DeepSeek claims to have built its models highly efficiently and quickly (though some are skeptical of these claims), and is providing these models at a fraction of the price American AI companies charge. The development has rattled not only tech giants but the highest levels of the U.S. government, which fear that China is pulling ahead in the AI arms race.
“I wouldn’t be surprised if a lot of AI labs have war rooms going on right now,” said Robert Nishihara, the co-founder of AI infrastructure startup Anyscale, in an interview with TechCrunch.
The rise of DeepSeek marks an inflection point for Silicon Valley’s AI landscape. AI CEOs, founders, researchers, and investors tell TechCrunch that DeepSeek’s models have major implications for American AI policy. Moreover, these experts say, the models serve as an indicator of the accelerating rate of AI progress.
“Of course [DeepSeek] was over-hyped,” said Ravid Shwartz-Ziv, an assistant professor at NYU’s Center for Data Science, in an interview. “But it’s still very interesting, and there’s a lot we can take from it.”
New ways to get AI thinking
One of DeepSeek’s key innovations in creating its R1 model was “pure reinforcement learning,” a trial-and-error approach, according to Workera CEO and Stanford adjunct lecturer Kian Katanforoosh.
Katanforoosh compared DeepSeek’s breakthrough to a kid figuring out not to touch a hot plate by accidentally burning themselves.
“[A kid] might touch a hot plate, get burned, and quickly learn not to do it again,” Katanforoosh said via text. “That’s pure reinforcement learning — learning from trial and error based on feedback […] DeepSeek’s method is all about letting the model learn through experience alone.”
DeepSeek seems to have relied more heavily on reinforcement learning than other cutting edge AI models. OpenAI also used reinforcement learning techniques to develop o1, which the company revealed weeks before DeepSeek announced R1. OpenAI’s upcoming o3 model achieves even better performance using largely similar methods, but also additional compute, the company claims.
Reinforcement learning represents one of the most promising ways to improve AI foundation models today, according to Katanforoosh. The term “foundation models” generally refers to AI models trained on massive amounts of data, like images and text from the web. It seems likely that other AI labs will continue to push the limits of reinforcement learning to improve their AI models, especially given the success of DeepSeek.
Just a few months ago, AI companies found themselves struggling to boost the performance of their foundation models. But the success of methods such as reinforcement learning and others, like supervised fine-tuning and test-time scaling, indicate that AI progress may be picking back up.
“R1 has given me a lot more confidence in the pace of progress staying high,” said Nathan Lambert, a researcher at Ai2, in an interview with TechCrunch.
A turning point for AI policy
R1, which can be downloaded and run on any machine that meets the hardware requirements, matches or beats o1 on a number of AI benchmarks. While it’s not the first time we’ve seen the performance gap narrow between “closed” models like that of OpenAI and openly available models, the speed with which DeepSeek did it has taken the industry aback.
This may push the U.S. to increase its investment in open, or even fully open source, AI in order to compete with China. Martin Casado, a general partner at Andreessen Horowitz (a16z), tells TechCrunch that DeepSeek proves just how “wrongheaded” the regulatory rationale of the last two years has been.
“For AI, I think this just shows us that [the United States] is not alone in our technical capability,” Casado said in an interview. “Very competitive solutions can come from anywhere, but in particular, China. Rather than hampering U.S. innovation, we should invest strongly in it. Open source does not in some way enable China. In fact, disallowing our companies from doing open source means that our technology doesn’t proliferate as much.”
Casado seemed to be referring to former President Biden’s recently-repealed AI Executive Order and the vetoed California bill SB 1047, both of which a16z aggressively opposed. A16z has argued both measures prioritized preventing “outlandish” AI doomsday scenarios over American innovation. More broadly, Silicon Valley generally had success tamping down the “AI doom movement” in 2024. The real concern around AI, a16z and others have repeatedly said, is America losing its competitive edge to China.
That scenario seems much more tangible in light of DeepSeek’s rise.
Not for nothing, a16z is heavily invested in many of the open AI world’s largest players, including Databricks, Mistral, and Black Forest Labs. The VC firm may also play an outsized role advising the Trump Administration on AI. Former a16z partner Sriram Krishnan is now Trump’s senior policy advisor for AI.
President Trump said on Monday that DeepSeek should be a “wakeup call” for American AI companies, while praising the Chinese AI lab for its open approach. That lines up pretty closely with a16z’s stance on AI.
“DeepSeek R1 is AI’s Sputnik moment,” said a16z co-founder Marc Andreessen in a post on X, referencing the launch of the Soviet Union’s Earth-orbiting spacecraft decades ago that pushed the U.S. to seriously invest in its space program.
The rise of DeepSeek also appears to have changed the mind of open AI skeptics, like former Google CEO Eric Schmidt. Just last year, Schmidt expressed concern about the proliferation of Western open AI models around the globe. But in an op-ed published Tuesday, Schmidt said DeepSeek’s rise marks a “turning point” in the global AI race, and called for further investment in American open AI.
