China on Eigenform Articles

Learning the Language of Rain

Tue, 17 Oct 2023 00:00:00 +0800

TL;DR

Compares Huawei’s Pangu (200 billion parameters, reportedly barely used in practice) against Google’s much smaller MetNet/MetNet-2 (225 million parameters) - both reportedly beat physics-based weather models like ECMWF’s IFS, but by different margins and via very different approaches.
Frames forecast quality as a tradeoff between “skill” (how close the average prediction is to the true outcome) and “spread” (variance across the model’s own predictions), arguing that adding variables tends to raise both accuracy and entropy at once - accuracy compounds arithmetically while the entropy from variable interactions compounds geometrically.
Argues Pangu likely models more variable interactions per iteration than physics-based IFS, which could explain why its predictions may centre closer to the true outcome while being harder to pin down precisely - with some suggestive, though inconclusive, evidence from comparing skill-spread correlations.
The central puzzle: rain has historically been a poor predictor of future rain, yet MetNet performs strongly using raw rain-cloud images as its main input; the piece argues this only makes sense if the images are non-human-readably encoding other predictive variables (wind, humidity, and so on), cutting the model’s formal variable count - and the entropy that comes from variables interacting - while preserving the predictive information.
This is framed through an information-theory/compression argument: a system with richer pre-loaded context (the way a chemistry professor unpacks more from the letter “H” than a layperson) can compress more information into fewer formal units, which the piece offers as the underlying reason a smaller, well-designed model can outperform a larger, more literal one.
Closes with a speculative contrast in motive: it reads Pangu’s minimal design effort as a reflection of limited genuine corporate interest beyond political optics, versus MetNet as a more deliberate, first-principles rebuild of what a weather model can be - while explicitly stating this isn’t intended as a criticism of Pangu’s underlying technical capability, and that the two systems aren’t being benchmarked like-for-like.

Key Takeaways

Artificial Intelligences in the Guanzi and the Han Feizi

Mon, 20 Mar 2023 00:00:00 +0800

TL;DR

Confucians learn by accumulating proven methods; Daoists learn by eliminating disproven ones through trial and error, which the piece treats as a form of gradient descent that needs no comprehension, only a lower cost of failure.
Both traditions imply that learning does not require a brain: a dog, a river finding its route to the sea, or a market setting a price all “learn” in the same mechanistic sense.
Eigenform’s 2021 agent-based model of the Guanzi’s redistribution system shows a simple wealth-and-generosity following rule producing progressively more stable political hierarchies over time, without any agent understanding what stability is.
That Guanzi-style system only learns one task in one sandbox; the Han Feizi describes a more general design built on two parts - an ungamable universal reward (territory) for solving any problem, and a “central decider” who need not understand a solution, only whether it succeeded.
The state of Qin’s real-world reward system for military conquest is offered as historical evidence that this filtering-and-reward loop actually produced compounding, self-reinforcing gains in state capacity.
The essay proposes porting this design to software directly: an AI that treats unclaimed disk space as territory, writes and tests code to annex it, and retrains itself on whichever scripts succeed - a route to general rather than narrowly task-bound self-improvement.

Key Takeaways

Procedural Models of Political Order

Fri, 01 Oct 2021 00:00:00 +0800

TL;DR

Opens from a modelling principle: a good model isolates the necessary and sufficient inputs for an outcome, and parsimony matters because individual-level noise - like irrational decisions in the homo oeconomicus model - often has no effect on macro-level, emergent outcomes.
Recaps the authors’ earlier two-model theory of political order: an Anglo-European model built on threat and protection, where followers hedge by backing the second-strongest actor, producing frequent leadership turnover and a “monopoly of legitimate violence”; against a Chinese model built on benevolence and redistribution, where followers back whoever offers the greatest livelihood benefit, producing a “monopoly of legitimate benevolence.”
To test these theories against charges of oversimplification, the authors built two agent-based simulations in which agents follow either the highest or second-highest surplus-times-generosity score, and found the simulations reproduced not just the predicted structures but unprogrammed emergent behaviours - increasingly long periods of stability over time, and faster network-rebuilding under economic hardship, matching predictions from the original classical texts.
The Chinese-model simulation converges on a single flat pyramid where a leader must maximise inequality with followers while minimising inequality among followers, since that configuration best resists internal challengers - a result the piece stresses emerges from selection among surviving configurations, since none of the agents are capable of any such reasoning themselves.
The Anglo-European-model simulation unexpectedly produces not genuinely separate competing parties but a single dominant bloc split into two or three internally similar factions, which the piece reads as a structural explanation for why establishment parties often converge and jointly resist independent outsiders, even when voters might benefit from real competition.
Both models are used to argue that political structure is defined by follower selection rules rather than leader intent, extended to explain why transplanted foreign political systems tend to either collapse or become disguised copies of the pre-existing order, with the Enlightenment and the Warring States shift toward meritocracy singled out as rare successful reforms because they redirected existing tendencies rather than rewriting agents’ preferences outright, and benefited both agents and system at once.

Key Takeaways

“Someone spoke to the King in Zheng”: using high tech methods to solve an ancient Chinese mystery

Mon, 22 Feb 2021 00:00:00 +0800

TL;DR

Opens from a genuine historical puzzle in the Han Feizi: Han Fei delivers a speech to King Zheng of Qin that appears to betray his own state, then argues the opposite position soon after, is denounced by rival minister Li Si, and dies via a diplomatic mission to Han that seems designed to fail - leaving open why he acted as he did and why Han refused to receive Li Si’s embassy.
Proposes to resolve part of this by testing, with computational text analysis, whether an anonymous chapter called “Speaking to the King in Zheng” in the separately compiled Stratagems of the Warring States is actually a missing Han Fei speech that supplies the missing backstory.
Builds a qualitative case first: the chapter shares Han Fei’s stylistic hallmarks, but more tellingly shares his documented lack of self-preservation instinct - insulting people capable of having him killed - which the piece argues is far harder for a later imitator to fake than surface stylistic tics.
Runs cosine-similarity comparisons, with and without punctuation, showing the chapter sits closer to the Han Feizi corpus than to any section of the Stratagems, including other Han-related chapters, and that including it in the Han Feizi group barely dents that group’s internal consistency while including it in the Stratagems group dents that group’s consistency far more.
Repeats the comparison against several unrelated classical texts and, separately, controls for shared topic vocabulary by clustering on the full text and then again on only the words common to every text - the chapter stays closer to the Han Feizi corpus under every version of the test.
Concludes that if the chapter is a displaced Han Fei speech, it supplies the historical background explaining his apparently contradictory behaviour: not betrayal, but a calculated, ultimately failed attempt to position himself as a bridge between Han and Qin, undone by Li Si’s personal rivalry and deliberately aggressive diplomacy toward Han.

Key Takeaways