The Chinese AI laboratory DeepSeek triggered a significant reaction in Silicon Valley as it unveiled open-access AI models that rival the cutting-edge technologies developed by OpenAI, Meta, and Google in early 2025.
DeepSeek asserts that it has developed its models rapidly and efficiently (though some industry experts harbor doubts about these assertions) and offers them at a fraction of the costs charged by American AI firms. This development has alarmed not just technology giants but also high-ranking officials in the U.S. government, who are concerned that China may be gaining an advantage in the AI arms race.
“I wouldn’t be surprised if many AI labs have activated crisis strategies at this point,” remarked Robert Nishihara, co-founder of AI infrastructure company Anyscale, in a conversation with TechCrunch.
The emergence of DeepSeek signifies a pivotal moment in the landscape of AI in Silicon Valley. AI CEOs, founders, researchers, and investors have conveyed to TechCrunch that the implications of DeepSeek’s models could be substantial for U.S. AI policy. Furthermore, these specialists note that the models indicate the speeding pace of advancements in AI technology.
“Certainly, [DeepSeek] received excessive hype,” commented Ravid Shwartz-Ziv, an assistant professor at NYU’s Center for Data Science, during an interview. “However, it is quite fascinating, and there is much to learn from it.”
Innovative Approaches to AI Learning
A fundamental advancement in DeepSeek’s R1 model is “pure reinforcement learning,” which employs a trial-and-error methodology, as noted by Kian Katanforoosh, CEO of Workera and adjunct lecturer at Stanford.
Katanforoosh likened DeepSeek’s progress to a child who learns not to touch a hot stove after getting burned.
“When a child touches a hot stove, they get burned and quickly learn not to repeat the action,” Katanforoosh stated in a text message. “This is pure reinforcement learning—gaining knowledge from trial and error based on feedback. DeepSeek’s approach centers around allowing the model to learn solely from experience.”
DeepSeek appears to have utilized reinforcement learning more extensively than other leading AI models. OpenAI also incorporated reinforcement learning techniques to create o1, which the company introduced a few weeks prior to DeepSeek’s announcement of R1. OpenAI’s forthcoming o3 model is claimed to offer even improved performance through largely similar yet more computational resources.
Reinforcement learning is regarded as one of the most promising strategies to enhance AI foundational models today, according to Katanforoosh. Typically, “foundation models” refer to AI models trained on vast datasets, such as images and text obtained from the web. It appears likely that other AI laboratories will continue exploring the boundaries of reinforcement learning to advance their models following DeepSeek’s success.
Just months ago, AI firms were grappling with challenges in enhancing the performance of their foundational models. However, the effectiveness of methods such as reinforcement learning, along with others like supervised fine-tuning and test-time scaling, suggests that AI advancement may be on the rise again.
“R1 has given me renewed confidence in the sustained pace of progress,” said Nathan Lambert, a researcher at Ai2, during a conversation with TechCrunch.
A Defining Moment for AI Policy
R1, which can be downloaded and executed on any compatible hardware, performs similarly or even superiorly to o1 across various AI benchmarks. Although this isn’t the first time we have seen a narrowing performance gap between “closed” models like those from OpenAI and openly accessible alternatives, the swift manner in which DeepSeek accomplished this has left the industry astonished.
This development may drive the U.S. to bolster its investment in either open or fully open-source AI to maintain competitiveness with China. Martin Casado, a general partner at Andreessen Horowitz (a16z), informed TechCrunch that DeepSeek highlights just how misguided the regulatory approaches of the past two years have been.
“This makes it clear that [the United States] is not the sole entity with technical capabilities,” Casado stated in an interview. “Highly competitive solutions can arise from various locations, particularly China. Instead of stifling U.S. innovation, we should enhance our investments in it. Open-source initiatives do not facilitate China’s advancements; rather, restricting our companies from pursuing open-source means our technology does not proliferate as effectively.”
Casado seemed to refer to the recently repealed AI executive order from former President Biden and the vetoed California bill SB 1047, both of which a16z actively contested. The firm has argued that these measures prioritized curtailing “outlandish” AI dystopian scenarios over promoting American innovation. In broader terms, Silicon Valley has managed to undermine the “AI doom movement” throughout 2024. The true risk concerning AI, a16z and others have repeatedly emphasized, is the potential for America to lose its competitive position relative to China.
This prospect seems increasingly plausible in light of DeepSeek’s emergence.
It’s worth mentioning that a16z has significant investments in many of the largest players in the open AI sector, such as Databricks, Mistral, and Black Forest Labs. The venture capital firm may also have a substantial impact on advising the Trump administration regarding AI matters, with former a16z partner Sriram Krishnan now serving as Trump’s senior policy advisor for AI.
President Trump declared on Monday that DeepSeek serves as a “wake-up call” for U.S. AI firms and commended the Chinese lab for its open methodology. This aligns closely with a16z’s perspective on AI.
“DeepSeek R1 represents AI’s Sputnik moment,” stated a16z co-founder Marc Andreessen in a post on X, drawing a parallel to the launch of the Soviet Union’s satellite that prompted the U.S. to invest earnestly in its space initiatives.
The rise of DeepSeek also seems to have influenced the views of open AI detractors, including former Google CEO Eric Schmidt. Only a year ago, Schmidt expressed trepidation about the spread of Western open AI models worldwide. However, in an op-ed published Tuesday, he remarked that DeepSeek’s ascent signifies a “turning point” in the global AI race, advocating for increased investment in American open AI initiatives.
Future Perspectives
It is crucial to avoid overstating the achievements of DeepSeek.
For instance, some analysts question DeepSeek’s assertion that it trained one of its leading models, DeepSeek V3, for just $5.6 million—a minimal sum in the AI sector—using approximately 2,000 older Nvidia GPUs. The lab did not emerge overnight, and it reportedly has a stockpile of over 50,000 more advanced Nvidia Hopper GPUs.
Moreover, DeepSeek’s models carry inaccuracies. A test conducted by the information-veracity organization NewsGuard revealed that R1 produced incorrect or incomplete answers 83% of the time when queried about news-related content. A separate evaluation indicated that R1 declined to answer 85% of queries related to China, potentially due to the government censorship imposed on AI models in the nation.
Additionally, there are allegations of intellectual property theft. OpenAI claims to possess evidence that DeepSeek trained its models using data from its own AI models through a process known as distillation. If true, this would breach OpenAI’s policies and lessen the significance of DeepSeek’s achievements. For instance, researchers at Berkeley recently created a distilled reasoning model for merely $450. (It is worth noting that OpenAI is currently facing lawsuits from several parties for alleged copyright infringement concerning its training methods.)
Nevertheless, DeepSeek has made strides with more efficient models and demonstrated innovation. Lambert pointed out that, unlike o1, R1 allows users to view its “reasoning process.” He noted that some users tend to trust or ascribe greater legitimacy to AI reasoning models when they are privy to their internal operations, during which the models “explain their rationale.”
Going forward, it remains to be seen how U.S. policymakers and AI laboratories will respond to these developments.
TechCrunch offers a newsletter focused on AI! Subscribe here to receive it in your inbox every Wednesday.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
