In early 2025, a wave of breathless optimism swept through the artificial intelligence industry. Foundation models were getting bigger. Benchmarks were being shattered. Capital was flooding in at rates that made the dot-com era look restrained. And then, quietly at first, the cracks started to show.
Lee Han Chung, an AI engineer and researcher, published a provocative essay drawing an uncomfortable historical parallel: the current AI boom, he argued, bears a disturbing resemblance to Mao Zedong’s Great Leap Forward — the catastrophic 1958–1962 campaign in which China’s leadership, intoxicated by ideological fervor and fabricated metrics, drove the country into famine and economic ruin. The comparison is deliberately jarring. It’s also, on closer inspection, uncomfortably apt.
The thesis isn’t that AI will cause mass starvation. It’s subtler and more structural than that. What Chung identifies is a pattern of systemic dishonesty — inflated benchmarks, overpromised capabilities, misaligned incentives between those building AI systems and those deploying them — that mirrors the dynamics of a command economy running on false reports. In Mao’s China, local cadres reported impossible grain yields to please Beijing. In today’s AI industry, companies report impossible benchmark scores to please investors.
The Benchmark Problem: When the Map Replaces the Territory
At the center of Chung’s argument is a crisis of measurement. AI benchmarks — standardized tests like MMLU, HumanEval, and others — have become the primary currency by which foundation models are evaluated, funded, and marketed. The problem? They’ve become targets rather than measures.
This is Goodhart’s Law in its purest form: when a measure becomes a target, it ceases to be a good measure. And the AI industry has Goodharted itself into a corner.
Consider the progression. In 2023, GPT-4 stunned the world with its performance on the bar exam and various academic benchmarks. By mid-2024, virtually every major lab — OpenAI, Anthropic, Google DeepMind, Meta, Mistral — was releasing models that matched or exceeded GPT-4’s scores on those same benchmarks. The improvements looked extraordinary on paper. But users weren’t experiencing extraordinary improvements. Customer satisfaction surveys and real-world deployment data told a different story: models were getting marginally better at some tasks, worse at others, and dramatically better at taking tests.
Sound familiar? During the Great Leap Forward, steel production numbers soared on paper while the actual output was unusable pig iron smelted in backyard furnaces. The metrics looked spectacular. The reality was worthless.
Chung’s essay points to a specific mechanism: benchmark contamination. Training data increasingly includes the benchmarks themselves, or close derivatives. Models aren’t learning to reason better — they’re memorizing the test. Some labs have been more transparent about this than others. Most haven’t been transparent at all.
And the incentive structure makes honesty almost impossible. A lab that reports honest, lower benchmark scores will see its valuation drop, its recruiting pipeline dry up, and its enterprise customers defect to competitors with shinier numbers. So everyone inflates. Everyone knows everyone inflates. And the inflation continues because no one can afford to stop first.
This is the exact dynamic that drove Mao’s cadres to report grain yields of 10,000 jin per mu when actual yields were a fraction of that. Not because they were stupid. Because the system punished honesty and rewarded fabrication.
The consequences in AI aren’t famine. They’re subtler: enterprises deploying AI systems based on benchmark promises that don’t hold up in production; billions in capital allocated based on capability claims that evaporate under scrutiny; and a growing credibility gap between what the industry says its technology can do and what it actually does.
The Capital Misallocation Machine
The financial dimensions of this parallel are staggering. In 2024 and into 2025, AI-related venture capital and corporate investment reached levels that defied conventional analysis. Microsoft committed over $10 billion to OpenAI. Google, Amazon, and others poured comparable sums into their own efforts. Startups with minimal revenue commanded valuations in the tens of billions.
Chung frames this as a misallocation problem analogous to the Great Leap Forward’s diversion of agricultural labor into steel production. During that campaign, Mao ordered peasants to abandon their fields and build backyard furnaces, convinced that China could leapfrog Britain’s industrial output through sheer will and mass mobilization. The result: agricultural output collapsed while the steel produced was largely useless.
The AI parallel isn’t exact, but the structural similarity is real. Enormous resources — talent, compute, electricity, capital — are being concentrated in a narrow set of activities (training ever-larger foundation models) based on the assumption that scale alone will produce artificial general intelligence or something close to it. The scaling hypothesis. More data, more parameters, more compute equals more capability. Always.
Except the evidence for continued scaling returns is weakening. Multiple reports in late 2024 and early 2025 suggested that the major labs were hitting diminishing returns on pre-training alone. The era of easy gains from simply making models bigger appeared to be ending. Labs pivoted to post-training techniques — reinforcement learning from human feedback, chain-of-thought prompting, inference-time compute scaling — but these approaches have their own limitations and costs.
Meanwhile, the capital keeps flowing. Because stopping the flow would mean acknowledging that the scaling hypothesis has limits, which would mean acknowledging that current valuations are unsupportable, which would mean a repricing event that nobody in the industry wants to trigger.
So the backyard furnaces keep burning.
This dynamic extends beyond pure AI labs. Enterprises across every sector have launched AI transformation initiatives, many of them driven more by fear of being left behind than by clear-eyed assessments of what the technology can actually deliver today. Consulting firms have sold billions in AI strategy engagements. Corporations have hired chief AI officers. Press releases tout AI-powered everything. But the gap between announcement and implementation remains vast.
A Wall Street Journal analysis of corporate earnings calls in early 2025 found that mentions of AI had increased roughly fivefold compared to two years earlier, while actual AI-attributable revenue remained a rounding error for most companies outside the hyperscalers. The talk-to-walk ratio is extreme.
Chung draws another parallel here: the propaganda function. During the Great Leap Forward, state media amplified the fabricated production numbers, creating a feedback loop in which leadership believed its own propaganda. In the AI industry, the role of state media is played by a combination of tech press, social media influencers, and the companies’ own marketing machines. Benchmark results get amplified. Limitations get buried. Demo-ware gets presented as production-ready capability.
The result is a collective delusion — not universal, not total, but pervasive enough to distort capital allocation at a massive scale.
There are dissenters, of course. Researchers like Gary Marcus have been vocal about the limitations of current approaches. Yann LeCun at Meta has argued that autoregressive language models are fundamentally limited and that new architectures are needed. But these voices are often dismissed as pessimists or contrarians, much as Peng Dehuai — the Chinese general who dared to criticize the Great Leap Forward at the 1959 Lushan Conference — was purged for his honesty.
The industry doesn’t purge critics literally. But it marginalizes them. Funding dries up for research directions that challenge the dominant paradigm. Researchers who question scaling laws find themselves outside the inner circle. The social and professional costs of skepticism are real, even if they’re not as severe as what Peng Dehuai faced.
Where the Analogy Breaks — and Where It Holds
It’s worth being precise about where Chung’s analogy works and where it doesn’t. The Great Leap Forward killed an estimated 15 to 55 million people. The AI boom, whatever its excesses, is not going to produce anything remotely comparable in human suffering. Comparing software benchmark inflation to mass famine is, at one level, obscene.
But Chung isn’t making a moral equivalence argument. He’s making a structural one. The mechanisms are the same: a system that punishes honest reporting and rewards exaggeration; leadership that has committed to a vision so completely that contradictory evidence gets filtered out; a measurement apparatus that has been captured by the thing it’s supposed to measure; and a massive misallocation of resources driven by false signals.
The structural argument holds.
And there’s one more element that makes the comparison particularly sharp: the role of true believers. The Great Leap Forward wasn’t driven solely by cynicism. Many cadres genuinely believed that the revolutionary spirit of the Chinese people could overcome material constraints. Many AI leaders genuinely believe that scaling will produce AGI. The sincerity of the belief doesn’t make the outcomes less dangerous — in fact, it makes them more so, because true believers are harder to correct than cynics.
So what happens next? If the analogy holds, the AI industry is somewhere in 1959 — past the initial euphoria, entering the period where reality starts to assert itself but the system hasn’t yet adjusted. The Great Leap Forward didn’t end with a sudden crash. It ended with a slow, grinding recognition that the numbers were fake, the strategy was flawed, and the costs were mounting. The Chinese Communist Party eventually acknowledged the disaster — years later, partially, and with enormous reluctance.
The AI industry’s correction, if it comes, will likely follow a similar pattern. Not a sudden crash but a gradual repricing. Enterprise customers will start demanding proof of ROI rather than accepting benchmark scores. Investors will start asking harder questions about unit economics. The gap between demo and deployment will become too large to paper over.
Some companies will survive this correction. The ones building real products that solve real problems for real customers — not just chasing benchmark leaderboards. Others won’t. And the capital that was misallocated will be written off, the way it always is after a bubble.
But here’s what makes the current moment different from a typical technology bubble: the underlying technology is genuinely powerful. AI isn’t tulips. Large language models can do things that were impossible five years ago. The problem isn’t that the technology is fake. The problem is that the industry’s relationship with the truth has become badly distorted, and that distortion is leading to decisions — about investment, about deployment, about strategy — that don’t match reality.
The Great Leap Forward didn’t fail because China lacked potential for industrialization. It failed because the system for translating potential into progress was broken by dishonesty and ideological rigidity. The AI industry has the same disease. Whether it proves fatal to the current boom or merely debilitating depends on whether the industry can do something that Mao’s China couldn’t: correct course before the damage becomes irreversible.
The benchmarks are lying. The capital is flowing to the wrong places. The true believers are in charge. And the honest voices are being marginalized.
The question isn’t whether a correction is coming. It’s whether anyone in a position of power will listen before it arrives on its own terms.
WebProNews is an iEntry Publication