With the release of the new Chinese open-source AI model DeepSeek, there has been much speculation on the possible implications of the model - both in Israel and around the globe.
“A chatbot owned by a Chinese entity - that’s a concern. We don’t know where the data is going or who’s looking at it,” Moshe Zalcberg, CEO of Veriest, told The Jerusalem Post.
Veriest, part of IT specialist Aman Group, specializes in the design of ASIC chips and provides a complete set of professional engineering services. The company's engineering teams in Israel, Serbia, Hungary, and the UK number over 150 engineers.
Commenting on the implications of the release of DeepSeek for the Israeli AI industry, Zalcberg added, “If I forget for a moment that they are in China, the very fact that a company was able to do more with less is good news.”
“DeepSeek did what we Israelis like to do, which is to come up with a better solution to a problem.”
“There are lots of Israeli companies trying to do the same thing. To do things in a different way, in a smarter and cheaper way with a lower energy footprint and less electricity. From an industry perspective, I think it’s good news,” he said.
DeepSeek claimed that their latest model can do the same as other chatbots like ChatGPT, but in a much more time and cost-efficient way. They have created their model with just $5.5 million, according to the company. Comparatively, Meta spent around 10 times as much to develop their own AI model.
The first variant of DeepSeek, DeepSeek-V3, was released on December 26, and the second, DeepSeek-R1, on January 20.
They used Nvidia’s H800 chips, similar to H100 chips but manufactured specifically for the Chinese market to comply with US regulations introduced by Biden in 2022. The US had previously placed restrictions on the export of advanced AI chips to China in an effort to maintain American AI dominance.
Concerns surrounding DeepSeek
One concern for the US, in particular, is that DeepSeek may have managed to circumvent American limitations and regulations, designed to maintain US supremacy in the field, and achieved a similar level of quality to other American models - including ChatGPT.
OpenAI, the owner of ChatGPT, believes that DeepSeek has used a ‘distillation’ technique to train their model, in which developers use a shortcut by taking the output of pre-existing AI models in order to train new ones. OpenAI claims that their models were used by DeepSeek to train their own chatbot, which would violate OpenAI’s terms of service.