Google is the latest to praise — and criticize — DeepSeek
During Alphabet’s earnings call last week, CEO Sundar Pichai was mostly effusive about DeepSeek, the Chinese company whose AI model has upended much of what American AI firms thought was possible for the price.
“I think [they are] a tremendous team. I think they’ve done very, very good work,” Pichai said, before touting Google’s own bona fides.
On Bloomberg today, Google DeepMind leader Demis Hassabis was a little more cutting, saying the company might have underestimated its costs and exaggerated its innovation. Here’s a slightly trimmed-down transcript:
“It’s a very impressive model, a very impressive piece of work. I think the team is probably the best team that I’ve seen come out of China. That said, I think a lot of the claims are exaggerated and a little bit misleading.
First of all, when you report how much it costs to do a training run, they seem to have reported just their final training run, which is only a fraction of what it costs to explore and train and do all the tests before you do your final run.
They seem to have relied on some Western models to distill from or to basically fine-tune against the outputs of.
Finally, it’s an impressive piece of work but we don’t see any silver-bullet new technologies, techniques that we haven’t seen before or haven’t invented before. They’ve just applied it very well.
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. We just don’t talk about that very much.”
The leaders of Nvidia, OpenAI, Microsoft, and Tesla have followed a similar playbook when commenting on the Chinese AI company.
On Bloomberg today, Google DeepMind leader Demis Hassabis was a little more cutting, saying the company might have underestimated its costs and exaggerated its innovation. Here’s a slightly trimmed-down transcript:
“It’s a very impressive model, a very impressive piece of work. I think the team is probably the best team that I’ve seen come out of China. That said, I think a lot of the claims are exaggerated and a little bit misleading.
First of all, when you report how much it costs to do a training run, they seem to have reported just their final training run, which is only a fraction of what it costs to explore and train and do all the tests before you do your final run.
They seem to have relied on some Western models to distill from or to basically fine-tune against the outputs of.
Finally, it’s an impressive piece of work but we don’t see any silver-bullet new technologies, techniques that we haven’t seen before or haven’t invented before. They’ve just applied it very well.
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. We just don’t talk about that very much.”
The leaders of Nvidia, OpenAI, Microsoft, and Tesla have followed a similar playbook when commenting on the Chinese AI company.
.gif)