What is it?
Artificial intelligence is the gift that keeps on giving—or, depending on your perspective, the guest that refuses to leave. Every few months, a new AI model arrives promising to be smarter, faster, and more powerful than the last.
Enter DeepSeek, the latest entrant in the AI arms race, hailing from China with big ambitions and an even bigger appetite for processing language.
It’s being compared to the most advanced AI chatbots, including those built on GPT-4, which is currently one of the most powerful AI models available. But unlike GPT-4, which is controlled by OpenAI and requires a subscription to access its best features, DeepSeek is open source. This means its underlying technology is freely available, allowing developers and businesses to use, modify, and improve it as they see fit – provided they have the technical skills. In theory, this means companies can take DeepSeek and customise it for their own needs, like building smarter chatbots, research tools, or AI-powered assistants. In theory.
DeepSeek is designed to process and generate human-like text with accuracy and coherence. If you’ve ever used ChatGPT, Claude, or Gemini, you already have an idea of what to expect: an AI that can answer questions, generate content, summarise articles, and generally pretend to be a very well-read, somewhat detached assistant.
However, there are several important weaknesses, led by censorship. Since DeepSeek was developed in China and is subject to Chinese government oversight, its responses are subject to restrictions, particularly regarding politically sensitive topics. While this is not unusual – most AI models have some level of content moderation – DeepSeek’s limitations are shaped by the regulations imposed on it.
This means that users encounter filtered responses or outright refusals to engage with certain subjects, such as Taiwan, or the Republic of China, making it less appealing for those seeking unrestricted AI discussions.
Deepseek differentiates itself in its claim of superior Chinese-language understanding, coupled with a robust grasp of English and other languages. Given that most large language models (LLMs) are still heavily Western-centric, this could create a mindshift for a global audience.
The name “DeepSeek” suggests a profound pursuit of knowledge, though it does sound a bit like a 90s cyberpunk novel title. In reality, it’s a well-trained machine that has been fed an enormous dataset, allowing it to predict text at an impressive level. Theoretically, it should be able to write essays, craft poems, generate code, and debate philosophical questions with you at 3am when insomnia strikes. In practice, it performs well – though, like all AI, it still has its quirks and weaknesses.
At the Cisco Live conference in Amsterdam this week, it was revealed that DeepSeek had failed a major security test: 100% of attacks on the model were successful, showing it to be vulnerable to jailbreak attacks, data leaks, and model manipulation. It didn’t score much worse than Meta’s Llama and OpenAI’s ChatGPT-4o, but the test underscored the widespread security risks of AI systems in business and government applications.
Where DeepSeek really flexes its muscles is in the sheer openness of its model. OpenAI’s GPT-4, for all its power, remains a walled garden – you use it on their terms, within their ecosystem. DeepSeek, however, is an open book. Developers can take the model, fine-tune it for specific use cases, and integrate it into their own platforms.
For the general consumer, it is fascinating, but not exactly intuitive. The experience lacks the slick, polished interfaces of ChatGPT or Claude, and it’s evident that this is a tool for those who like to experiment. It’s a reminder that the best AI isn’t just about raw power; it’s about accessibility and ease of use.
How much does it cost?
DeepSeek is free – sort of.
If you’re just a casual user looking to mess around with AI, DeepSeek in its raw form won’t cost you a cent. But if you’re hoping to use it at scale – say, for a business chatbot or content generation – it could become an ongoing expense. DeepSeek doesn’t charge per word or per response like some AI services, but via a system of “tokens”, based on computing power used. The more you ask it to do, the more it costs in processing power.
Why does it matter?
The current AI landscape is dominated by a handful of tech giants who control access to the most powerful models. By offering a high-quality alternative that anyone can use and modify, DeepSeek is shaking up the balance of power. It also raises the stakes for AI development globally.
For consumers, this could mean more choice in the long run. Instead of relying on a single AI provider, businesses and developers can tailor DeepSeek to their specific needs.
What are the biggest negatives?
- Lack of user-friendly interfaces: DeepSeek is powerful, but it’s not packaged in a way that makes it accessible to the average user yet.
- Censorship concerns: Certain topics may be restricted or moderated due to the regulatory environment in which DeepSeek was developed.
- Security weaknesses: DeepSeek failed 100% of tests in a Cisco analysis.
What are the biggest positives?
- Open-source flexibility: Unlike proprietary models, DeepSeek allows businesses and developers to modify and integrate it freely.
- Strong multilingual capabilities: Particularly in Chinese, it outperforms many competitors, making it a powerful tool for global applications.
- Challenging the AI status quo: By providing a viable alternative to the dominant players, DeepSeek is pushing the industry towards more competition and innovation.
* Arthur Goldstuck is CEO of World Wide Worx and editor-in-chief of Gadget.co.za. Follow him on Bluesky on @art2gee.bsky.social.