DeepSeek has made ripples across the tech world since its release earlier this year, but top US tech labs continue to not be very impressed.
Anthropic’s Jack Clark, one of the co-founders of the AI safety and research company, has offered a candid assessment of DeepSeek’s capabilities. Clark believes there is a significant performance gap between DeepSeek and frontier labs like Anthropic, Google, and OpenAI. He also revealed that Anthropic had tested DeepSeek’s technologies in its national security assessments.
“I think the DeepSeek hype is perhaps a bit overblown,” Clark stated. He characterized DeepSeek as “a competent group of engineers and researchers that have read all of the same research papers all the frontier labs have made and have clean-sheeted a new system off of that which has some clever algorithmic ideas.” However, he tempered this by adding, “but if you actually analyze it, it looks 6 to 8 months behind where us frontier companies are.”
Clark further highlighted DeepSeek’s compute limitations. “Additionally,” he noted, “the DeepSeek founder said the main thing limiting his company is compute. So if they had access to arbitrarily large amounts of compute, they might be closer.”
Finally, he offered a revealing glimpse into how these new AI models are being scrutinized. “Just to add one layer of detail,” Clark explained, “we ran the same national security tests on DeepSeek’s technology that we ran on our own, and it was significantly behind our own. Neither our own technology nor DeepSeek’s technology are yet what you might call concerning to national security people. It’s more like they’re symptoms of progress that we should track closely.”
Clark isn’t the only US-based executive that’s downplayed DeepSeek’s models. Anthropic CEO Dario Amodei has said that DeepSeek had more GPUs than was initially believed, and that it had no blocks against generating harmful information. Google DeepMind CEO Demis Hassabis said that that while DeepSeek is impressive, it hasn’t invented anything new.
But while DeepSeek might not have made radically new breakthroughs, what makes it significant is that it comes out of China. Since the beginning of the AI revolution, US labs had believed that they were far ahead of everyone else in the field. DeepSeek’s R1 release in late 2024 shattered that illusion, and showed how capable China was in AI research. DeepSeek’s model wasn’t only extremely capable, but the company was also offering it at a price much lower than many top US labs. And DeepSeek managed to win over consumers as well — it briefly became the most downloaded app on the US app store. Since DeepSeek’s release, even more Chinese companies have come out with impressive models of their own, showing that DeepSeek wasn’t a mere flash in the Chinese pan. And with AI likely to be a major determinant of geopolitical and strategic dominance in the years to come, the DeepSeek-led Chinese progress in AI research might end up being quite consequential in the grand scheme of things.