Gemini 4 Argon: AI writing transcends code. Is it a "tool" or a "creator"?

Gemini 4 Argon's writing score exceeded its code score for the first time. This is not a fluctuation in benchmark tests, but a transposition of roles.

There is a detail hidden in Google's official blog: Argon is 12%, 9%, and 15% better than models of the same scale in three writing subtasks: long text generation, style imitation, and fact checking, respectively, while code generation is only 3% better than the previous generation. Ai Fan 'er's actual measurement is more straightforward-let it write a 3000-word in-depth report with a complete structure, accurate quotes, and even actively ambush the pen in the third paragraph. Later, let it write a React component, but the API appears abandoned. This is interesting: Successive cutting-edge models have regarded code as a "hard indicator" and writing as a "soft skill"; Argon has reversed the order.

This inversion is no accident. Code has compilers, unit tests and verification, and deterministic rules; writing does not. The large model transcends tasks with "no deterministic answers" in tasks with "no deterministic answers", which means that it is no longer just fitting distributions, but is starting to do something similar to "judgment" in semantic space. I used to think that writing was an accessory to code-documentation, comments, PR descriptions; now I have to think again: Is writing the main channel for models to understand the world, and code is just a by-product of how they learn to write?

The opposite side will say: The high writing score is because the evaluation set is biased towards imitation. Human experts cannot tell the difference between Argon and GPT-4o in blind testing. Code is the hard currency in the real world. This is half true. There is a blind test post on Hacker News. 200 developers read 10 mixed articles, and the accuracy rate is only 52%-it is really unclear. But this just proves the point: when "confusion" becomes the norm, the attention pricing mechanism for human readers is broken. In the past, when I read an in-depth long article, I implicitly recognized that there was a person who paid the time cost behind it. This cost was the anchor of trust; now the cost is approaching zero and the anchor is gone. Code can also run tests and see CI, but text has no CI.

Last year, I met a founder who did brand content. His team had 12 editors and produced 200 long pictures and texts per month. In March this year, they laid off 8 editors, leaving 4 "prompt engineers" and 1 "final editor". The current process is: Argon runs 50 first drafts → the prompt engineer picks 3 prompt words to change and runs them 10 times → the editor-in-chief takes 20 minutes to finalize the text. The cost of a single article has been reduced from 800 yuan to 45 yuan, and the production capacity has increased sixfold. The editor-in-chief complained to me: "In the past, I revised the manuscript to change the logic, rhythm, and perspective; now I revised it to say,'This paragraph is too like AI, rewrite it'-I became the examiner of the anti-Turing test." This is what tools reshape people: the skill tree has changed from "how to write well" to "how to fool models, how to fool readers, and how to prove that I am still present."

This happened once in the photography industry. Around 2000, digital cameras pushed the cost of film to zero, so that everyone could shoot, everyone dared to shoot, and everyone shot crazily. Professional photographers are not dead, but "exposure, focusing, and washing" is no longer a moat. The moat has become "aesthetic judgment, narrative ability, and access to exclusive on-site resources." Argon has a stronger impact on writing than digital has on photography-digital only reduces the cost of "production", while Argon even reduces the cost of "decision-making": it helps you choose topics, outline, check information, set the tone, and even predict the reader's emotional curve. All that is left for people is "approval" and "responsibility".

Commercial accounts are more heart-warming. Content platforms used to rely on "scarce high-quality content" to retain users and sell advertisements; now high-quality content can be produced infinitely, and what is scarce is "deterministic sources" and "emotional connections." The "original tags" on Weixin Official Accounts, Xiaohongshu's "real-name authentication", and Substack's "subscription system" are essentially selling the same thing: human endorsements. Argon completely overflows the production side, and value anchors can only be migrated to the distribution side and trust side. I once read an internal PPT of an MCN: They are breaking up 200 contracted authors into an "IP matrix"-each IP is bound to a real face, real social account, and offline meeting. The content is all produced in batches by Argon. The operation team is only responsible for "human design maintenance." This model has worked, with a monthly turnover of 3 million yuan. But I have also seen another one: the same model, with all accounts dead for three months, because readers found that "this person never responds to comments, never broadcasts, and never makes mistakes"-too perfect is false.

Tools reshaping people has never been a linear substitution, but a move of "scarcity". Printing technology has moved "copying" to "editing"; cameras have moved "realistic painting" to "composition aesthetics"; search engines have moved "memory" to "retrieval and judgment"; and Argon is moving "text production" to "intention setting and responsibility covering." The difference is that the previous few moves left humans with a buffer period of decades, but this buffer period may only last for a few quarters. Before Argon opened the API, the number of internal beta quotas was sold to 50,000 yuan each by scalpers-not because it was well written, but because whoever got it first could first zero the "production cost" and seize the distribution channel.

The most ironic thing is that Argon writes code instead becomes weaker. This high probability is the result of the weight allocation of training data-Google feeds it more high-quality long texts (books, papers, in-depth reports), and the proportion of code corpus has dropped relatively. But this also exposes a deeper truth: big model manufacturers are beginning to optimize for "universal assistants" rather than "co-drivers for programmers." There are only tens of millions of programmers, and there are billions of people who can write, want to write, and are forced to write. Business accounts work: Argon can turn every small red book blogger, every civil servant who writes materials, every middle school student writes compositions, and every person who sells and writes solutions into a "high-yield author." Code generation is B2B, writing generation is C2C, and the latter's TAM is two orders of magnitude larger.

Then the question comes: When "writing well" is no longer scarce, and "what do you want to write","dare to write" and "are you willing to be responsible for the word" become the only scarcity, how can we distinguish creators from content farms? How to price "human endorsement"? How to prevent trust anchors from being forged in batches? Argon has no answer, and Google has no answer. The answer can only grow in the next round of game-like photographers learning to "compose", like programmers learning to "structure", like editors learning to "judge".

Writing transcends code, not because models have become smarter, but because the market has chosen a more profitable direction. Tools are never neutral. They move scarcity and people. The next one to be reshaped may be you and me who are reading this article.