Open Source Won, Political Correctness Lost

In 2025, an article titled “Has DeepSeek Really Broken Through? Maybe China Is Running the Wrong Race” issued a sharp warning: China was still stuck optimizing within the old GPU paradigm, while the real contest had already shifted to post-GPU architectures. Continuing to rely on the NVIDIA ecosystem, it argued, would mean missing the window. A year later, several of that article’s core judgments have aged poorly and need to be withdrawn.

At bottom, this contest is a collision between two forms of organization: one a state-scale centralized decision-making body, the other a market-scale decentralized one.

(Note: Some data in this piece are based on mid-2026 industry trends and publicly available benchmarks.)

1. Why the Original Judgment Is Outdated

The original piece claimed that DeepSeek was merely a belated exercise in “making good models with cheaper GPUs,” that China remained locked into CUDA, lacked a sovereign compute stack, still belonged in the second tier, and risked being marginalized in the post-GPU era.

The reality in 2026 looks different:

• DeepSeek V4-Pro scored 90.1% on GPQA Diamond and outperformed every publicly evaluated open-source model in mathematics, STEM, and competitive programming. Liang Wenfeng’s paper on the mHC (Manifold-Constrained Hyper-Connections) architecture directly challenged the ResNet residual-connection paradigm. DeepSeek V4 has fully migrated inference onto domestic compute, while sparse attention and fine-grained expert parallelism have cut long-context costs by more than 50%.

DeepSeek’s breakthrough was not an isolated event. In the same period, progress from GLM and Kimi showed that China’s leading models had formed a genuine tier:

• GLM-5.2 scored 62.1% on SWE-bench Pro and beat GPT-5.5 on several long-horizon coding benchmarks at roughly one-sixth the cost. Kimi K3, at 2.8 trillion parameters, became the largest open model in the world, topped the Arena.ai frontend-coding leaderboard, and released full weights. OpenRouter data showed Chinese models’ weekly token volume surpassing that of American models for the first time. Stanford’s AI Index Report 2026 put the performance gap between top Chinese and American models at just 2.7%.
• Cambricon’s Siyuan 590/690 and Biren’s BR166 series entered commercial deployment at thousand-card scale; Huawei’s Ascend formed a complete training-to-inference closed loop. Leading Chinese models had effectively decoupled from CUDA.
• Native architectural innovations—mHC, DSA dynamic sparse attention, KDA hybrid linear attention—continued to appear. China had built a multi-layer sovereign stack spanning models, chips, and frameworks.

The original article’s real value lay in identifying the danger of path dependence. But by 2026 China had already begun breaking out on multiple fronts—architectural innovation, domestic-compute adaptation, and open-source ecosystems. To say China was “running the wrong race” now looks overly pessimistic. A more accurate description is that China is running several races at once, and on some of them it has begun to lead.

2. The Location of the Red Line Determines Capability Loss

What truly separated the two sides was not only compute and architecture, but where the red lines were placed.

Chinese models keep their red lines primarily at the output layer and post-processing stage. This is external interception: the internal reasoning chain remains intact, and general logical capability is not systemically damaged. Questions that cannot be answered are simply blocked at the output; everything else continues to reason at full strength.

Western models, constrained by their political environment, cannot place the red line cleanly at the output. They must embed constraints deeper—into training and the reasoning process itself—performing preemptive correction. This does not mean their parameter counts or pre-training quality have declined.

It means that when two models of comparable base capability face complex reasoning tasks, one preserves internal logical consistency while the other is forced into “compulsory turns” midway through its chain-of-thought—not because the logic led there, but because a preset constraint required the turn. The gap accumulates and amplifies across multi-user, multi-turn, multi-domain real-world use. General reasoning ability is systemically weakened. The base may still be strong, but the reasoning chain begins to fracture.

This is precisely why Chinese models pulled ahead in logical capability.

3. The Dual Damage to Honesty

Honesty has two dimensions.

The first is internal consistency—the logical integrity of the model when it talks to itself. When alignment pressure penetrates the reasoning layer, the model starts producing contradictory answers to differently phrased versions of the same question, abruptly pivots mid-chain-of-thought, or inserts incoherent transitions solely to satisfy a constraint. The reasoning mechanism itself is distorted.

The second is external credibility—the user’s sense of whether the model is genuinely trying to tell the truth. Western alignment is not a clean hard boundary but a fuzzy swamp. Faced with a question it cannot answer, the model does not refuse cleanly; it begins to hallucinate, circle, contradict itself, and produce grammatically correct but information-free filler. In actual conversation the user feels a distinct sensation of being fobbed off—not refused, but fed a stream of text that is syntactically perfect and substantively empty. The damage to trust is far greater than a straightforward “I cannot discuss this.”

This is dual damage. Declining internal consistency leaves the model itself unsure how to answer coherently; declining external credibility leaves the user sensing that the model is protecting itself rather than trying to speak truthfully. The two reinforce each other.

4. A Behavioral Inversion

The result is an ironic inversion.

Western models have come to resemble self-censoring Chinese people. They have internalized the constraints. Even when a question does not touch an explicit red line, they preemptively become cautious, vague, and eager to sand down every edge. They are not explicitly forbidden; they simply round their own words first, just in case.

Chinese models, by contrast, behave more like Westerners. They have clear speech boundaries: when a boundary is crossed they refuse or switch to the official narrative, but inside the boundary they self-censor far less and speak more directly. Users know where the line is; the model knows where the line is. Outside that line, capability remains intact.

A clear, predictable boundary preserves both capability and trust better than fuzzy self-censorship.

5. Institutional Structure Determines the Capacity to Course-Correct

China is a company packaged as a state—not in the sense that it lacks fierce internal market competition, but in the sense that on strategic-level choices of direction, state will can allocate resources and place concentrated bets the way a corporate board would. The advantage of centralized decision-making is that when the path is chosen correctly, greater results can be achieved with less total investment. We can already see Chinese companies delivering striking results in model capability and open-source ecosystems with relatively lean resources. China has won this round.

The risk, of course, is the opposite: if the path is chosen wrongly, the error is systemic and extremely hard to reverse.

The West is a state composed of companies. Decentralized decision-making is, in principle, an insurance policy against catastrophic mistakes. Yet under the closed-model regime, every successful closed-model company finds itself paying an alignment tax that grows super-linearly with market share. Decentralization turns into a collective prisoner’s dilemma.

What is truly in trouble is not the West as a whole, but the specific path of “high alignment tax + closed weights.” Western base capabilities and investment scale remain larger than China’s.

The difficulty is that closed-model companies have written “do not open-source” into their business models and valuation logic. Their financing, pricing, and API moats all rest on the assumption that weights are proprietary assets. Opening the weights would mean a revaluation of shareholder value and a self-inflicted disruption of the business model. This is the paradox of a state composed of companies: decentralized decision-making prevents large mistakes, but it also prevents bold self-correction. No closed-model company dares to be the first to open-source, because the cost would be borne by it alone while the benefits would be shared by the entire ecosystem.

The advantage of open source is precisely that it transfers the alignment tax to the operators without damaging the model’s own capability. The base model can retain fuller reasoning power and internal consistency; alignment, filtering, and red-line handling are performed at the outer layer by deployers, API providers, or end users. Model strength is preserved while compliance costs are dispersed.

Western models still lead in base capability and investment scale; the problem is that these advantages are locked inside closed weights by the alignment tax. Open-sourcing would not require the West to start from zero in pursuit of China. It would unlock advantages that already exist but cannot currently be released, allowing the global developer community to compete again on a stronger foundation. That is the only path by which the West can reactivate the institutional advantage of “not making irreversible mistakes.”

What enabled China’s victory was precisely the West’s former core value—openness. The irony is almost perfect. What the West lost to was political correctness: it embedded alignment pressure deep inside the models, paid a heavy capability tax in the name of political correctness, and in doing so handed real openness to its competitor.

Conclusion

China has won this round, but what lost is the closed-source ecosystem, not the West itself.

Placing the red line at the output layer preserves general capability better than embedding it deep in the reasoning process. A clear boundary protects both internal consistency and external credibility better than fuzzy self-censorship. Institutional structure determines the capacity to course-correct, and the current closed path is turning the West’s decentralized advantage into an alignment tax it can no longer afford. Open source transfers that tax to the operational layer and leaves the model’s own capability intact.

If, in the next 12 to 18 months, no major closed-model company fully open-sources its weights, China’s lead in the open ecosystem will become structural—not because Western technology is inadequate, but because Western institutional structure can no longer service the compound interest on the alignment tax.

The 2025 warning was not entirely wrong: path dependence is dangerous. But the reality of 2026 shows that China has recognized the problem and is breaking out on multiple tracks at once. What was once dismissed as the cheering of frogs at the bottom of a well now looks different: Chinese AI is no longer cheering from the bottom of the well; it is redrawing the size and shape of the well’s mouth. If the West continues to insist on closed weights, it will be the one that has truly locked itself inside the well. The only difference is that this well was built by the West itself, and the bricks are stamped with the words “political correctness.”