DeepSeek R1：正在嘎嘎作響矽谷的AI地震

隨著DeepSeek R1的引入，人工智能（AI）的景觀經歷了重大的動盪，這種模型不僅是一個新的參賽者，而且是潛在的遊戲改變者。

Artificial Intelligence (AI) has taken a new turn with the introduction of DeepSeek R1, a model that is not just a new entrant but a potential game-changer in the AI industry. In a recent development that has sent shockwaves through the tech world, DeepSeek R1 has managed to match the performance of AI giants like OpenAI’s o1 at a mere 3-5% of the cost. This efficiency is not just a market disruptor but a technical marvel that challenges the very foundations of how AI models have been developed and deployed.

人工智能（AI）隨著DeepSeek R1的推出，這是一個新的轉變，該模型不僅是新進入者，而且是AI行業的潛在遊戲改變者。在最近通過技術界引發衝擊波的發展中，DeepSeek R1設法以Openai的O1（例如Openai的O1）的表現，僅佔成本的3-5％。這種效率不僅是一個市場破壞者，而且是一個技術奇蹟，它挑戰瞭如何開發和部署AI模型的基礎。

DeepSeek R1’s benchmark performances are nothing short of impressive. “On the AIME mathematics test, it scored 79.8% compared to OpenAI’s 79.2%,” Siegler highlighted, underscoring its capability. The model also achieved a 97.3% accuracy on the MATH-500 benchmark, surpassing OpenAI’s 96.4%. These achievements come with a dramatic reduction in operational costs, with DeepSeek R1 running at “55 cents per million token inputs and $219 per million token outputs,” in stark contrast to OpenAI’s higher rates. This cost-performance ratio is a wake-up call for the industry, suggesting a shift towards more economically viable AI solutions.

DeepSeek R1的基準表演令人印象深刻。 Siegler強調說：“在AIME數學測試中，它的得分為79.8％，而OpenAI的79.2％則得到了79.2％的速度。”該模型還達到了Math-500基準的97.3％精度，超過了Openai的96.4％。這些成就的運營成本急劇降低，DeepSeek R1的運行方式“每百萬個代幣投入55美分，每百萬個代幣產量為219美元”，與Openai的較高利率形成鮮明對比。這種成本績效比率是對該行業的喚醒呼籲，這表明向更經濟可行的AI解決方案轉變。

The market has responded with what can only be described as shock. Siegler pointed out, “In pre-market trading, Nvidia was down 10 to 11%,” with other tech behemoths like Microsoft and Google also witnessing significant drops. This market reaction signals a potential reevaluation of investment in AI infrastructure, particularly in hardware like Nvidia’s GPUs, which have been at the heart of AI’s scaling narrative.

市場的反應只能被描述為震驚。西格勒指出：“在市場前交易中，NVIDIA下降了10％至11％，”其他技術龐然大物等其他技術龐然大物也見證了大量下降。這種市場反應標誌著對AI基礎設施的投資的潛在重新評估，尤其是在Nvidia的GPU等硬件中，這是AI擴展敘事的核心。

From a technical standpoint, DeepSeek R1’s architecture is a testament to innovation under constraint. “It’s based on a mixture-of-experts architecture,” Siegler explained, allowing the model to activate only necessary parameters for each query, thus optimizing for both speed and efficiency. This approach contrasts with the monolithic models that activate all parameters regardless of the task at hand, leading to higher computational and energy costs.

從技術角度來看，DeepSeek R1的體系結構證明了受到限制的創新。 Siegler解釋說：“它基於Experts體系結構的混合物。”該模型僅激活每個查詢的必要參數，從而優化了速度和效率。這種方法與單層模型形成鮮明對比，該模型激活了所有參數，而不論手頭的任務如何，從而導致更高的計算和能源成本。

The model’s development involved a process of distillation from larger models to create compact yet potent versions. “They took, for example, a Llama model with 70 billion parameters and distilled it down,” said Siegler, outlining how DeepSeek managed to maintain high performance with fewer resources.

該模型的開發涉及從較大模型的蒸餾過程，以創建緊湊而有效的版本。 Siegler說：“例如，他們以700億個參數的Llama模型，並將其蒸餾出來。”

DeepSeek R1 diverges from the prevalent self-supervised learning methods by employing pure reinforcement learning (RL). “The models tend to figure out what’s the right answer on their own,” noted Siegler, indicating that this self-guided learning approach not only reduces the need for vast labeled datasets but also fosters unique reasoning capabilities within the model. This RL focus has allowed DeepSeek to fine-tune models through trial and error, improving their reasoning without the need for extensive human annotation, which is both cost and time-intensive.

DeepSeek R1通過採用純強化學習（RL）來與普遍的自我監督學習方法不同。 Siegler指出：“這些模型傾向於弄清楚正確的答案。”他表明，這種自導的學習方法不僅減少了對龐大的標記數據集的需求，而且還促進了模型中獨特的推理能力。這種RL重點使DeepSeek通過反複試驗來微調模型，改善了其推理，而無需大量的人類註釋，這既是成本又是時間密集的。

The scaling hypothesis, which posits that performance increases with more compute, data, and time, is now under scrutiny. “DeepSeek has shown you can actually do all this without that,” Siegler remarked, suggesting that the era of simply scaling up might be nearing an end. This could potentially reduce the dependency on massive hardware investments, redirecting focus towards smarter, more efficient AI development strategies.

縮放假設認為，隨著更多的計算，數據和時間，績效的增加，現在正在受到審查。 Siegler表示：“ DeepSeek表明您實際上可以做所有這一切。”他表明，簡單擴展的時代可能即將結束。這可能有可能減少對大規模硬件投資的依賴，從而將注意力轉向更智能，更有效的AI開發策略。

The immediate market fallout has been significant, with Nvidia’s stock plummeting. “It’s going to be pretty hard for this day at least,” Siegler observed, reflecting on the market’s knee-jerk reaction. However, some see this as a long-term opportunity for companies like Nvidia, where increased efficiency might spur demand for more specialized, less resource-heavy AI hardware.

NVIDIA的股票下降，直接市場的影響很大。 Siegler觀察到，對市場的膝蓋反應進行了反思：“至少這一天將非常困難。”但是，對於Nvidia等公司來說，這是一個長期的機會，在這種情況下，提高效率可能會刺激對更專業，資源不足的AI硬件的需求。

The business implications are profound. Companies like Microsoft and Google, which have been integrating AI into their ecosystems, now face a dilemma. “If the underlying economics just totally changed overnight, what does that do to their models?” Siegler questioned. This might push these companies towards reimagining their AI offerings, possibly leading to price adjustments or new service models to align with the new cost structures.

業務影響是深遠的。像Microsoft和Google這樣的公司已經將AI集成到其生態系統中，現在面臨困境。 “如果基本經濟學在一夜之間完全改變，那對他們的模型有什麼影響？”西格勒質疑。這可能會促使這些公司重新構想其AI產品，這可能會導致價格調整或新的服務模型，以與新的成本結構保持一致。

There’s a dichotomy in how this development is perceived. On one hand, there’s optimism that efficiency will lead to broader adoption and innovation. On the other, there’s caution about the implications for companies that have invested heavily in scaling. “Do we continue to spend billions for marginal gains, or do we leverage this efficiency to push towards practical AI applications?” Siegler pondered.

這種發展的感知方式有二分法。一方面，有樂觀的效率將導致更廣泛的採用和創新。另一方面，人們對在擴展方面投入大量投資的公司的影響有所謹慎。 “我們是否繼續花費數十億美元來獲得邊際收益，還是利用這種效率來推動實用的AI應用？”西格勒（Siegler）思考。

In response, tech leaders are attempting to calm the markets with narratives around increased efficiency leading to higher usage, with Nadella citing Jevons Paradox. “It feels like there’s a group text going on,” Siegler said, hinting at a coordinated message to reassure investors.

作為回應，技術領導者試圖以提高效率的敘述來平息市場，納德拉引用傑文斯（Jevons）的悖論。西格勒說：“感覺就像有一個小組文字正在進行。”

The ultimate test for DeepSeek R1 and similar models will be their application in real-world scenarios. “We need to see AI applications like we need to see an economy that takes use of

DeepSeek R1和類似模型的最終測試將是它們在實際情況下的應用。 “我們需要看到AI應用程序，就像我們需要看到的經濟