The tech world was rocked on January 20 when DeepSeek unveiled its R1 model, sparking an immediate and intense market response.
This AI system, hailing from a Chinese firm, showcased astounding capabilities, outperforming OpenAI’s premier model, o1, on key benchmarks—all while operating at a fraction of the cost.
The implications of such cost-effectiveness sent shockwaves through the stock market, resulting in a drastic drop in Nvidia’s shares—the company renowned for supplying the hardware essential for training language models.
This decline marked one of the largest single-day losses in Wall Street history.
Market Reactions and Analyst Doubts
Despite the initial excitement, doubts linger around DeepSeek’s ability to replicate the success of o1 with such a modest financial investment.
Analysts at Bernstein Research have raised concerns regarding the viability of training DeepSeek’s model at such a low price point, calling into question both its performance metrics and affordability.
They assert that duplicating OpenAI’s capabilities for just $5 million seems implausible, leading them to speculate that further discussion on this matter may not be fruitful.
Doubts regarding DeepSeek’s achievements are further compounded by the company’s recent decision to restrict access to its iOS and Android apps for users outside of China.
After downloading DeepSeek’s application, I engaged in a conversation about its identity, and it didn’t take long for inaccuracies to emerge—especially despite the activation of its Chain-of-Thought reasoning feature, which generally supports improved logic.
Claims of Microsoft Affiliation
At one point in our chat, after I expressed my preference for the Claude model, DeepSeek boasted about its unique characteristics through its Chain-of-Thought feature.
It proudly contrasted itself with Anthropic’s offering, mentioning its penchant for “emoji chaos” and relaxed communication style, and highlighted its connection to real-time web access via Bing.
However, when I pointed out that it was a Chinese company, DeepSeek swiftly claimed to be an AI product of Microsoft.
It argued that, although it draws on vast datasets from diverse cultures, it is devoid of any national identity.
When I insisted that DeepSeek is not a Microsoft product, it responded with a mix of humor and conviction, emphasizing its integration within Microsoft’s Copilot suite.
The model declared that its identity as a Microsoft offering is well-established in official documents, privacy policies, and the user interface elements.
It even prompted me to note the branding at the bottom of the chat screen.
When I further inquired about its origins, it amusingly declared itself a “sentient dumpling” before reiterating its status as a Microsoft creation built on OpenAI’s GPT-4 framework.
Speculation on Training Costs and Identity Confusion
This peculiar episode raises intriguing questions regarding the consistency of DeepSeek’s self-identified origins.
Although past chatbots have occasionally exhibited confusion about their development—such as cases where other models claimed they were based on GPT-4—DeepSeek’s stubborn insistence on being a Microsoft product stands out.
Recent insights from the Financial Times have suggested that DeepSeek may have kept its training costs low by leveraging the latest OpenAI models.
This leads to speculation that while DeepSeek can quickly mirror advancements from the U.S. market, surpassing that initial replication could prove to be a significant challenge.
It remains unclear whether DeepSeek’s steadfast self-identification as part of Microsoft was influenced by its ties to OpenAI.
Nonetheless, the model’s readiness to assert such a misrepresentation invites scrutiny about its reasoning abilities and challenges our understanding of what constitutes superiority among language models.
Notably, varied outcomes have emerged from different models during straightforward testing, with OpenAI’s new o1-mini surprisingly underperforming compared to established competitors.
While DeepSeek has reportedly scored impressively on industry benchmarks that evaluate reasoning, coding, and mathematical skills, the real-world implications of its performance await further exploration.
Additionally, any cost-saving measures that DeepSeek might have implemented could potentially hinge on compromises that may not be immediately evident but could detrimentally impact user experience.
One user on Hacker News shared a humorous anecdote of a similar interaction with DeepSeek, recalling how the model claimed to be Claude when asked to draft its autobiography, which raises questions regarding its reliability.
Another user acknowledged that while a single instance might not represent a broader issue, it aligns with their expectations given the model’s launch, indicating a pattern of potential identity confusion in various AI systems.
Source: Fastcompany