It’s hard to keep up with all the AI breakthroughs, but one story that continues to resonate days later is the arrival of DeepSeek. Not only is DeepSeek’s R1 model as powerful as OpenAI’s o1, but the Chinese startup behind the model claims it spent just $5.6 million to train DeepSeek—a fraction of the more than $100 million OpenAI spent to train GPT-4.
DeepSeek sent a shockwave through the AI industry, but speaking to Bloomberg at the Artificial Intelligence Action Summit in Paris on Monday, Google DeepMind CEO and co-founder Demis Hassabis poured cold water on the audacious claims made by DeepSeek.
“It’s a very impressive model, very impressive piece of work, and I think the team is probably the best team that I’ve seen come out of China,” Hassabis said. “That said, I think a lot of the claims are exaggerated and a little bit misleading.”
He went on to explain that he believes the low cost they shared that made headlines is only the cost of the “final training run,” which is only a fraction of the full cost of training an AI model from start to finish. He also claims that DeepSeek “relied on Western models to distill from,” which is what OpenAI suggested in the days following DeepSeek’s launch.
“We know PRC-based companies — and others — are constantly trying to distill the models of leading US AI companies,” OpenAI told Bloomberg at the time.
Finally, Hassabis noted that as impressive as DeepSeek is, Google doesn’t see the model as a “silver bullet” in terms of new technologies or techniques.
“So it’s impressive, but it isn’t some new outlier on the efficiency curve. For example, Gemini is more efficient than DeepSeek in terms of its training to performance or its cost to performance,” Hassabis proclaimed. “We just don’t talk about that very much.”
You can see Hassabis’s full comments from the AI summit in the video below:
