DeepSeek Just Exposed The Rot At The Core Of The AI Industry

Every few decades, something comes along that obliterates the status quo. In the ’50s, it was Sputnik. In the ’90s, it was the internet. And in the modern ’20s, it appears to be DeepSeek, a Chinese AI aimed squarely at challenging OpenAI. So what makes DeepSeek so damn special? Well, it’s cheaper, more efficient, and generally higher quality than anything produced in the West. This model only cost $6 million to build, in comparison to OpenAI, which spent well over $3 billion on AI training last year alone and has required $13 billion in investment from Microsoft just to stay afloat. And it wasn’t just cheaper to build. The cost of actually running DeepSeek is over 96% cheaper than OpenAI’s o1 model! Unlike any OpenAI model, you can actually use DeepSeek for free. Yet, in third-party bench tests, DeepSeek outperformed every OpenAI model by some margin. Unsurprisingly, upon launch, DeepSeek became the top-rated free app on the App Store, and when news of this product broke, $1 trillion was wiped off the American tech stock market. In doing so, it has exposed the rotten ideology at the core of the AI industry. So, how has DeepSeek achieved this? What does it mean for the future of AI and the American tech industry? And, more importantly, what does this mean for you?

Let’s start with the technical stuff. How have they built such a capable model on such a small budget?

In this instance, necessity seemed to be the mother of innovation. The US has restricted the sale of high-end AI-optimised supercomputer GPUs to China for some time now. The idea was that these are essential infrastructures for developing AI, and this restriction would keep the US one step ahead of the game (so much for free market capitalism, hey?). As a result, DeepSeek was only able to use weaker, cheaper GPUs designed for gaming to develop their model. This forced them to be hyper-efficient.

Consequently, DeepSeek’s model is only 38% the size of OpenAI’s older ChatGPT-4 model, and it has been optimised to run efficiently on these less powerful chips. But this tiny size meant that if they took the same approach as OpenAI, their model would underperform dramatically! So, they made two critical changes.

Firstly, the architecture. OpenAI uses an AI architecture known as “fully dense.” This basically means that the architecture is comprised of a single, vast network that processes every request with all its parameters and data points. This is incredibly computationally dense, but the idea is that it can make it more capable in a broader application. DeepSeek is instead much more picky and uses a “mixture of experts” architecture. In this approach, the AI is split into many models designed to be better at answering certain queries, and there is a front-facing AI that can understand what kind of query it is being asked and triage it to the AI model best suited to answer it. This is a far more efficient model, as it only engages the parts of the AI needed to solve a problem and not the whole AI. It also means it requires less training, which is very costly, as these expert AIs are more focused and restricted in scope. However, as a trade-off, this approach, in theory, should make a model have a less broad application than those with fully dense architecture.

But we don’t see this application restriction when it comes to bench testing DeepSeek against its rivals or in the real world. How come?

Well, this is thanks to the second change DeepSeek made — ensuring the finished product is open-source and not closed-source like most Western AI models.

Closed-source AIs are developed in secret, and then use cases are found for the model once it is released, hence why OpenAI needs to use the costly “fully dense model” and why it seems like AI is being shoehorned into every possible corner of our lives, even if it doesn’t add any value. These closed-source models are expensive to build, and tech companies need to find applications to justify their astronomical expenditure on them. But open-source works the other way around. Users have specific use cases in mind and work with the developers to develop the AI to be usable in that specific application. This enables far more focused, efficient, and cheaper development that results in an AI that is more useful in areas that actually matter than in closed-source models.

So, why are most Western AI companies using a closed-sourced approach? Well, this creates a proprietary AI that is wholly owned by the AI company. This enables them to charge more for its use and also helps them raise more investment, as it creates a more secure asset. After all, a tech giant or investment bank isn’t going to bankroll an open-source model that they can’t have total power and control over or sell on for billions of dollars. More on this critical point in a minute.

I have seen a few rumours and articles claiming that DeepSeek’s dramatic efficiency is a vast leap forward for AI, and this will make human-like AI possible in the near future and resolve the technical issues in the AI world. This simply isn’t true. DeepSeek still suffers from the same scaling issues, lack of actual cognition, and query errors that all AIs suffer from. It is no better a tool than those from OpenAI, and its cheaper cost won’t open the door to far better tools. It is fundamentally the same technology with the same flaws; it’s just been managed and built in a non-moronic way. I have covered the limitations of AI extensively; to read more about it, read those articles here, here, and here.

So, the question has to be asked: why has the West gotten so caught up in wildly expensive and overly costly closed-source models?

Well, this is, I think, the main insight from DeepSeek. It’s not that China could dominate the AI race. It’s not that OpenAI isn’t that unique. Instead, it is that the US economy is no longer based on the real world and therefore doesn’t value genuine innovation. In short, it exposes how far America has fallen into late-stage capitalism.

OpenAI was originally meant to be open-source, but it changed to closed-source. Why? Well, Western venture capitalists see AI as a way for them to replace the human workforce with their own AI in basically every sector of the economy, enabling them to wield significant control and reap huge profits. This is the hype that has been pushed for years now, despite scientists and AI developers themselves saying that AI can’t, and might not ever be able, to do this. But these venture capitalists didn’t care; they knew that investing in AI and pushing this false narrative would increase the stocks and make them money, whether AI can achieve this feat or not. As such, OpenAI and almost every Western AI project switched to closed-source to enable such investment, as it enriches the investors and the executives of these AI companies.

In fact, this is the real reason why America restricted GPU sales to China — not to get ahead in the AI race but to protect American investment and hegemonic domination of the sector.

In short, they are making AI a worse and more expensive product than it needs to be, have removed any true free market economics for the sector at the cost of true innovation, while pushing misinformation about the technology all too specifically to enrich investors at the cost of the consumer. This isn’t so much late-stage capitalism and more akin to a messy, plutocratic, post-capitalist market intervention and all.

This is why DeepSeek wiped $1 trillion off the stock market despite not having proven itself in the real world and having some serious security issues. It represented real-world economics encroaching on the heavily protected and falsely inflated American bubble.

And this economic rot doesn’t just exist in the AI industry. It is now widespread across the entire Western economy. Indeed, this is why the middle class is being squeezed so hard, as the inbuilt laws and motions of capitalism that once protected them are being overridden to enrich the 1%.

So, yes, DeepSeek is a Sputnik or World Wide Web moment. Just not for AI. They haven’t actually changed the world of AI all that much, as all the horrific issues and fatal flaws of the technology still very much remain. But it exposes the economic rot at the heart of Western economies that is causing us so much pain, strife, and disenfranchisement. It is the wake-up call we need to move past the bullshit we are swimming in and get back to the real world. After all, if a quasi-Communist country can achieve market economics better than the US, something seriously wrong is afoot.

Thanks for reading. Content like this doesn’t happen without your support. If you want to support content like this, or read articles early, go and follow me and my project Planet Earth & Beyond or follow me on X.

Sources: Sky, The Independent, The Guardian, BI, Sky, TensorOps, Capacity Media, Al Jazeera, India Startup News, Reuters, Axios

详情

每隔几十年，总会有一些颠覆现状的事件发生。50年代，是斯普特尼克；90年代，是互联网；而在现代的20年代，似乎轮到了 DeepSeek——一款直面挑战 OpenAI 的中国人工智能。那么，DeepSeek 究竟有何过人之处？首先，它的成本更低、效率更高，而且整体质量远超西方任何同类产品。该模型的构建成本仅为 600 万美元，而相比之下，OpenAI 去年在 AI 训练上的投入就超过 30 亿美元，仅为了维持运营，还从微软获得了 130 亿美元的投资。而且，不仅构建成本低，DeepSeek 的实际运行成本比 OpenAI 的 o1 模型低了 96% 以上！与所有 OpenAI 模型不同，你甚至可以免费使用 DeepSeek。更为关键的是，在第三方的基准测试中，DeepSeek 明显超越了所有 OpenAI 模型。毫不奇怪，DeepSeek 一经推出便成为 App Store 上评价最高的免费应用，而这一消息一出，便使美国科技股市蒸发了 1 万亿美元。这一切揭露了 AI 行业核心深处的腐朽意识形态。那么，DeepSeek 是如何做到这一点的？这对 AI 以及美国科技产业的未来意味着什么？更重要的是，这对你又意味着什么？

让我们先从技术层面谈起：他们究竟是如何在如此有限的预算下打造出如此强大的模型的？

在这种情况下，迫不得已往往能催生创新。美国早已限制向中国出售高端、专为 AI 优化的超级计算机 GPU，其初衷在于这些 GPU 是开发 AI 的关键基础设施，而这一限制则意在让美国始终领先一步 (自由市场资本主义又算什么呢？)。因此，DeepSeek 只能依靠那些性能较弱、成本更低、专为游戏设计的 GPU 来开发其模型，这迫使他们必须做到极致高效。

结果是，DeepSeek 的模型规模仅为 OpenAI 早期 ChatGPT-4 模型的 38%，并经过特别优化，能够在这些性能不足的芯片上高效运行。但正因为模型如此小巧，如果照搬 OpenAI 的做法，其性能将大打折扣！因此，他们做出了两项关键改变。

首先是架构。OpenAI 采用了一种被称为“全密集（fully dense）”的 AI 架构，这意味着整个架构由一个庞大的单一网络构成，每次请求都调用所有的参数和数据点进行处理。这种方式计算密集，但旨在使模型在更广泛的应用中展现更强能力。而 DeepSeek 则更为挑剔，采用了“专家混合（mixture of experts）”架构。在这种模式下，AI 被拆分为多个擅长解决特定问题的模型，同时配备一个前端 AI 来判断请求的性质，并将其分流到最合适的专家模型。这种设计更为高效，因为它只激活解决问题所必需的部分，而非整个系统。同时，也意味着训练成本大幅降低——毕竟，专门化的专家模型所需的训练规模和范围都较为有限。作为代价，从理论上讲，这种方式可能会使模型的应用范围不如全密集架构那般广泛。

然而，无论是在对比测试中，还是在实际应用中，我们都没有发现这种应用范围的受限现象。这究竟是为何？

答案在于 DeepSeek 的第二项改变——确保最终产品是开源的，而不像大多数西方 AI 模型那样闭源。

闭源 AI 的开发过程常常秘密进行，模型发布后才去寻找合适的应用场景，这正是 OpenAI 不得不采用昂贵“全密集模型”的原因，也让人感觉 AI 正被强行塞进我们生活的各个角落，哪怕有时并无实际价值。这些闭源模型造价高昂，科技公司需要不断探索应用场景来证明其巨额投入的合理性。而开源模式则截然相反：用户往往已有明确的应用需求，并与开发者合作，量身打造出适用于特定场景的 AI。这种模式使开发更加专注、高效且成本更低，从而产出的 AI 在真正重要的领域内，比闭源模型更为实用。

那么，为什么大多数西方 AI 公司仍然偏爱闭源模式呢？原因在于，这种模式能够打造出完全归公司所有的专有 AI，使他们在使用上可以收取更高费用，同时也便于吸引更多投资，因为它构成了一项更为安全的资产。毕竟，没有哪个科技巨头或投资银行会为一个无法完全掌控，亦难以以数十亿美元出售的开源模型买单。关于这一关键点，我们稍后再详细讨论。

我曾见过一些传闻和文章宣称，DeepSeek 显著的效率代表了 AI 领域的一大飞跃，预示着类似人类的 AI 即将在不久的将来出现，并解决所有技术难题。但这根本不成立。DeepSeek 依然存在所有 AI 普遍面临的扩展性问题、缺乏真正认知能力以及查询错误等缺陷。它并不比 OpenAI 的工具更出色，更低的成本也不会为我们带来质的飞跃。归根结底，它依然是拥有相同缺陷的同一技术；只是其管理和构建方式更为得当罢了。关于 AI 局限性的更多讨论，请阅读以下文章：这里、这里和这里。

因此，我们不得不问：为什么西方如此痴迷于那些成本极高、价格昂贵的闭源模型？

我认为，这正是 DeepSeek 给我们的主要启示。问题不在于中国能否主导 AI 竞赛，也不在于 OpenAI 是否独一无二，而在于美国经济已不再立足于现实世界，从而也不再重视真正的创新。简言之，这揭示了美国已经深陷晚期资本主义的泥潭。

最初，OpenAI 本应是开源的，但后来转而采用闭源模式。为何如此？原因在于，西方风险投资家视 AI 为一种手段，借此在经济几乎每个领域用自己的 AI 替代人力，从而掌握更大控制权并获取巨额利润。尽管科学家和 AI 开发者早已指出 AI 无法（甚至可能永远无法）实现这一目标，但这种炒作已持续多年。这些投资家根本不在意；他们深知，无论 AI 能否实现这一目标，投资并推动这一虚假叙事都能抬高股价，为他们赚取丰厚回报。因此，OpenAI 以及几乎所有西方 AI 项目都转向闭源模式，以吸引这类投资，从而使投资者和企业高管获利。

事实上，这正是美国限制向中国出售 GPU 的真正原因——并非为了在 AI 竞赛中取得领先，而是为了保护美国的投资利益及其在该领域的霸权地位。

简而言之，他们正将 AI 做得比实际所需的更糟、更昂贵，牺牲真正的创新剥夺了自由市场经济，同时大肆散布有关这项技术的错误信息，以便在牺牲消费者利益的同时令投资者受益。这不仅仅是晚期资本主义的问题，更像是一种混乱的、寡头统治的后资本主义市场干预。

也正因如此，尽管 DeepSeek 在现实中尚未完全证明自己，且存在一些严重的安全隐患，却仍使股市蒸发了 1 万亿美元。这代表着现实经济正逐步侵入那被严密保护、虚高膨胀的美国泡沫。

而这种经济腐败不仅仅存在于 AI 行业，它如今已蔓延至整个西方经济。正因如此，中产阶级正受到前所未有的挤压，因为曾经保护他们的资本主义内在规律正被颠覆，以便让那 1% 的富豪更富有。

所以，是的，DeepSeek 确实可以被视为“斯普特尼克时刻”或“万维网时刻”，但这并非针对 AI 本身。它并没有真正改变 AI 的世界，因为所有那些可怕的问题和致命缺陷依然存在。但它揭露了西方经济核心的腐败，这正是导致我们饱受痛苦、纷争和边缘化的根源。这是一记警钟，提醒我们必须摆脱充斥的虚假繁文缛节，回归真实世界。毕竟，如果一个半共产主义国家都能比美国更好地实现市场经济，那么某些严重的问题必然存在。

感谢阅读。 没有你的支持，就不会有这样的内容。如果你也想支持此类内容，或希望抢先阅读更多文章，请关注我和我的项目 Planet Earth & Beyond 或关注我在 X 上的动态。

来源: Sky, The Independent, The Guardian, BI, Sky, TensorOps, Capacity Media, Al Jazeera, India Startup News, Reuters, Axios

欢迎关注我公众号：AI悦创，有更多更好玩的等你发现！

公众号：AI悦创【二维码】

.