America’s Big tech’s enormous spendings on developing AI models and data centers have come under scrutiny, and the latest developments have sparked concerns about whether America’s leadership in artificial intelligence is eroding.
Silicon Valley is in a panic after a little-known Chinese AI lab, DeepSeek, released AI models that can surpass the best in the United States, despite being constructed more cheaply and with less powerful hardware.
In late December 2024, DeepSeek released a powerful large-language model (LLM) that is free and open-source. According to the lab, it took less than $6 million and two months to develop, utilizing Nvidia’s H800 reducer chips.
DeepSeek’s model beat Meta’s Llama 3.1, OpenAI’s GPT-4o, and Anthropic’s Claude Sonnet 3.5 in a series of independent benchmark tests that measured accuracy in everything from math and coding to sophisticated problem-solving. In several of those third-party tests, DeepSeek’s reasoning model, r1, which was released on Monday, also performed better than OpenAI’s most recent o1.
Credit: AIBC World
“To see the DeepSeek new model, it’s super impressive in terms of both how they have really effectively done an open-source model that does this inference-time compute, and is super-compute efficient,” Microsoft CEO Satya Nadella said at the World Economic Forum in Davos, Switzerland, on Wednesday. “We should take the developments out of China very, very seriously.”
US AI and Semiconductor Regulations
In addition, DeepSeek had to circumvent the stringent semiconductor regulations that the US government had imposed on China. The restrictions prevent China from obtaining the most potent chips, such as Nvidia’s H100s.
Recent developments imply that either DeepSeek managed to circumvent the regulations or that Washington did not intend to strictly enforce the regulations.
According to Benchmark General Partner Chetan Puttagunta, “they can take a really good, big model and use a process called distillation.
“In essence, you utilize a very large model to help your little model become more intelligent at the task you want it to perform. In actuality, that’s quite economical.”
DeepSeek’s Profile
The lab and its creator, Liang WenFeng, are not well known. As to media sources, DeepSeek originated from a Chinese hedge fund named High-Flyer Quant, which oversees over $8 billion in assets.
DeepSeek is not the only Chinese startup that is gaining traction though. Leading AI researcher Kai-Fu Lee claims that under $3 million was used to train his model 01.ai.
On Wednesday, ByteDance, the parent company of TikTok, published an upgrade to its model that is said to beat OpenAI’s o1 in a crucial benchmark test. “Necessity is the mother of invention,” Aravind Srinivas, CEO of Perplexity, remarked.” Because they had to figure out work-arounds, they actually ended up building something a lot more efficient.”
