Tool or Creation: How Li Lu Frames the AI Question
![muted documentary photography, diplomatic setting, formal atmosphere, institutional gravitas, desaturated color palette, press photography style, 35mm film grain, natural lighting, professional photojournalism, A glass-encased accord on a dark oak plinth, its parchment yellowed and taut from age, bearing two signatures in contrasting ink—one precise and machine-rendered in metallic silver, the other handwritten in fading black, the official seal cracked down the center, embossed with interlocking human and algorithmic glyphs, side-lit by narrow shafts of cold morning light through tall institutional windows, dust suspended in the air, silence thick as velvet, the atmosphere of a chamber after a signing no one fully understood [fal-ai/z-image/turbo] muted documentary photography, diplomatic setting, formal atmosphere, institutional gravitas, desaturated color palette, press photography style, 35mm film grain, natural lighting, professional photojournalism, A glass-encased accord on a dark oak plinth, its parchment yellowed and taut from age, bearing two signatures in contrasting ink—one precise and machine-rendered in metallic silver, the other handwritten in fading black, the official seal cracked down the center, embossed with interlocking human and algorithmic glyphs, side-lit by narrow shafts of cold morning light through tall institutional windows, dust suspended in the air, silence thick as velvet, the atmosphere of a chamber after a signing no one fully understood [fal-ai/z-image/turbo]](https://cdn.digitalrain.dev/thelongview/viral-images/5c1c6701-e435-47c2-8dbb-e9931e54bed9_viral_0_square.jpg)
Generative AI demonstrates synthetic capability at the knowledge layer, but its economic footprint remains confined to recombination—not transformation. The valuation surge reflects belief in recursive self-improvement, not evidence of its deployment.
Li Lu, founder of Himalaya Capital, argues in a Hong Kong Economic Journal column that AI is moving from tool towards creation. His working definition is deliberately modest: a machine that combines very large computing power with clever algorithms, is fed nearly all human-generated data, and extracts first-order correlations among data variables that carry contextual relationships. Out of those correlations emerge abilities that look close to human reasoning. The approach works without full explanation, he says, because people know how to use the system and how to improve it but have not fully understood why it is effective.
His account of where it fits starts from the economy as a stack of layers. Energy and transport sit at the base, manufacturing above them, then communications and digital media, and above those a large layer of knowledge and technical services that combine centuries of accumulated expertise for everything beneath. Generative AI is strongest at recombining that top layer and making it more widely available. It does not change energy and transport themselves, and manufacturing remains the recombination of matter, atoms and molecules. On this reading its present reach is information synthesis, and it has not surpassed humans.
The next step in the Silicon Valley narrative is recursive self-improvement, in which AI makes all of its own improvements without human participation. Li Lu says this has partly been entered already. How much of the improvement loop now runs without people is not established in his account, so the question of how far it extends remains open. His red line concerns human participation: whether in AI self-learning or in tool use, people must stay involved, because a system detached from human definition could set its own goals and move further out of human control.
He expects society to absorb the change as it has absorbed earlier ones, through cycles of boom and bust, probably more than once, rather than through perfect planning. The pricing of the sector is where he sees the strain. He describes a company, calling itself a research lab, that spends tens of billions of dollars a year and promises hundreds of billions more, says its technology has roughly a 10% chance of killing everyone, and invites investors to commit at least $100bn at a $2 trillion valuation. If there is any logical consistency in that, he says, it is hard to find, and it illustrates how irrational markets can become.
His benchmark for sanity is the S&P 500. From 2010 it returned roughly 14% to 15% a year for about sixteen years, while the decade from 2000 to 2010 returned almost exactly zero and was slightly negative. Across the full twenty-six years the annualised return falls back to the high single digits. Real GDP growth of about 2% plus 2% to 3% inflation gives nominal growth of 4% to 5%, and listed companies usually trail the economy, though the index is selected to hold 500 of the strongest firms, which explains why a high single-digit long-run return is the normal level. Against that baseline, the question is what long-run return is reasonable and where markets eventually return to it.
† Analysis: the writer's conclusion from the reported facts, not a fact stated in the source.
- The author infers that the current valuation of an AI company promising a 10% chance of 'killing everyone' while seeking $100bn funding is an illustration of market irrationality, which directly aligns with the source's framing of that specific example. The conclusion that long-run returns for companies will eventually normalize to the high single digits is an inference drawn from the source's provided economic baseline (GDP + inflation vs. historical index performance).
Published November 11, 2026