GuShiio.com Wen| technological whirlpool
In the past few weeks, discussions on DeepSeek have been extremely lively in the U.S. technology community, focusing on chip supply and technical barriers. People have speculated about how many chips DeepSeek has hoarded and what methods it has used to circumvent U.S. export controls. These discussions not only reflect the United States ‘concerns about its own technological leadership, but also reveal the strategic limitations of its long-standing reliance on closed technology and export controls.
Multiple faculty members from Stanford University’s People-centered Artificial Intelligence Institute (HAI) came together to discuss the technological advancement of publishing the DeepSeek model, and also provided an in-depth analysis of its broader impact on global academia, industry and society.
At the heart of the discussion is how DeepSeek challenges traditional concepts about the funding and computing resources needed to achieve major advances in AI. The smart engineering and algorithmic innovations demonstrated by DeepSeek show that even organizations with fewer resources can compete on meaningful projects. This clever design, coupled with open source weights and technical detail papers, has created an open innovation environment that has promoted technological progress for decades.
DeepSeek has rekindled discussions about open source, legal liability, geopolitical power shifts, privacy issues and more. In this set of perspectives, senior researchers at Stanford HAI provide a multidisciplinary perspective on the significance of DeepSeek for the field of artificial intelligence and society as a whole. It is worth noting that several professors pointed out that the United States ‘long-standing strategy of relying on closed technology and export controls is facing unprecedented challenges. DeepSeek’s success not only demonstrates the strength of China companies in technological innovation, but also reveals the vulnerability and anxiety of the American technology community.
Russ Altman and other professors discussed the opportunities and challenges brought by DeepSeek from different perspectives, including the importance of open source sharing concepts, the value of algorithm innovation, and how to balance copyright and other intellectual property issues in data acquisition. In addition, Professor Michele Elam highlighted the profound impact of DeepSeek’s ambiguity in privacy protection, data sources and copyright on the art field, and called attention to the technology’s role in shaping the way we perceive the real world. These discussions reveal the disconnect between technological progress and the actual social impact, one of the most pressing challenges in the field today.
Russ Altman’s comments
Russ Altman is Professor Kenneth Fong, Professor of Bioengineering, Professor of Genetics, Professor of Medicine, Professor of Biomedical Data Science, and is also a Senior Fellow and Professor of Computer Science (concurrently) at Stanford HAI
First, the commitment to open source (adopted by both Meta and DeepSeek) seems to transcend geopolitical boundaries and DeepSeek and Llama(from Meta) provide scholars with an opportunity to independently examine, evaluate, and improve existing methods. ldquo; The closed-source movement faces some challenges in justifying it. Of course, there are still legitimate concerns (e.g., that bad actors use open source models to do bad things), but these concerns are best addressed through open access to the tools used by these actors so that academia, industry, and government can collaborate to innovate and mitigate risks.
Second, smart engineering and algorithmic innovation can reduce the capital requirements of serious AI systems, which means that less-funded academic efforts (and other areas) can compete and contribute in building certain systems. Many people believe that we need to wait until the next generation of cheap AI hardware to democratize AI, and this may still be true. But before that, we had unexpectedly discovered that software innovation can also be an important source of efficiency improvements and cost reductions. Taken together, we can now imagine non-trivial and relevant real-world AI systems built by organizations with relatively limited resources.
Third, DeepSeek’s advances combined with advancements in agent-based AI systems make it easier to imagine creating specialized AI agents and mixing them to create capable AI systems. Monolithic generic AI may still be of academic interest, but from a cost-effective and better engineering perspective (e.g. modularity), it is more appropriate to create a system composed of components that can be built, tested, maintained, and deployed separately. The model of cooperation among AI agents and with humans replicates the concept of a human team that solves problems. Sometimes problems can be solved by a single overall genius, but this is usually not the best strategy. So DeepSeek helps restore balance by validating open source shared ideas (data is another matter), demonstrating the power of continued algorithmic innovation, and making it possible to economically create AI agents that can be economically combined to produce useful and powerful AI systems. Of course, there are still questions that need to be answered:
● How to democratize access to the big data needed to build models while respecting copyright and other intellectual property rights?
● How can we build a professional model when the amount of data in certain professional disciplines is insufficient?
● How to evaluate a system that uses multiple AI agents to ensure it is running correctly? Even if individual agents are verified, does that mean they will be effective when combined?
Yejin Choi’s comments
Yejin Choi is Professor of HAI at the Dieter Schwartz Foundation, Professor of Computer Science, and Senior Fellow at HAI at Stanford
The success of the DeepSeek R1 model shows that when there is a proof of the existence of a solution (as shown in OpenAI’s o1), it is only a matter of time before others find a solution. DeepSeek’s decision to share detailed plans for R1 training and open source weighting models of different sizes has profound implications, as it will further accelerate progress and we will see new open source efforts continue to replicate and enhance R1. This shift marks the end of the era of violent expansion and ushers in a new phase that focuses on continued expansion through data synthesis, new learning frameworks, and new inference algorithms.
However, a major issue at present is how to use these powerful artificial intelligence systems to benefit all mankind. A model’s excellence on mathematical benchmarks does not immediately translate into solutions to difficult challenges facing mankind, including growing political tensions, natural disasters, or the continued spread of false information. The disconnect between technical capabilities and actual social impact remains one of the most pressing challenges in the field.
Michele Elam’s comments
Michele Elam is the William Robertson Coe Professor of Humanities, a Senior Fellow at Stanford HAI, and an Undergraduate Education Fellow at Bass University
Faced with the recent release of DeepSeek AI chat robot by China, which is significantly cheaper, has lower computing requirements, and has less environmental burden, amid the anxiety and confusion in the United States, few people have considered its impact on AI in the art field. In fact, DeepSeek’s significance to literature, performing arts, visual culture, etc. seems completely irrelevant in the face of seemingly higher-level anxieties such as national security, the economic devaluation of the U.S. AI industry, and the pros and cons of open source on innovation.
But in reality, DeepSeek’s complete opacity in terms of privacy protection, data sources and copyright debate has had a huge impact on the art world. ldquo; Opaque is actually a permissive term: DeepSeek takes a lazy attitude towards these issues. Needless to say, the SAG-AFTRA strike in the creative industries, The New York Times and other ongoing lawsuits.
In many ways, it is our own fault that DeepSeek was able to escape this blatant perfunctory attitude. The popularity of its chatbots is a magnifying reflection and taking advantage of the growing tendency of American consumers themselves to ignore these issues, which the industry actively encourages through its business model, diverting our attention in the name of return on investment.
Like TikTok, DeepSeek capitalizes on the past few years of our trend of gradually relinquishing privacy as we click on constantly updated and increasingly vague terms of service on our devices (often for the sake of personalization of so-called magic marketing euphemisms).
Arguably, as many have pointed out, DeepSeek’s greedy consumption of private and sensitive data takes advantage of the country’s lack of any regulation of AI (unlike the UK and EU) and puts the country at risk in many ways because we uphold the belief that regulation impedes innovation.
But, as far as art is concerned, we should focus on the way DeepSeek controls the keys to our imagination through pre-censorship, consistency with its nationalist ideology, and our unwitting or unconscious consent to its reality-modeling algorithms. Stanford University is currently adapting a more secure version of DeepSeek for experimentation through the Microsoft Azure project and warns the community not to use the commercial version because of security and privacy issues. Regardless, DeepSeek’s release highlights the technology’s huge impact, especially in shaping the way we experience reality and even what we think of as reality itself. As the early debate between Plato and Aristotle about the influence of drama and poetry suggested, this is where the power of art.
Comments from Mykel Kochenderfer
Mykel Kochenderfer is an associate professor in the Department of Aeronautics and Astronautics at Stanford University and a senior fellow at Stanford HAI
Artificial intelligence is increasingly being used to support critical safety or high-risk scenarios, from autonomous vehicles to clinical decision support. However, reconciling the unexplainability of current AI systems with safety engineering standards in high-risk applications remains a challenge. A particularly eye-catching feature of DeepSeek R1 is its reasoning transparency when responding to complex queries. The level of detail it provides can facilitate audits and help build trust in what it generates. This transparent reasoning provided by the language model when asking questions is called interpretability when reasoning. Although the interpretability of language models is still in its infancy and requires a lot of development to mature, the small steps we see today may lead to future systems that can assist humans more safely and reliably.
Another obstacle is the large amount of data and computing resources required to apply recent advances in artificial intelligence to many applications. DeepSeek demonstrates the huge potential of developing new methods to reduce reliance on large data sets and heavy computing resources. I hope that collaboration between academia and industry will accelerate these innovations. By creating more efficient algorithms, we can make language models more popular on edge devices, eliminating the need for continuous connectivity to expensive infrastructure. With a lot of common sense knowledge embedded in these language models, we can develop smarter, more useful, and more resilient applications, especially when the risks are highest.
Comments by James Landay
James Landay is Professor of Computer Science at Stanford University, Professor of Anand Rajaraman and Venky Harinarayan Schools of Engineering, and Co-Director of Stanford HAI
DeepSeek is a good thing for the field. They made their research public, and the model was released with open source weights, which meant that others could modify and run it on their own servers. They have reduced the cost of AI, which is beneficial to promoting the development of AI research and applications. One of the biggest criticisms of AI is the sustainability impact of training large basic models and supporting queries/reasoning for these models. DeepSeek demonstrates many useful optimizations that reduce computing costs in both areas. This is beneficial for the entire field because other companies or researchers can use the same optimizations (they are both documented in technical reports and the code is open source).
“The practice of sharing innovation through technical reports and open source code continues the tradition of open research that has driven the development of computer science over the past 40 years. rdquo;
This practice of innovation through technical reports and open source code sharing continues the tradition of open research that has driven the development of computer science over the past 40 years. As a research field, we should welcome this type of work. It will help everyone’s job become better. While many U.S. companies prefer proprietary models and still have questions about data privacy and security, DeepSeek’s open approach promotes broader participation, benefits the global AI community, and promotes iteration, progress and innovation.
Percy Liang’s comments
Percy Liang is associate professor of computer science at Stanford University, director of the Center for Fundamental Models Research (CRFM), and senior fellow at Stanford HAI
DeepSeek R1 shows that advanced AI will be widely available, difficult to control, and has no borders. This also shows that in addition to having a lot of computing resources, ingenuity and engineering design are equally important. For academia, the availability of more powerful open source weighting models is a boon because it allows for reproducibility, privacy protection, and research into advanced AI interiors.
Comments by Christopher Manning
Christopher Manning is the chairman of Thomas M., Department of Linguistics and Computer Science at Stanford University. Siebel Professor of Machine Learning and Deputy Director of Stanford HAI
People see this as some kind of sudden surprise, but in fact it’s not surprising if you’ve been following open source AI. DeepSeek has been publicly releasing open source models and detailed technical research reports for more than a year. Training costs for DeepSeek V3 will be announced in December 2024;R1-Lite-Preview will be released in November 2024.
“As an open country that has long promoted open science and engineering, the best way to understand the details of modern LLM design and engineering is to read detailed technical reports from China companies, which is a sad state of affairs. rdquo;
This release shows that so-called American cutting-edge AI companies do not have a huge technology moat. There are many excellent China Large Language Models (LLMs). At most, these companies are six months ahead, and maybe only OpenAI is truly ahead. As an open country that has long promoted open science and engineering, the best way to understand the details of modern LLM design and engineering is to read detailed technical reports from China companies. This is a sad state of affairs.
DeepSeek does some very good data engineering that minimizes data flow and allows efficient and stable training under fp8. They have made modest progress in technology, using a unique form of multi-head potential attention, a large number of experts in a mixed expert system, and their own simple and efficient form of reinforcement learning (RL), which is different from some people’s preference for rules-based rewards. But this is not exactly the next generation of technology. DeepSeek uses methods and models similar to those of other companies, and Deepseek-R1 has made a breakthrough in quickly catching up and delivering products of similar quality to OpenAI o1. This is not a new breakthrough in capabilities.
However, the release of DeepSeek-R1 does significantly advance the forefront of open source LLM and shows that the United States cannot curb the development of strong open source LLM. It may also mean that more U.S. companies will start using China’s LLM in their products, after before that they have generally avoided using these models, preferring to use Meta’s Llama model or other models from companies such as Databricks.
Comments by Julian Nyarko
Julian Nyarko is a professor at Stanford Law School and deputy director of Stanford HAI
Large Language Models (LLMs) are a general purpose technology used in many areas. Some companies create these models, while others use them for specific purposes. A key current debate is who is responsible for harmful model behavior, whether it is the developers who develop these models or the organizations that use them. In this context, the new DeepSeek model developed by a China startup highlights how the global nature of AI development complicates regulatory responses, especially when different countries have different legal norms and cultural understandings. While export controls are considered an important tool to ensure that leading AI implementations are consistent with our laws and value systems, DeepSeek’s success highlights the limitations of such measures when competing countries are able to independently develop and release state-of-the-art models. The nature of DeepSeek’s open source release further complicates the issue of legal liability. Because models can be freely obtained, modified, and deployed, the idea that model developers can and will effectively address the risks posed by their models may become increasingly unrealistic. Instead, the focus of regulation may need to shift to the downstream consequences of model use, which may place more responsibility on those who deploy the model.
Comments by Amy Zegart
Amy Zegart is a senior fellow at Morris Arnold and Nona Jean Cox at the Hoover Institution, a senior fellow at the Freeman Spogley Institute of International Studies, Professor of HAI at Stanford (part-time), and Professor of Political Science (part-time)
Over the past few weeks, discussions about DeepSeek have focused on chips and moats. How much does DeepSeek hoarding, smuggling, or innovating to circumvent U.S. export controls? Given DeepSeek’s progress, how many and what chips do researchers need now to innovate in cutting-edge areas? Did U.S. hyperscale companies like OpenAI end up spending billions of dollars building a competitive moat or simply a Maginot Line that provides a false sense of security? These are important questions, but the answers will take time.
“Almost all of the 200 engineers who wrote the breakthrough R1 paper last month were educated at China universities, and about half had only studied and worked in China. This should be a warning to U.S. policymakers. rdquo;
However, three serious geopolitical implications are already evident:
1. Success of local talents:DeepSeek has successfully leveraged local talent. Almost all of the 200 engineers who wrote the breakthrough R1 paper last month were educated at China universities, and about half had only studied and worked in China. This should be a warning to U.S. policymakers. In the era of science and technology, talents are an important source of national strength. ldquo; Although the slogan that the United States attracts the world’s best talents is often mentioned, it is increasingly inaccurate. Improvements in higher education levels and significant improvements in higher education institutions in China and elsewhere are redrawing the map of knowledge power. Meanwhile, the U.S. K-12 education system is in chaos, with a 15-year-old student ranking 34th in mathematics on a recent international test, behind Slovenia and Vietnam.
2. Imitate American universities rather than companies:DeepSeek imitates American universities, not companies. The startup hired young engineers rather than experienced industry veterans and gave them the freedom and resources to do crazy science aimed at long-term discovery rather than product development for the next quarter. Commercialization is an important part of innovation, but breakthroughs often start with basic research without clear product or profit goals. This basic research is the lifeline of universities and has been a pillar of American innovation leadership for decades, spawning everything from CubeSats to COVID-19 vaccines. However, today China is investing in basic research six times faster than the U.S. government. If current trends continue, China will overtake the U.S. within a decade. This is a crucial long-term innovation battlefield, and the United States is abandoning it.
3. Market turmoil:DeepSeek’s announcement disrupted U.S. markets, sending the Nasdaq Composite Index down 3%, and Nvidia’s shares down 17%, wiping out $600 billion in value. This was the largest corporate loss in a single day in U.S. history, equivalent to 65% of the U.S. annual defense budget. Today’s unintended consequences may be intentional actions tomorrow. Imagine an adversary deliberately announcing a real or fraudulent technological advancement to punish a particular company or shake another country’s capital market. Such gray area economic weapons can strike accurately or have a huge impact and may be difficult or even impossible to attribute to deliberate activities. And it works best without warning.
DeepSeek doesn’t just release new AI developments; it reveals the contours of an emerging geopolitical era with new sources of national power and new battlefields.
conclusion
DeepSeek’s release not only demonstrates China’s strong rise in artificial intelligence, but also triggers a global rethink of open source, legal liability, geopolitical power shift and privacy issues. The multidisciplinary perspective of Stanford University experts reveals that the United States ‘long-standing strategy of relying on closed technology and export controls is facing challenges. At the same time, DeepSeek’s success highlights the advantages and potential of the open innovation model, but also exposes its shortcomings in privacy protection and data transparency.
This dual discussion at the technical and policy levels reminds us that when enjoying the convenience brought by AI, we must pay more attention to its long-term impact and social responsibilities. As global AI competition intensifies, how to establish a unified and effective regulatory mechanism on a global scale to ensure that technological innovation and social well-being are in parallel will become an important issue that needs to be solved urgently in the future. DeepSeek’s success is not only a technological milestone, but also a profound inspiration for the future global science and technology landscape. This requires countries to pursue technological breakthroughs while also strengthening international cooperation and dialogue to jointly respond to the opportunities and challenges brought by this new era.
