4 min Applications

DeepSeek, hot on OpenAI’s heels, hit by cyberattack

DeepSeek, hot on OpenAI’s heels, hit by cyberattack

DeepSeek is taking the tech industry by storm. After surprising all with its “reasoning” model R1, the company has now launched a competitive AI image generator: Janus-Pro. However, the Chinese company is quickly learning the challenges that come with popularity, as its service is already experiencing large-scale cyber attacks.

DeepSeek has temporarily suspended new signups for its chatbot service (via chat.deepseek.com), though existing account holders can still access the platform. Updates are available on their status page, which notably runs on Atlassian Statuspage – the same system used by OpenAI.

Hurdle to take

Growing pains are to be expected for a service that has experienced such explosive growth. DeepSeek transformed from what many considered an obscure player to the #1 spot on the U.S. App Store in just a week’s time. Most users are accessing the technology through the DeepSeek AI Assistant app rather than running the open-source model via HuggingFace or GitHub locally. Since DeepSeek’s rise caught everyone off guard, scaling the service may prove challenging. Due to export restrictions on Nvidia chips, the Chinese company might need to get creative to find additional computing capacity. However, the most demanding work – AI training – is already complete, as AI inferencing (running the model to get outputs) for millions of users requires significantly less computing power.

The identity of the attackers remains unknown. Speaking somewhat tongue-in-cheek, some Nvidia shareholders might have the right motive to carry out DDoS attacks on DeepSeek. Like other western tech companies, Nvidia took a hit in the stock market yesterday, suffering a historic loss of approximately $600 billion (!) – nearly 17 percent of its market value. Broadcom’s stock took an even bigger dive of 17.4 percent. The AI competition is clearly rattled. OpenAI appears particularly vulnerable, as its model o1, once considered a key asset, is now closely matched by DeepSeek-R1.

Vision model

As if OpenAI didn’t have enough concerns, DeepSeek has now unveiled Janus-Pro. This new open-source AI model is multimodal and focused on visual generation – essentially an AI image maker similar to OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion. Released under the MIT license, other companies can build commercial products around it.

Janus-Pro does have limitations, particularly its restricted input resolution of 384×384. While DeepSeek claims the model competes with major AI market players in benchmarks, it hasn’t created the same shock waves as DeepSeek-R1. Vision models come in many varieties, and Janus-Pro’s outputs aren’t consistently high quality across all metrics. Nevertheless, it serves as another warning to U.S. AI companies that DeepSeek’s capabilities deserve serious attention – and provides valuable lessons for the open-source AI community.

Walls torn down

DeepSeek made its global debut this week. It took roughly a week for news of the R1 model to reach the broader public and Wall Street in particular. While this isn’t quite a “ChatGPT moment” – as it doesn’t represent an entirely new level of generative AI quality – this highly efficient and open model has put OpenAI in its most challenging position since ChatGPT’s launch. Previous competitors, including Google Bard (later Gemini), Claude, Microsoft’s Copilot suite, AI agents, and numerous others, failed to threaten OpenAI’s core advantages. Since late 2022, OpenAI has maintained the largest user base and consistently offered premium, top-tier AI models through its API.

An even more significant development may be on the horizon: what if DeepSeek scales its innovations to a model larger than 671 billion parameters? Or what if competitors like Meta, Anthropic, or Google apply similar principles to their own LLMs?

The next breakthrough could also come from OpenAI itself. Its CEO Sam Altman has called DeepSeek “an impressive model,” adding, “we will obviously deliver much better models and also it’s legit invigorating to have a new competitor! we will pull up some releases.”

Also read: DeepSeek introduces series of LLMs with high reasoning capabilities