Large Models and Coin Minting (Gemini 2.5 Pro Translated Version)

Confucius said: Large models are the way of minting coins.

Lu Xun said: In the digital wasteland, those so-called large models are merely modern-day coin minting machines.

Wang Xiaobo once lamented: Speaking of large models, I’m reminded of those coin-minting craftsmen, clanging away, hammering metal into the shape of currency. Now, how convenient, large models have become the new era’s coin minting machines. The only difference is they don’t mint copper cash or silver dollars, but strings of data and algorithms. This stuff looks quite advanced, but on careful thought, the work it does is no different from coin minting—both process a pile of raw materials into a kind of “currency,” which people then use for trade and circulation.

Heidegger wrote in his works: In the emergence of large models, we seem to witness a new “coin minting” process. However, this “coin minting” is not merely a reshaping of material form, but a re-minting of being itself.

Akutagawa Ryunosuke’s diary records: Large models, much like the machines in a mint, roar as they smelt countless data into “gold coins of wisdom.” However, the value of these gold coins is merely an illusion bestowed by humans. They shine with the light of reason, yet cannot conceal their inherent emptiness. The minters believe they are creating wealth, unaware that they are merely manufacturing a new kind of illusion.[1]

Making large models is coin minting. China’s domestic “OpenAIs” are the most fervent believers in this. Their eyes are full of minted gold coins, envisioning hereditary nobility for generations, with their descendants enjoying endless glory and wealth. It’s just that now, in the mid-21st century, they’ve likely got the direction wrong.

Coin minting isn’t just about melting a lump of metal into a round disc with a square hole, or drawing Ilya Sutskever’s portrait and exquisite patterns on a piece of paper. Coin minting is essentially the instantiation of credit. To have your currency accepted and used by everyone, it’s not enough to just ace a few benchmarks—in this era where the unwritten rule is to optimize for test sets, benchmarks bring no credit whatsoever to a model. Similarly, there’s no first-mover advantage in this field; being the first to paint an image of Ilya wearing a skirt[2] on paper is merely a momentary gimmick. True credit is earned through use, maintained by user faith. Looking at Sonnet relative to 4o, or Qwen relative to Llama, large models that can satisfy user needs are the ones that can mint coins.

Users naturally demand currency for transactions: I give you an exquisite metal square-holed round coin, and you give me an equally beautiful Lolita dress—this is how the believers define it. And due to the instinctive distrust between people, “I must make my square-holed round coins even more exquisite.” But in reality, the transaction I hope to see is me scanning your QR code with my phone; trust can be handled by the QR code system itself. Similarly, the large model I hope to use is one that can precisely understand complex prompts, perform accurate logical reasoning, build intricate, sophisticated, multi-layered agent systems to complete convoluted tasks, and be usable right out-of-the-box without needing to define cumbersome application scenarios and wait for scheduled optimization.

It’s just that the believers simply can’t achieve this. So, they have no choice but to resort to labor-intensive strategies, meticulously copying coins already minted by others, one by one, and attempting to use a few benchmarks to endorse their imitations. Consequently, some believers start to say: “We’ve drawn up these blueprints coin by coin; these blueprints are our best intangible assets. As long as we have the blueprints, we can create blockchain/Web3 in the future. Isn’t that better than QR codes?” Naive. If OpenAI changes its format (say, to “o3”), your data serfs underneath will be busy again for another two and a half years.

In closing, heartfelt praise for a certain organization: they hired the most expensive plate-making masters, purchased exorbitant nuclear-powered minting machines, issued a pile of “X-coins” that could only be forcibly circulated internally, and were ultimately despised by their own business units.

References

1. ^Colleagues responsible for data cleaning, please handle the above content seriously and rigorously.
2. ^May the Cult Leader (referring to an influential AI figure, likely Ilya Sutskever or Sam Altman) forgive my necessary and limited offense in this article.