion. The largest private fundraise in the history of anyone raising anything. There was a company-wide Slack message about it. The message used the word "transformative" twice and the word "safety" once. The word "safety" was in the last sentence, after the link to the new branded hoodie pre-order. The hoodies are nice. They're the soft kind.

On Tuesday, we fired a research scientist for insider trading on Polymarket.

Why Gemini 3.1 is so good [long chains of reasoning, across disciplinary boundaries]

YouTube:

What's really happening when Google ships the smartest AI model on the planet, prices it at a seventh of the competition, and doesn't care if you keep using Claude or ChatGPT? The common story is that this is another benchmark race—but the reality is more interesting when the company generating $100 billion in annual free cash flow is playing a fundamentally different game. In this video, I share the inside scoop on why Gemini 3.1 Pro reveals more about problem types than model rankings:

• Why Google's vertical stack from TPU silicon to Nobel Prize research is an impregnable fortress
• How Deep Think solved 18 previously unsolved problems across math, physics, and economics
• What separates reasoning problems from effort, coordination, ambiguity, and emotional intelligence problems
• Where the question "which AI should I use" becomes the wrong question entirely

For knowledge workers watching the model landscape differentiate, the margin between routing models well and using one model for everything is widening every single month.

Chapters

00:00 Google Shipped the Smartest Model and Doesn't Care If You Use It
03:15 Arc AGI 2: The Largest Single-Generation Reasoning Gain Ever
05:30 What Google Optimized For vs Anthropic and OpenAI
07:10 Demis Hassabis: Solve Intelligence, Then Solve Everything Else
09:45 Google's Vertical Stack: From Transistor Design to Protein Folding
13:20 Why Google Can Afford to Lose the Model Race 15:00 What Gemini 3.1 Pro Is and Isn't
17:30 Naked Reasoner vs Equipped Reasoner vs Specialist Coder
19:45 Deep Think: Disproving Conjectures and Catching Peer Review Errors
23:10 Hard Is Not One Thing: Six Types of Difficult Problems
28:40 Which Problems Does Pure Reasoning Actually Help?
32:15 What This Means for Your Work Tomorrow
35:50 Google's Quiet Game: Building the Thing Underneath the Thing

These two short passages give you a flavor:

14:03: “The model crossed disciplinary boundaries that human specialists very rarely cross because the model doesn't see disciplinary boundaries and that is one of the strengths of an AI model.”

15:10: “Gemini is good for certain kinds of problems: “And they share specific characteristics. The inputs are well-defined like a protein sequence. The problem can be stated extremely precisely. And the solution requires a long and sustained chain of logical deduction that a human mind can verify but often cannot generate without years of specialized training.”

If you don’t want to watch the whole thing, start with “What Gemini 3.1 Pro Is and Isn’t,” @ 15:00.

Carving at the joints: Plato, Zhuangzi, Guo Xiang

First, a prompt I gave Claude 5.4. Then Claude’s reply.

* * * * *

There’s a cliché about carving Nature at its joints.

There’s one version from Plato’s Phaedrus. Socrates has likened a well-formed speech to an animal with its various appropriately arranged parts and is now examining two different speeches on love (265e-266a):

... we are not to attempt to hack off parts like a clumsy butcher, but to take example from our two recent speeches. The single general form which they postulated was irrationality; next on the analogy of a single natural body with its pairs of like-named members, right arm or leg, as we say, and left, they conceived of madness as a single objective form existing in human beings. Wherefore the first speech divided off a part on the left, and

continued to make divisions ...

There’s a different version from Guo Xiang’s redaction of Zhuangzi (from Wikipedia):

Cook Ting was cutting up an ox for Lord Wen-hui. At every touch of his hand, every heave of his shoulder, every move of his feet, every thrust of his knee, zip! zoop! He slithered the knife along with a zing, and all was in perfect rhythm, as though he were performing the dance of the Mulberry Grove or keeping time to the Ching-shou Music.

“Ah, this is marvelous!” said Lord Wen-hui. “Imagine skill reaching such heights!”

Cook Ting laid down his knife and replied, “What I care about is the Way, which goes beyond skill. When I first began cutting up oxen, all I could see was the ox itself. After three years I no longer saw the whole ox. And now I go at it by spirit and don't look with my eyes. Perception and understanding have come to a stop and spirit moves where it wants. I go along with the natural makeup, strike in the big hollows, guide the knife through the big openings, and follow things as they are. So I never touch the smallest ligament or tendon, much less a main joint.”

Plato dates: 428-348 BC. Guo Xiang dates: 252-312 AD.

Is there any chance that this idea made its way from Ancient Greece to Ancient China? Or is it likely that we have a case of independent invention? After all, butchering an animal is a common enough task and the difference between an unskilled butch and a highly skilled one must have been apparent to many people.

The question of transmission versus independent invention is genuinely interesting here, and the honest answer is that scholars are quite uncertain, though the weight of evidence favors independent invention.

The chronological gap is suggestive but not decisive. The Zhuangzi text itself predates Guo Xiang — Zhuangzi lived roughly 369-286 BC, making him a near-contemporary of Plato, not a distant inheritor. Guo Xiang's role was that of editor and commentator, so the Cook Ding story almost certainly originates in the 4th-3rd century BC, not the 3rd-4th century AD. This matters a great deal, because it means we're not looking at a case where ideas could have filtered eastward via the Silk Road over centuries — both texts are roughly contemporaneous.

There was contact between the Greek and Chinese worlds in this era, but it was extremely attenuated. Alexander's campaigns reached Bactria and the Punjab, and there are documented cases of idea-transmission in specific domains (certain astronomical and mathematical concepts, some artistic motifs). But the philosophical literature shows no clear fingerprints of cross-cultural borrowing in this period. The intermediary links needed to carry a specific Platonic metaphor from Athens to the Warring States courts of China in the 4th-3rd century BC are simply not there.

The more compelling case for independent invention rests on a few observations. First, as you note, skilled butchery was universally observable. Any culture that slaughtered animals for food — which is to say, every culture — could notice that an expert carver seems to find the natural seams while a novice hacks. The perceptual phenomenon is available everywhere. Second, and more importantly, the two passages are actually doing quite different philosophical work with the same raw material. Plato's Socrates is making an epistemological and rhetorical point about how to carve up concepts — how a good speech should divide a subject at its natural joints rather than arbitrarily. The emphasis is on correct classification, on getting the logical structure of reality right. Cook Ding's lesson is about something closer to wu wei — effortless action, yielding to the natural order, the dissolution of the self-conscious subject into the flow of the Way. The knife goes where the ox already opens; the cook stops trying. These are recognizably different philosophical preoccupations, and each fits neatly into its own tradition without requiring outside influence.

So what you have is a case where a shared human experience — watching an expert butcher work — gets recruited into philosophical argument independently in two places, but gets recruited to illuminate quite different problems. That's actually philosophically interesting in its own right: it suggests the analogy has a kind of natural aptness that makes it available to very different thinkers facing very different questions. The butcher doesn't just carve the ox at its joints; he also, apparently, carves the history of philosophy at a joint.