deepseek: everything you need to know about the ai chatbot app

DeepSeek's Rapid Ascent to Prominence

The AI company DeepSeek has recently experienced a surge in popularity.

This week, the Chinese AI laboratory DeepSeek gained significant attention as its chatbot application achieved the number one position on both the Apple App Store and Google Play Store charts.

The Implications of DeepSeek's Technology

DeepSeek’s AI models, developed through methods focused on computational efficiency, have prompted analysts on Wall Street and within the technology sector to reassess the United States’ potential to remain at the forefront of AI development.

Furthermore, questions are being raised regarding the long-term sustainability of demand for AI chips.

Origins and Rise of DeepSeek

The swift rise of DeepSeek to international recognition has naturally led to inquiries about the company’s background and the factors contributing to its rapid success.

Understanding the company’s origins is key to grasping its current position in the competitive landscape of artificial intelligence.

The Founding of DeepSeek and its Connection to High-Flyer Capital

DeepSeek’s origins are closely tied to High-Flyer Capital Management, a China-based quantitative hedge fund. This fund leverages AI to enhance its investment strategies and trading decisions.

Liang Wenfeng, a dedicated proponent of artificial intelligence, was instrumental in establishing High-Flyer in 2015. Wenfeng’s initial exploration of trading began during his studies at Zhejiang University.

High-Flyer Capital Management was formally launched as a hedge fund in 2019. Its primary focus was the creation and implementation of AI algorithms for financial markets.

In 2023, High-Flyer initiated DeepSeek as a dedicated research laboratory. This lab was designed to explore AI tools independently from the fund’s core financial operations.

Subsequently, with High-Flyer contributing as an investor, the lab transitioned into a standalone entity, also named DeepSeek.

Infrastructure and Challenges Faced by DeepSeek

From its inception, DeepSeek has invested in building its own data center infrastructure for the purpose of model training. However, the company, like many other AI firms in China, has encountered difficulties due to U.S. restrictions on hardware exports.

Specifically, to facilitate the training of a recent model, DeepSeek was required to utilize Nvidia H800 chips. These chips represent a less potent alternative to the H100, which is accessible to companies within the United States.

DeepSeek’s Team and Recruitment Strategy

The technical workforce at DeepSeek is characterized by its youthfulness. The company is known to actively seek out and recruit doctoral-level AI researchers from leading Chinese universities.

Furthermore, DeepSeek employs individuals from diverse academic backgrounds, even those without formal training in computer science. This approach aims to broaden the technical team’s understanding across various disciplines, as reported by The New York Times.

This diverse hiring strategy is intended to ensure the AI developed is well-rounded and applicable to a wider range of subjects.

DeepSeek’s Advanced Models

DeepSeek initially introduced its model suite – encompassing DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat – in November 2023. However, significant attention within the AI sector wasn't garnered until the following spring with the release of the next-generation DeepSeek-V2 model family.

The DeepSeek-V2 system, designed for general-purpose text and image analysis, demonstrated strong performance across a range of AI benchmarks. Notably, it offered a more cost-effective operational expense compared to its competitors at the time.

This performance prompted adjustments from DeepSeek’s Chinese competitors, including ByteDance and Alibaba, who responded by reducing usage prices and even offering some models without charge.

Further solidifying DeepSeek’s position, the launch of DeepSeek-V3 in December 2024 significantly enhanced the company’s reputation.

Internal benchmarks conducted by DeepSeek indicate that DeepSeek V3 surpasses both openly accessible models, such as Meta’s Llama, and API-accessed, “closed” models like OpenAI’s GPT-4o in overall performance.

The R1 “reasoning” model from DeepSeek is also particularly noteworthy. Released in January, DeepSeek asserts that R1 achieves comparable results to OpenAI’s o1 model on crucial benchmarks.

The Benefits of a Reasoning Model

As a reasoning model, R1 incorporates a self-fact-checking mechanism, mitigating common errors encountered in standard AI models. While reasoning models may require slightly longer processing times – typically seconds to minutes – they generally exhibit greater reliability in fields like physics, mathematics, and the sciences.

However, DeepSeek’s models, including R1, DeepSeek V3, and others, are subject to scrutiny by Chinese internet regulators. This ensures their responses align with “core socialist values.”

Consequently, the DeepSeek chatbot application, utilizing R1, will not address inquiries concerning sensitive topics such as Tiananmen Square or the autonomy of Taiwan.

In March, DeepSeek recorded over 16.5 million visits. According to David Carr, editor at Similarweb, “For March, DeepSeek is in second place, despite seeing traffic drop 25% from where it was in February, based on daily visits.”

Despite this strong showing, DeepSeek’s traffic remains lower than that of ChatGPT, which exceeded 500 million weekly active users in March.

During May, DeepSeek made an updated version of its R1 reasoning AI model available on the Hugging Face developer platform.

In September, DeepSeek introduced V3.2-exp, a new experimental model engineered for substantially reduced inference costs during long-context operations.

A Novel Competitive Strategy

The precise business model employed by DeepSeek remains somewhat ambiguous. The company’s pricing for its offerings consistently falls below prevailing market rates, with numerous products being provided at no cost.

Notably, DeepSeek has eschewed venture capital funding despite significant interest from investors.

DeepSeek attributes its ability to maintain such aggressive pricing to substantial gains in operational efficiency. However, the validity of the company’s reported figures has been questioned by certain industry analysts.

Despite not being strictly open source, DeepSeek’s models are accessible under licenses that permit commercial applications, leading to widespread adoption by developers. Clem Delangue, CEO of Hugging Face, notes that over 500 derivative models based on DeepSeek’s R1 have been created on their platform, accumulating 2.5 million downloads.

The impact of DeepSeek on established AI competitors has been characterized as both “disruptive” and “exaggerated.” The company’s emergence contributed to an 18% decline in Nvidia’s stock price in January and prompted a public statement from OpenAI’s CEO, Sam Altman.

In March, Reuters reported that U.S. Commerce Department bureaus instructed staff to prohibit DeepSeek’s models on government-issued devices.

Microsoft has announced the availability of DeepSeek through its Azure AI Foundry service, consolidating enterprise AI solutions. During Meta’s first-quarter earnings call, CEO Mark Zuckerberg affirmed that investment in AI infrastructure would remain a key “strategic advantage” for the company, when questioned about DeepSeek’s influence.

OpenAI labeled DeepSeek as “state-subsidized” and “state-controlled” in March, advocating for a potential U.S. government ban on its models.

Nvidia CEO Jensen Huang highlighted DeepSeek’s “excellent innovation” during the company’s fourth-quarter earnings call, emphasizing that models requiring substantial computational resources, like DeepSeek, benefit Nvidia.

Conversely, several entities, including entire nations and governments – such as South Korea – have initiated bans on DeepSeek’s use. The state of New York has also prohibited its utilization on government systems.

Brad Smith, Microsoft’s vice chairman and president, testified in a Senate hearing in May that Microsoft employees are restricted from using DeepSeek due to concerns regarding data security and the potential for propaganda.

The future trajectory of DeepSeek remains uncertain. Further model improvements are anticipated. However, increasing scrutiny from the U.S. government regarding perceived foreign influence is evident. Reports in The Wall Street Journal in March indicated a likely ban on DeepSeek for government use within the U.S.

This article was first published on January 28, 2025, and will undergo periodic updates.