How Disruptive Is DeepSeek? Stanford HAI Faculty Discuss China’s New Model

In recent weeks, the emergence of China’s DeepSeek — a powerful and cost-efficient open-source language model — has stirred considerable discourse among scholars and industry researchers. At the Stanford Institute for Human-Centered AI (HAI), faculty are examining not merely the model’s technical advances but also the broader implications for academia, industry, and society globally. 

Central to the conversation is how DeepSeek has challenged the preconceived notions regarding the capital and computational resources necessary for serious advancements in AI. The capacity for clever engineering and algorithmic innovation demonstrated by DeepSeek may empower less-resourced organizations to compete on meaningful projects. This clever engineering, combined with the open-source weights and a detailed technical paper, fosters an environment of innovation that has driven technical advances for decades.  

While the open weight model and detailed technical paper is a step forward for the open-source community, DeepSeek is noticeably opaque when it comes to privacy protection, data-sourcing, and copyright, adding to concerns about AI's impact on the arts, regulation, and national security. The fact that DeepSeek was released by a Chinese organization emphasizes the need to think strategically about regulatory measures and geopolitical implications within a global AI ecosystem where not all players have the same norms and where mechanisms like export controls do not have the same impact. 

DeepSeek has reignited discussions of open source, legal liability, geopolitical power shifts, privacy concerns, and more. In this collection of perspectives, Stanford HAI senior fellows offer a multidisciplinary discussion of what DeepSeek means for the field of artificial intelligence and society at large.

Russ Altman

Kenneth Fong Professor and Professor of Bioengineering, of Genetics, of Medicine, of Biomedical Data Science, Stanford HAI Senior Fellow, and Professor, by courtesy, of Computer Science

We at HAI are academics, and there are elements of the DeepSeek development that provide important lessons and opportunities for the academic community.  

First, the commitment to open source (embraced by Meta and also adopted by DeepSeek) seems to transcend geopolitical boundaries — both DeepSeek and Llama (from Meta) provide an opportunity for academics to inspect, assess, evaluate, and improve on existing methods, from an independent perspective. The “closed source” movement now has some challenges in justifying the approach—of course there continue to be legitimate concerns (e.g., bad actors using open-source models to do bad things), but even these are arguably best combated with open access to the tools these actors are using so that folks in academia, industry, and government can collaborate and innovate in ways to mitigate their risks.

Second, the demonstration that clever engineering and algorithmic innovation can bring down the capital requirements for serious AI systems means that less well-capitalized efforts in academia (and elsewhere) may be able to compete and contribute in some types of system building. Many of us thought that we would have to wait until the next generation of inexpensive AI hardware to democratize AI — this may still be the case. But even before that, we have the unexpected demonstration that software innovations can also be important sources of efficiency and reduced cost. Taken together, we can now imagine non-trivial and relevant real-world AI systems built by organizations with more modest resources. 

Third, the progress of DeepSeek coupled with advances in agent-based AI systems makes it easier to imagine the widespread creation of specialized AI agents that are mixed and matched to create capable AI systems. The monolithic “general AI” may still be of academic interest, but it will be more cost-effective and better engineering (e.g., modular) to create systems made of components that can be built, tested, maintained, and deployed before merging. A model of AI agents cooperating with one another (and with humans) replicates the idea of human “teams” that solve problems. Sometimes problems are solved by a single monolithic genius, but this is usually not the right bet. Thus, DeepSeek helps restore balance by validating open-source sharing of ideas (data is another matter, admittedly), demonstrating the power of continued algorithmic innovation, and enabling the economic creation of AI agents that can be mixed and matched economically to produce useful and robust AI systems. Of course, questions remain:

• How can we democratize the access to huge amounts of data required to build models, while respecting copyright and other intellectual property?

• How do we build specialized models when the volume of data for some specialized disciplines is not sufficiently large?

• How do we evaluate a system that uses more than one AI agent to ensure that it functions correctly? Even if the individual agents are validated, does that mean they are validated in combination?

Yejin Choi

Dieter Schwarz Foundation HAI Professor, Professor of Computer Science, and Stanford HAI Senior Fellow

The success of DeepSeek's R1 model shows that when there’s a “proof of existence of a solution” (as demonstrated by OpenAI’s o1), it becomes merely a matter of time before others find the solution as well. DeepSeek’s decision to share the detailed recipe of R1 training and open weight models of varying size has profound implications, as this will likely escalate the speed of progress even further — we are about to witness a proliferation of new open-source efforts replicating and enhancing R1. This shift signals that the era of brute-force scale is coming to an end, giving way to a new phase focused on algorithmic innovations to continue scaling through data synthesis, new learning frameworks, and new inference algorithms.

However, a major question we face right now is how to harness these powerful artificial intelligence systems to benefit humanity at large. The fact that a model excels at math benchmarks does not immediately translate to solutions for the hard challenges humanity struggles with, including escalating political tensions, natural disasters, or the persistent spread of misinformation. This disconnect between technical capabilities and practical societal impact remains one of the field’s most pressing challenges.

Michele Elam

William Robertson Coe Professor in the Humanities, Stanford HAI Senior Fellow, Bass University Fellow in Undergraduate Education

Amidst all the U.S. handwringing and twisted knickers over the recent Chinese drop of the apparently (wildly) less expensive, less compute-hungry, less environmentally insulting DeepSeek AI chatbot, to date few have considered what this means for AI’s impact on the arts. In fact, what DeepSeek means for literature, the performing arts, visual culture, etc., can seem utterly irrelevant in the face of what may appear like much higher-order anxieties regarding national security, economic devaluation of the U.S. AI industry, and the benefits or not of open source for innovation.

But, actually, DeepSeek’s total opacity when it comes to privacy protection, data sourcing and scraping, and NIL and copyright debates has an outsized impact on the arts. Actually, “opacity” is a generous term: DeepSeek is a “can’t-even-be-bothered” response to these concerns. Never mind the SAG-AFTRA strikes in the creative industry, the ongoing lawsuits by The New York Times, and many others.

In many ways, the fact that DeepSeek can get away with its blatantly shoulder-shrugging approach is our fault. The very popularity of its chatbot is an amplified reflection of — and capitalization on — American consumers’ own increasing tendency to turn a blind eye to these issues, a tendency aggressively encouraged by an industry whose business models intentionally turn our attention from such unpleasantries in the name of return-on-investment.

Like TikTok, DeepSeek leverages the creep of our acculturation over the last several years to giving away our privacy rights with each click of the ever-updated ever-more obscure terms of contract on our devices (usually in the name of that marvelous marketing euphemism, “personalization”).

Arguably, as many have already noted, DeepSeek’s omnivorous consumption of private and sensitive data exploits the national failure to have any regulation of AI, unlike the U.K. and the E.U., and puts the country at risk in many ways because of our mantra that “regulation impedes innovation.”

But as it relates to the arts, we would be well-served to pay attention to the way DeepSeek controls the keys to our imagination through its preemptive censorship, its alignment with nationalist ideologies, our unknowing or unthinking consent to its algorithmic modeling of reality – that is, its ability to shape how we see and act in the world. Stanford has currently adapted, via Microsoft’s Azure program, a “safer” version of DeepSeek with which to experiment and warns the community not to use the commercial versions because of safety and security concerns. But, regardless, the release of DeepSeek highlights the risks and rewards of this technology’s outsized ability to influence our experience of reality in particular — what we even come to think of as reality. As the early debates between Plato and Aristotle about the influential civic power of the theatre and poetry signaled, that is also precisely the power of the arts. 

Mykel Kochenderfer

Associate Professor of Aeronautics and Astronautics at Stanford University, Stanford HAI Senior Fellow

AI is increasingly being used to support safety-critical or high-stakes scenarios, ranging from automated vehicles to clinical decision support. However, reconciling the lack of explainability in current AI systems with the safety engineering standards in high-stakes applications remains a challenge. A particularly compelling aspect of DeepSeek R1 is its apparent transparency in reasoning when responding to complex queries. The level of detail it provides can facilitate auditing and help foster trust in what it generates. This transparent reasoning at the time a question is asked of a language model is referred to as interference-time explainability. While inference-time explainability in language models is still in its infancy and will require significant development to reach maturity, the baby steps we see today may help lead to future systems that safely and reliably assist humans.

Another barrier in applying recent advances in artificial intelligence to many applications is the huge amounts of data and compute required. DeepSeek demonstrates that there is still enormous potential for developing new methods that reduce reliance on both large datasets and heavy computational resources. I hope that academia — in collaboration with industry — can help accelerate these innovations. By creating more efficient algorithms, we can make language models more accessible on edge devices, eliminating the need for a continuous connection to high-cost infrastructure. With the tremendous amount of common-sense knowledge that can be embedded in these language models, we can develop applications that are smarter, more helpful, and more resilient — especially important when the stakes are highest.

James Landay

Professor of Computer Science and the Anand Rajaraman and Venky Harinarayan Professor in the School of Engineering at Stanford University, Stanford HAI Co-Director

DeepSeek is a good thing for the field. They are publishing their work. Their model is released with open weights, which means others can modify it and also run it on their own servers. They are bringing the costs of AI down. This is all good for moving AI research and application forward. One of the biggest critiques of AI has been the sustainability impacts of training large foundation models and serving the queries/inferences from these models. DeepSeek has shown many useful optimizations that reduce the costs in terms of computation on both of these sides of the AI sustainability equation. This is good for the field as every other company or researcher can use the same optimizations (they are both documented in a technical report and the code is open sourced). 

“The practice of sharing innovations through technical reports and open-source code continues the tradition of open research that has been essential to driving computing forward for the past 40 years.”

The practice of sharing innovations through technical reports and open-source code continues the tradition of open research that has been essential to driving computing forward for the past 40 years. As a research field, we should welcome this type of work. It will help make everyone’s work better. While many U.S. companies have leaned toward proprietary models and questions remain, especially around data privacy and security, DeepSeek’s open approach fosters broader engagement benefiting the global AI community, fostering iteration, progress, and innovation.

Percy Liang

Associate Professor of Computer Science at Stanford University, Director of the Center for Research on Foundation Models (CRFM), Stanford HAI Senior Fellow

DeepSeek R1 showed that advanced AI will be broadly available to everyone and will be difficult to control, and also that there are no national borders. It also shows that ingenuity and engineering do matter, in addition to having large amounts of compute. For academia, the availability of more strong open-weight models is a boon because it allows for reproducibility, privacy, and allows the study of the internals of advanced AI.

Christopher Manning

The Thomas M. Siebel Professor in Machine Learning in the Departments of Linguistics and Computer Science at Stanford University, and Stanford HAI Associate Director

People treated this as some kind of out-of-the-blue surprise, but it really wasn’t if you were actively following open-source AI. DeepSeek has been publicly releasing open models and detailed technical research papers for over a year. The cost of training DeepSeek V3 came out in December 2024; an R1-Lite-Preview release came out in November 2024.

“It’s a sad state of affairs for what has long been an open country advancing open science and engineering that the best way to learn about the details of modern LLM design and engineering is currently to read the thorough technical reports of Chinese companies.”

This release underlines that the U.S. so-called “frontier” AI companies do not have some huge technical moat. There are now many excellent Chinese large language models (LLMs). At most these companies are six months ahead, and maybe it’s only OpenAI that is ahead at all. It’s a sad state of affairs for what has long been an open country advancing open science and engineering that the best way to learn about the details of modern LLM design and engineering is currently to read the thorough technical reports of Chinese companies. 

DeepSeek has done some very good data engineering, minimizing data flow and allowing efficient and stable training in fp8. They have some modest technical advances, using a distinctive form of multi-head latent attention, a large number of experts in a mixture-of-experts, and their own simple, efficient form of reinforcement learning (RL), which goes against some people’s thinking in preferring rule-based rewards. But there’s nothing totally next generation here. DeepSeek uses similar methods and models to others, and Deepseek-R1 is a breakthrough in nimbly catching up to provide something similar in quality to OpenAI o1. It’s not a new breakthrough in capabilities.

The DeepSeek-R1 release does noticeably advance the frontier of open-source LLMs, however, and suggests the impossibility of the U.S. being able to contain the development of powerful open-source LLMs. It may well also mean that more U.S. companies will start using Chinese LLMs within their own products, whereas until now they have generally avoided them, preferring to use Meta’s Llama models or others from Databricks, etc.

Hear more from Prof. Manning on DeepSeek in this talk with AIX Ventures.

Julian Nyarko

Professor of Law at Stanford Law School, Stanford HAI Associate Director

LLMs are a “general purpose technology” used in many fields. Some companies create these models, while others use them for specific purposes. A key debate right now is who should be liable for harmful model behavior—the developers who build the models or the organizations that use them. In this context, DeepSeek’s new models, developed by a Chinese startup, highlight how the global nature of AI development could complicate regulatory responses, especially when different countries have distinct legal norms and cultural understandings.  While export controls have been thought of as an important tool to ensure that leading AI implementations adhere to our laws and value systems, the success of DeepSeek underscores the limitations of such measures when competing nations can develop and release state-of-the-art models (somewhat) independently. The open-source nature of DeepDeek’s releases further complicates the question of legal liability. With the models freely available for modification and deployment, the idea that model developers can and will effectively address the risks posed by their models could become increasingly unrealistic. Instead, regulatory focus may need to shift towards the downstream consequences of model use — potentially placing more responsibility on those who deploy the models. 

Amy Zegart

Morris Arnold and Nona Jean Cox Senior Fellow at the Hoover Institution, Senior Fellow at the Freeman Spogli Institute for International Studies, at Stanford HAI, and Professor, by courtesy, of Political Science

The past few weeks of DeepSeek deep freak have focused on chips and moats. How much did DeepSeek stockpile, smuggle, or innovate its way around U.S. export controls? How many and what kind of chips are needed for researchers to innovate on the frontier now, in light of DeepSeek’s advances? Did U.S. hyperscalers like OpenAI end up spending billions building competitive moats or a Maginot line that merely gave the illusion of security? These are all important questions, and the answers will take time.

“Nearly all of the 200 engineers authoring the breakthrough R1 paper last month were educated at Chinese universities, and about half have studied and worked nowhere else. This should be a red flag for U.S. policymakers.”

However, three serious geopolitical implications are already apparent. First, DeepSeek succeeded with homegrown talent. Nearly all of the 200 engineers authoring the breakthrough R1 paper last month were educated at Chinese universities, and about half have studied and worked nowhere else. This should be a red flag for U.S. policymakers. In the tech era, talent is a major source of national power.  The mantra “the U.S. attracts the world’s best talent” is frequently uttered but it’s increasingly wrong. Rising educational levels and dramatic improvements in higher education institutions in China and elsewhere around the world are redrawing the knowledge power map. Meanwhile America’s K-12 education is in shambles, with U.S. 15-year-olds scoring a dismal 34th in math during the last international test — behind Slovenia and Vietnam.

Second, DeepSeek did not copy U.S. companies. It copied U.S. universities. The startup hired young engineers, not experienced industry hands, and gave them freedom and resources to do “mad science” aimed at long-term discovery for its own sake, not product development for next quarter. Commercialization is an essential part of innovation. But breakthroughs often begin with fundamental research that has no foreseeable product or profit in mind. This kind of fundamental research is the lifeblood of universities, and it has underpinned U.S. innovation leadership for decades — giving rise to everything from cube satellites to COVID-19 vaccines. Yet today, China is investing six times faster in fundamental research than the U.S. government and, if current trends continue, China will out-invest the U.S. within a decade. This is a crucial long-term innovation battleground, and the U.S. is ceding it.

Third, DeepSeek’s announcement roiled U.S. markets, leading to a 3% decline in the NASDAQ composite and a 17% decline in NVIDIA shares, erasing $600 billion in value. It was the largest single-day loss of a company in U.S. history and a figure so massive it’s equivalent to 65% of the annual U.S. defense budget. This unintended consequence today could be the intended consequence of tomorrow. Imagine an adversary deliberately announces a real or fraudulent technological advance to punish a specific company or rattle the capital markets of another nation. This gray zone economic weapon could be precisely targeted or massive. It could be difficult, perhaps impossible, to attribute as a deliberate activity. And it works best if it comes without warning.

Far-fetched? The past decade has seen the rise of stunning gray zone activity in other domains, from Russia’s “little green men” marching into Crimea without uniforms claiming to be local Ukrainian self-defense units to cyber-enabled influence operations conducted by Russia, China, Iran, and other nations to inflame public opinion and shape elections around the world.

DeepSeek didn’t just release new AI advances; it revealed the contours of a burgeoning geopolitical era that has new sources of national power and new battlegrounds.