TV
Chinese AI startup DeepSeek announced on Tuesday that it would temporarily limit user registrations due to large-scale malicious attacks on its services even though existing users can continue to log in without any issues.
A warning banner on its website read “Due to large-scale malicious attacks on DeepSeek's services, we are temporarily limiting registrations to ensure continued service.”
Notably, the cyberattack on DeepSeek came just as it gained considerable attention as a fast-growing rival to OpenAI's ChatGPT and Google's Gemini. Moreover, the Chinese company, which released its R1 model last week, has risen to the top of app store charts, dethroning Open AI’s ChatGPT.
The malicious attacks forced DeepSeek to investigate and temporarily limit new registrations. According to their status page, the attacks caused degraded performance for users globally. Several independent researchers mentioned on social media sites like X (formerly Twitter) that the attacks are likely large-scale Distributed Denial-of-Service (DDoS) Attacks.
Meanwhile, the company reported that the issue was identified and that they were continuing to monitor for further problems.
On the other hand, a Forbes report from Tuesday cited a cyber intelligence company, Kela, that raised concerns about the vulnerabilities in DeepSeek R1. Their researchers reportedly found that the model could be tricked into performing malicious tasks like coding ransomware and suggesting illicit activities.
ALSO READ: 'Bring it on!' We asked ChatGPT what it thinks about DeepSeek
Despite such claims, DeepSeek R1's success has thrown the global AI market haywire. The model was reportedly developed for a fraction of the cost compared to its American competitors. As such, not long after news of its success broke out, billions of dollars worth of stock money were wiped off major tech companies like Nvidia, marking the biggest crash in its history.
What exactly is DeepSeek?
DeepSeek is a Chinese AI startup founded in 2023. It rapidly gained prominence in the AI industry with its DeepSeek-R1 model that was released last week.
Notably, R1 is an open-source large language model (LLM) and is comparable in performance to leading LLMs like OpenAI's GPT-4 while being developed at a significantly lower cost (around $6 million compared to estimated $41-78 million for GPT-4).
What did DeepSeek do? How did it disrupt the market?
This has disrupted the AI landscape, challenging the dominance of major tech companies and increasing accessibility to advanced AI technology with its capabilities rivalling those of models created by giants like Google and OpenAI.
DeepSeek’s AI assistant app recently surpassed ChatGPT as the most downloaded app on the Apple App Store, dethroning OpenAI’s Chat GPT.
The company’s sudden success in challenging the dominance of major tech companies has had significant repercussions on the US stock market, particularly affecting tech giants. Nvidia, for instance, experienced a staggering one-day loss of approximately $600 billion in market capitalization, marking the largest single-day drop for a listed company in US history.
Other tech companies, including Microsoft and Google, also saw their stock values decline by several percentage points
How does DeepSeek R1 achieve its cost-effectiveness?
DeepSeek's success is attributed to its innovative approach and cost-effective development. The company's DeepSeek-V3, a 671B Mixture of Experts (MoE) model, serves as the foundation for DeepSeek-R1.
This base model reportedly performs on par with advanced models like Sonnet 3.5 and GPT-4o. Further, the company claims to have trained its model for just $5.5 million, significantly less than its competitors. This efficiency is achieved through architectural innovations such as Multi Token Prediction (MTP), Multi-Head Latent Attention (MLA), and extensive hardware optimization.
DeepSeek-R1 employs a unique training approach, combining reinforcement learning techniques like Group Relative Policy Optimization (GRPO) with a 'cold start' fine-tuning phase to enhance reasoning skills and output quality.
