What you need to know about DeepSeek AI
Companies and organizations like Nvidia, OpenAI, Microsoft, Meta, Google, or Anthropic have dominated AI news in the past year. This dominance is now challenged by Chinese AI startup DeepSeek and its large language models.
What is DeepSeek?
The AI startup was founded by Liang Wenfeng in 2023. It received funding from the Chinese hedge fund High-Flyer, which was founded in 2015. Wenfeng is the co-founder of the hedge fund.
DeepSeek has developed several large language models, which it calls DeepSeek as well. The latest models are DeepSeek V3 and DeepSeek R1.
Why is DeepSeek in the news?
The startup claims that its latest large language model was developed in just two months at a cost of under $6 million. If true, development costs would be a fraction of the costs that current AI frontrunners have to pay to develop new models.
Additionally, DeepSeek V3, its latest large language model, has outperformed several models of US companies in publicly accessible benchmarks.
Chatbot Arena, a ranking website affiliated with UC Berkeley, has two DeepSeek models listed in the top ten. While ChatGPT and Gemini are placed above it in the leaderboard, competitors such as xAI's Grok or Anthropic's Claude have gone done in ranking as a consequence.
The startup's application for Apple devices has overtaken other AI apps in the productivity category on Apple's App Store. On Android, it has claimed a top 3 spot in the productivity category.
Why DeepSeek is challenging the dominance of US AI companies
DeepSeek's large language models appear to cost a lot less than other models. According to a research paper released last month, DeepSeek stated that it spend less than $6 million on the development of the V3 model.
Questions are now raised about the money that companies like OpenAI, Microsoft, or Google are spending on AI model development and data centers in comparison.
DeepSeek Limitations
While DeepSeek's AI model challenge models of competitors in most areas, it is facing other limitations than Western counterparts. One of these is that it ignores any topic that is critical of China according to reports.
DeepSeek requires an account, but the registration process seems to have technical problems at the time of writing. Attempts to sign-up using an email address are met with the "Sorry! We're having trouble sending you a verification code right now. Please try again later." error message.
Closing Words
The coming months will show whether DeepSeek is fueling another technical evolution in AI, one that could reduce the cost factor significantly and speed up development at the same time.
Have you tried DeepSeek already? What is your take on the AI models of the startup? Feel free to leave a comment down below.
The economics are not as impressive as many jump to conclusions about.
It is based on a ‘training cost only’ if they had to rent the similar amount of GPU hours at $2 per hour.
Now that they have proven a point with their work(/research) into this new training approach, cost would similarly drop at all the other big players as they begin to adapt the same methods.
Not that this makes nearly as big headlines. Not that this helps you make a even bigger one-day shorting of nvidia for huge profit and then pick up a profit on the bounce as people begin to look past the headlines and leathercoatman does some damagecontrol explaining: ~17% drop in one day, already a 9% rebound.
I bet one hundred euros that the EU commission is planning new regulations for Deepseek. It’s the only thing that the annoying EU is able to do nowadays. Pure garbage laws good for nothing.
Typical Chinese business: The founder of DeepSeek Liang, who had previously focused on applying AI to investing, had bought a “stockpile of Nvidia A100 chips,” a type of tech that is now banned from export to China. Those chips became the basis of DeepSeek, the MIT publication reported.”
Use banned technology to undermine US ingenuity. Clever. No doubt, those “banned” chips will soon be confiscated or purchased. Or the whole company, DeepSeek, will be a Chinese owned cooperative.
Or the company will become a negotiating tool for the Chinese to eliminate tariffs and Tik-Tok ban.
Liang is his surname.
Sounds like a cool name, another cool names are Liu and Feng.
“One of these is that it ignores any topic that is critical of China according to reports.”
So the hell what? How often do you look for information “critical of China”?
Do you think the US wouldn’t do the exact same thing? Why do you think the former head of the NSA is on OpenAI’s board of directors?
And today DeepSeek gets a DoS attack! Now who do you imagine might be behind that? Hint: It ain’t Russia.
> And today DeepSeek gets a DoS attack! Now who do you imagine might be behind that? Hint: It ain’t Russia.
Yeah. It’s Trump! And president Elongated Muskrat.
DeepSeek has released a new set of multimodal AI models that according to the company can outperform, among others, OpenAI’s DALL-E 3 and Stability AI‘s Stable Diffusion XL on two AI evaluation benchmarks, GenEval and DPG-Bench. The models (Janus-Pro) are available for download and can be used commercially without restriction.
What about the fact that any information gathered goes directly to the Chinese Government. Not to mention the censorship built directly into the bot.
Like western counterparts stealing data and censoring anything they deem politically incorrect.
I’d rather support China, which is why I have a Xiaomi phone. If my data can help them in some way, I would be happy, but I don’t see how data about what animes and games I like will help them.
Yeah I tried it for a couple of weeks. I know nothing about programming or math so I cant say anyting about that. But when an AI does what the regime dictates them to do I dont feel so confident that it wont manipulate other things to deliver chinese propaganda. Some china-critical questions it answers with manipulative answers that sounds right to people who know nothing about about the authotarian regime. Tiamanan square-questions it ignores.
Silicon Valley grifters: Please give us billions and billions and exempt us from copyright law so we can push some unreliable AI crap. Also, let us run our own nuclear power plants to waste energy on this crap.
China: We created unreliable AI crap for just 6 million and open-sourced it. It also uses less power.
I signed-up without any problems.
DeepSeek R1 is the most advanced LLM. It was released on January 21, 2025, about a month after V3’s release on December 27, 2024. According to benchmarks, DeepSeek R1 achieves state-of-the-art performance in math, code, and reasoning tasks, surpassing DeepSeek V3 and competing with top models like OpenAI’s o1.
Nope, can’t sign up. Getting the “sorry” message all the time.
I tried it out and it does the same thing as every other AI model but gives a more detailed replies.
“DeepSeek outage and performance issues keep hitting hard as users flock from ChatGPT”
https://www.neowin.net/news/deepseek-outage-and-performance-issues-keep-hitting-hard-as-users-flock-from-chatgpt/
Development is pretty cheap if you just steal 90%. Although I have yet to see an AI that isn’t piracy.
Pretty much all of this LLM technology is more or less public knowledge (starting with BERT and ending with RAG). So there is no “stealing” of any kind of high-tech technology required to come up with another model, you primarily just need a half-decent math background and the willingness to read a few textbooks on the topic.
The only thing, that is not public knowledge is the training data used, how the training was performed in detail, and the resulting weights and biases of the trained models. The technology, how everything works is public knowledge. All of this stuff is not real rocket science. Not totally trivial, but far, far removed from the true complexity of eg natural scientific research.
Read DeepSeek’s research paper. It is public available.