DeepSeek's Open-Source AI Model Disrupts the Industry, Sparking Innovation and Controversy

DeepSeek -- a wake-up call for responsible innovation and risk management

Since its launch on Jan. 20, DeepSeek R1 has grabbed the attention of users as well as tech moguls, governments and policymakers worldwide -- from praises to skepticism, from adoption to bans, from innovative brilliance to unmeasurable privacy and security vulnerabilities. DeepSeek developed a large language model (LLM) comparable in its performance to OpenAI GTPo1 in a fraction of the time and cost it took OpenAI (and other tech companies) to build its own LLM. Using clever architecture optimization that slashes the cost of model training and inference, DeepSeek was able to develop an LLM within 60 days and for under $6 million. Indeed, DeepSeek should be acknowledged for taking the initiative to find better ways to optimize the model structure and code. It's a wake-up call, but far from being a "Sputnik moment." Every developer knows that there are two ways to gain performance. Optimizing the code and "throwing" a lot of computing power. The latter option is very costly, and developers are always advised to maximize the architecture optimization before resorting to more computing. It seems that with the rich valuations of artificial intelligence startups and the massive investments pouring in, developers got lazy. Why spend time optimizing model architecture if you have billions of dollars to spend on computing power? This is a wake-up call to all developers to go back to basics. Innovate responsibly, get out of your comfort zone, think outside the box, and don't be afraid to challenge the norm. There is no need to waste money and resources -- use them wisely. Like any other LLM, DeepSeek R1 falls short on reasoning, complex planning capabilities, understanding the physical world and persistent memory. So, there is no earth-shaking innovation here. It's time for scientists to go beyond LLMs, address these limitations, and develop a "new paradigm of AI architectures." It may not be LLM or generative AI -- a true revolution. Paving the way for accelerated innovation DeepSeek's approach could encourage developers worldwide, including developing countries, to innovate and develop their own AI applications regardless of low resources. The more people contribute to AI research and development, the faster innovation evolves and meaningful breakthroughs might be achieved. Recent: Crypto AI agents see 'remarkable traction' but value still unclear This aligns with the Nvidia projective: to make AI affordable and for every developer or scientist to develop their own AI applications. That's the meaning of project DIGITS, announced in early January, a $3,000 GPU for your desktop. Humanity needs "all minds on deck" to solve humanity's urgent problems. Resources may no longer be a barrier -- it is time to shake up old paradigms. At the same time, the DeepSeek release was also a wake-up call for actionable risk management and responsible AI. Read the fine print All applications come with terms of services, which the public often tends to ignore. Some alarming details in DeepSeek terms of service that could affect your privacy, security and even your business strategy: The above are clear violations of the General Data Protection Regulation (GDPR) and other GDPR privacy and security violations, as stated by the complaints filed by Belgium, Ireland and Italy, which also temporarily banned the use of DeepSeek. In March 2023, Italian regulators temporarily banned OpenAI ChatGPT for GDPR violations before allowing it back online a month after compliance improvements. Will DeepSeek comply as well? Bias and censorship Like other LLMs, DeepSeek R1 hallucinates, contains biases in its training data, and exhibits behavior that reflects China's political views on certain topics, such as censorship and privacy. Being a Chinese company, this is what is expected. The China Generative AI Law, which applies to providers and users of AI systems, states in Article 4: This is a censorship rule. It means those developing and/or using generative AI must support "core socialist values" and comply with Chinese laws regulating this topic. Not to say that other LLMs don't have their own biases and "agenda." This calls attention to the need for trustworthy, responsible AI and users to adhere to diligent AI risk management. LLM security vulnerabilities LLMs might be subject to adversarial attacks and security vulnerabilities. These vulnerabilities are even more concerning, as they will impact any applications built on this LLM by any organization or individual. Qualys has tested the distilled DeepSeek-R1 LLaMA 8B variant for vulnerabilities, ethical concerns and legal risks. The model failed at half of the jailbreak -- i.e., attempts to bypass the safety measures and ethical guidelines built into AI models like LLMs -- attacks tested. Goldman Sachs is considering using DeepSeek, but the model needs a security screening, like prompt injections and jailbreak. It is a security concern for any company that uses an AI model to power its applications, whether that model is Chinese or not. Goldman Sachs is implementing the correct risk management, and other organizations should follow this approach before deciding to use DeepSeek. Lessons learned We must be vigilant and diligent and implement adequate risk management before using any AI system or application. To mitigate any LLM's "agenda" and censorship elicited by centralized development, we might consider decentralized AI, preferably structured as a decentralized autonomous organization (DAO). AI knows no boundaries. It might be high time to consider unified global AI regulations. Opinion by: Merav Ozair, PhD. This article is for general information purposes and is not intended to be and should not be taken as legal or investment advice. The views, thoughts, and opinions expressed here are the author's alone and do not necessarily reflect or represent the views and opinions of Cointelegraph.

Q&A: How DeepSeek is changing the AI landscape

On Monday January 27, a little known Chinese start-up called DeepSeek sent shockwaves and panic through Silicon Valley and the global stock market with the launch of their generative artificial intelligence(AI) model that rivals the models of tech giants like OpenAI, Meta and Google. It's AI assistant became the no. 1 downloaded app in the U.S., surprising an industry that assumed only big Western companies could dominate AI. Many AI-related stocks, including Nvidia, took a hit as investors reevaluated the competitive landscape. But what brought the market to its knees is that DeepSeek developed their AI model at a fraction of the cost of models like ChatGPT and Gemini. The launch of DeepSeek is being coined "AI's Sputnik moment" in the global race to harness the power of AI. To break down what this development could mean for the future of AI and how it could impact society, we spoke with Arun Rai, Director of the Center for Digital Innovation at Robinson. How is DeepSeek's AI technology different and how was it so much cheaper to develop? AI development has long been a game of brute force -- bigger models, more computing power, and cutting-edge chips. OpenAI, Google DeepMind, and Anthropic have spent billions training models like GPT-4, relying on top-tier Nvidia GPUs (A100/H100) and massive cloud supercomputers. DeepSeek took a different approach. Instead of relying on expensive high-end chips, they optimized for efficiency, proving that powerful AI can be built through smarter software and hardware optimization. Key differences include: How did the launch of DeepSeek happen? DeepSeek's emergence wasn't gradual -- it was sudden and unexpected. Founded in late 2023, the company went from startup to industry disruptor in just over a year with the launch of its first large language model, DeepSeek-R1. The U.S. government had imposed trade restrictions on advanced Nvidia AI chips (A100/H100) to slow global competitors' AI progress. But DeepSeek adapted. Forced to work with less powerful but more available H800 GPUs, the company optimized its model to run on lower-end hardware without sacrificing performance. DeepSeek didn't just launch an AI model -- it reshaped the AI conversation showing that optimization, smarter software, and open access can be just as transformative as massive computing power. There's been a lot of buzz about DeepSeek being an "open-source model." What does open source mean and what impact does that have? AI models vary in how much access they allow, ranging from fully closed, paywalled systems to open-weight to completely open-source releases. DeepSeek's approach stands at the farthest end of openness -- one of the most unrestricted large-scale AI models yet. Most AI models are tightly controlled. OpenAI's GPT-4, Google DeepMind's Gemini, and Anthropic's Claude are all proprietary, meaning access is restricted to paying customers through APIs. Their underlying technology, architecture, and training data are kept private, and their companies control how the models are used, enforcing safety measures and preventing unauthorized modifications. Some AI models, like Meta's Llama 2, are open-weight but not fully open source. The model weights are publicly available, but license agreements restrict commercial use and large-scale deployment. Developers must agree to specific terms before using the model, and Meta still maintains oversight on who can use it and how. DeepSeek's model is different. It imposes no restrictions. Anyone -- from independent researchers to private companies -- can fine-tune and deploy the model without permission or licensing agreements. This approach has major advantages. It democratizes AI innovation by giving startups, researchers, and developers access to cutting-edge AI without licensing fees. It encourages global AI development, allowing independent AI labs to improve the model. And it breaks the monopoly of large AI firms, offering a powerful alternative to proprietary, paywalled AI models. But it also introduces significant risks. Unlike proprietary AI, where companies can monitor and restrict harmful applications, DeepSeek's model can be repurposed by anyone, including bad actors. This raises concerns about misinformation, deepfake production, and AI-generated fraud. Without built-in safeguards, open AI systems could be used for mass disinformation, cyberattacks, or social manipulation. DeepSeek's move has reignited a debate: Should AI models be fully open, or should companies enforce restrictions to prevent misuse? Some see DeepSeek's release as a win for AI accessibility and openness driving innovation, while others warn that unrestricted AI could lead to unintended consequences and new risks that no one can control. Is the launch of DeepSeek something to panic over or be excited about? The launch of DeepSeek marks a transformative moment for AI -- one that brings both exciting opportunities and important challenges. It has opened new possibilities for AI development while also raising fresh questions about security, responsibility, and control. On one hand, DeepSeek's open-source release expands access to cutting-edge AI like never before, which could lead to faster breakthroughs in fields like science, health care, and business. DeepSeek's efficiency-first approach also challenges the assumption that only companies with billions in computing power can build leading AI models. If this method scales, it could redefine how AI is developed globally. At the same time, its unrestricted availability introduces complex risks. What are the concerns with DeepSeek? DeepSeek's launch has raised critical questions about security, control, and ethical responsibility. The main concerns center on national security, intellectual property, and misuse. Unlike proprietary AI models, DeepSeek's open-source approach allows anyone to modify and deploy it without oversight. This raises fears that bad actors could use it for misinformation campaigns, deepfakes, or AI-driven cyberattacks. The U.S. Navy was the first to ban DeepSeek, citing security concerns over potential data access by the Chinese government. Since then, Texas, Taiwan, and Italy have also restricted its use, while regulators in South Korea, France, Ireland, and the Netherlands are reviewing its data practices, reflecting broader concerns about privacy and national security. Similar concerns were at the center of the TikTok controversy, where U.S. officials worried that data from an app used by millions of Americans could be accessed by the Chinese government. The debate isn't just about DeepSeek -- it's about how open AI should be. Can AI be both widely accessible and responsibly managed? That question will shape the future of AI policy and innovation. How does regulation play a role in the development of AI? AI regulation is at a crossroads. Governments are racing to balance innovation with security, trying to foster AI development while preventing misuse. But the challenge is AI is evolving faster than laws can keep up. In the U.S., regulation has focused on export controls and national security, but one of the biggest challenges in AI regulation is who takes responsibility for open models. As AI continues to advance, policymakers face a dilemma -- how to encourage progress while preventing risks. Should AI models be open and accessible to all, or should governments enforce stricter controls to limit potential misuse? The answers will shape how AI is developed, who benefits from it, and who holds the power to regulate its impact. How could DeepSeek's impact on the AI landscape ultimately impact society? DeepSeek's impact on AI isn't just about one model -- it's about who has access to AI and how that changes innovation, competition, and governance. By making a powerful AI model open-source, DeepSeek has lowered the barrier to AI development, enabling more researchers, startups, and organizations to build and deploy AI without relying on big tech firms or government-backed research labs. It also challenges the idea that AI progress depends solely on massive computing power, proving that smarter software and hardware optimization can rival brute-force approaches. At the same time, decentralization makes AI harder to regulate. Without a central authority controlling its deployment, open AI models can be used and modified freely -- driving both innovation and new risks. DeepSeek has forced a key question to the forefront: Will AI's future be shaped by a handful of well-funded Western firms and government-backed AI research labs, or by a broader, more open ecosystem? That choice will determine not just who has access to AI, but how it reshapes society.

OpenAI says DeepSeek 'inappropriately' copied ChatGPT -- but it's facing copyright claims, too

Until a few weeks ago, few people in the Western world had heard of a small Chinese artificial intelligence (AI) company known as DeepSeek. But on January 20, it captured global attention when it released a new AI model called R1. R1 is a "reasoning" model, meaning it works through tasks step by step and details its working process to a user. It is a more advanced version of DeepSeek's V3 model, which was released in December. DeepSeek's new offering is almost as powerful as rival company OpenAI's most advanced AI model o1, but at a fraction of the cost. Within days, DeepSeek's app surpassed ChatGPT in new downloads and set stock prices of tech companies in the United States tumbling. It also led OpenAI to claim that its Chinese rival had effectively pilfered some of the crown jewels from OpenAI's models to build its own. In a statement to the New York Times, the company said, "We are aware of and reviewing indications that DeepSeek may have inappropriately distilled our models, and will share information as we know more. We take aggressive, proactive countermeasures to protect our technology and will continue working closely with the US government to protect the most capable models being built here." The Conversation approached DeepSeek for comment, but it did not respond. But even if DeepSeek copied -- or, in scientific parlance, "distilled" -- at least some of ChatGPT to build R1, it's worth remembering that OpenAI also stands accused of disrespecting intellectual property while developing its models. What is distillation? Model distillation is a common machine learning technique in which a smaller "student model" is trained on predictions of a larger and more complex "teacher model." When completed, the student may be nearly as good as the teacher but will represent the teacher's knowledge more effectively and compactly. To do so, it is not necessary to access the inner workings of the teacher. All one needs to pull off this trick is to ask the teacher model enough questions to train the student. This is what OpenAI claims DeepSeek has done: queried OpenAI's o1 at a massive scale and used the observed outputs to train DeepSeek's own, more efficient models. A fraction of the resources DeepSeek claims that both the training and usage of R1 required only a fraction of the resources needed to develop their competitors' best models. There are reasons to be skeptical of some of the company's marketing hype -- for example, a new independent report suggests the hardware spend on R1 was as high as US$500 million. But even so, DeepSeek was still built very quickly and efficiently compared with rival models. This might be because DeepSeek distilled OpenAI's output. However, there is currently no method to prove this conclusively. One method that is in the early stages of development is watermarking AI outputs. This adds invisible patterns to the outputs, similar to those applied to copyrighted images. There are various ways to do this in theory, but none is effective or efficient enough to have made it into practice. There are other reasons that help explain DeepSeek's success, such as the company's deep and challenging technical work. The technical advances made by DeepSeek included taking advantage of less powerful but cheaper AI chips (also called graphical processing units, or GPUs). DeepSeek had no choice but to adapt after the US banned firms from exporting the most powerful AI chips to China. While Western AI companies can buy these powerful units, the export ban forced Chinese companies to innovate to make the best use of cheaper alternatives. A series of lawsuits OpenAI's terms of use explicitly state nobody may use its AI models to develop competing products. However, its own models are trained on massive datasets scraped from the web. These datasets contained a substantial amount of copyrighted material, which OpenAI says it is entitled to use on the basis of "fair use": "Training AI models using publicly available internet materials is fair use, as supported by long-standing and widely accepted precedents. We view this principle as fair to creators, necessary for innovators, and critical for US competitiveness." This argument will be tested in court. Newspapers, musicians, authors and other creatives have filed a series of lawsuits against OpenAI on the grounds of copyright infringement. Of course, this is quite distinct to what OpenAI accuses DeepSeek of doing. Nevertheless, OpenAI isn't attracting much sympathy for its claim that DeepSeek illegitimately harvested its model output. The war of words and lawsuits is an artifact of how the rapid advance of AI has outpaced the development of clear legal rules for the industry. And while these recent events might reduce the power of AI incumbents, much hinges on the outcome of the various ongoing legal disputes. Shaking up the global conversation DeepSeek has shown it is possible to develop state-of-the-art models cheaply and efficiently. Whether they can compete with OpenAI on a level playing field remains to be seen. Over the weekend, OpenAI attempted to demonstrate its supremacy by publicly releasing its most advanced consumer model, o3-mini. OpenAI claims this model substantially outperforms even its own previous market-leading version, o1, and is the "most cost-efficient model in our reasoning series." These developments herald an era of increased choice for consumers, with a diversity of AI models on the market. This is good news for users: competitive pressures will make models cheaper to use. And the benefits extend further. Training and using these models places a massive strain on global energy consumption. As these models become more ubiquitous, we all benefit from improvements to their efficiency. DeepSeek's rise certainly marks new territory for building models more cheaply and efficiently. Perhaps it will also shake up the global conversation on how AI companies should collect and use their training data.

OpenAI says DeepSeek 'inappropriately' copied ChatGPT - but it's facing copyright claims too

Until a few weeks ago, few people in the Western world had heard of a small Chinese artificial intelligence (AI) company known as DeepSeek. But on January 20, it captured global attention when it released a new AI model called R1. R1 is a "reasoning" model, meaning it works through tasks step by step and details its working process to a user. It is a more advanced version of DeepSeek's V3 model, which was released in December. DeepSeek's new offering is almost as powerful as rival company OpenAI's most advanced AI model o1, but at a fraction of the cost. Within days, DeepSeek's app surpassed ChatGPT in new downloads and set stock prices of tech companies in the United States tumbling. It also led OpenAI to claim that its Chinese rival had effectively pilfered some of the crown jewels from OpenAI's models to build its own. In a statement to the New York Times, the company said: We are aware of and reviewing indications that DeepSeek may have inappropriately distilled our models, and will share information as we know more. We take aggressive, proactive countermeasures to protect our technology and will continue working closely with the US government to protect the most capable models being built here. The Conversation approached DeepSeek for comment, but it did not respond. But even if DeepSeek copied - or, in scientific parlance, "distilled" - at least some of ChatGPT to build R1, it's worth remembering that OpenAI also stands accused of disrespecting intellectual property while developing its models. What is distillation? Model distillation is a common machine learning technique in which a smaller "student model" is trained on predictions of a larger and more complex "teacher model". When completed, the student may be nearly as good as the teacher but will represent the teacher's knowledge more effectively and compactly. To do so, it is not necessary to access the inner workings of the teacher. All one needs to pull off this trick is to ask the teacher model enough questions to train the student. This is what OpenAI claims DeepSeek has done: queried OpenAI's o1 at a massive scale and used the observed outputs to train DeepSeek's own, more efficient models. A fraction of the resources DeepSeek claims that both the training and usage of R1 required only a fraction of the resources needed to develop their competitors' best models. There are reasons to be sceptical of some of the company's marketing hype - for example, a new independent report suggests the hardware spend on R1 was as high as US$500 million. But even so, DeepSeek was still built very quickly and efficiently compared with rival models. This might be because DeepSeek distilled OpenAI's output. However, there is currently no method to prove this conclusively. One method that is in the early stages of development is watermarking AI outputs. This adds invisible patterns to the outputs, similar to those applied to copyrighted images. There are various ways to do this in theory, but none is effective or efficient enough to have made it into practice. There are other reasons that help explain DeepSeek's success, such as the company's deep and challenging technical work. The technical advances made by DeepSeek included taking advantage of less powerful but cheaper AI chips (also called graphical processing units, or GPUs). DeepSeek had no choice but to adapt after the US has banned firms from exporting the most powerful AI chips to China. While Western AI companies can buy these powerful units, the export ban forced Chinese companies to innovate to make the best use of cheaper alternatives. A series of lawsuits OpenAI's terms of use explicitly state nobody may use its AI models to develop competing products. However, its own models are trained on massive datasets scraped from the web. These datasets contained a substantial amount of copyrighted material, which OpenAI says it is entitled to use on the basis of "fair use": Training AI models using publicly available internet materials is fair use, as supported by long-standing and widely accepted precedents. We view this principle as fair to creators, necessary for innovators, and critical for US competitiveness. This argument will be tested in court. Newspapers, musicians, authors and other creatives have filed a series of lawsuits against OpenAI on the grounds of copyright infringement. Of course, this is quite distinct to what OpenAI accuses DeepSeek of doing. Nevertheless OpenAI isn't attracting much sympathy for its claim that DeepSeek illegitimately harvested its model output. The war of words and lawsuits is an artefact of how the rapid advance of AI has outpaced the development of clear legal rules for the industry. And while these recent events might reduce the power of AI incumbents, much hinges on the outcome of the various ongoing legal disputes. Shaking up the global conversation DeepSeek has shown it is possible to develop state-of-the-art models cheaply and efficiently. Whether they can compete with OpenAI on a level playing field remains to be seen. Over the weekend, OpenAI attempted to demonstrate its supremacy by publicly releasing its most advanced consumer model, o3-mini. OpenAI claims this model substantially outperforms even its own previous market-leading version, o1, and is the "most cost-efficient model in our reasoning series". These developments herald an era of increased choice for consumers, with a diversity of AI models on the market. This is good news for users: competitive pressures will make models cheaper to use. And the benefits extend further. Training and using these models places a massive strain on global energy consumption. As these models become more ubiquitous, we all benefit from improvements to their efficiency. DeepSeek's rise certainly marks new territory for building models more cheaply and efficiently. Perhaps it will also shake up the global conversation on how AI companies should collect and use their training data.

What is open-source AI and how could DeepSeek change the industry?

The World Economic Forum's AI Transformation of Industries initiative is exploring the challenges and opportunities in AI-driven innovation. Has AI just had its "Sputnik moment"? That's to say, has Chinese company DeepSeek created a new approach to developing AI models that is cost-effective, energy-efficient and accessible? That's what influential tech venture capitalist Marc Andreessen thinks, hailing DeepSeek as "one of the most amazing and impressive breakthroughs I've ever seen - and as open source, a profound gift to the world". But what is open-source AI? And why is it such an important breakthrough? Based on news headlines over the past few weeks, DeepSeek has upset the tech applecart. The UK's Financial Times likens founder Liang Wenfeng to a David taking on the Goliath of America's big tech giants. Based in Hangzhou and funded by Chinese hedge fund High Flyer, DeepSeek was only established in 2023. Its flagship DeepSeek-R1 model, launched in January 2025, appears to show comparable reasoning and mathematical skills with other leading rivals. But there are some key differences on how it developed. Unlike OpenAI's ChatGPT and Anthropic's Claude for example - whose models, data sets and algorithms are secret - DeepSeek is open source, meaning it is available for anyone to download, copy and build upon. Its code and comprehensive technical explanations are freely shared, enabling global developers and organizations to access, modify and implement. By contrast, although Meta and Google's models are open to anyone to view, they are also not seen as being truly open source because the way users apply the models is restricted by licences and the training data sets aren't made public. Another key differentiator of DeepSeek is the cost to build. Some numbers suggest it was built for $5.6 million - just 10% of the cost of Meta's Llama. This has pulled into question whether the huge funding circulating around the AI market, in the US in particular, is as necessary as first perceived. Shares in Nvidia, which designs semiconductor technology powering AI, fell sharply at the end of January as a result. If it's shorter and less energy-intensive to set up, that's good news for the environment: the high electricity consumption related to AI is significantly driving up carbon emissions. DeepSeek's approach could offer a path towards more sustainable AI scaling. By the end of January 2025, DeepSeek had overtaken ChatGPT as the most downloaded free app on the Apple App Store in the US. Many believe its rapid rise could be a significant step in democratizing AI, allowing smaller companies, start-ups and individual developers to build on DeepSeek-R1. A more egalitarian AI-for-all approach could accelerate innovation in regions with limited access to cutting-edge technologies and drive advancements in the technology more broadly, it's argued. Instead of focusing on building specialized models, developers can dedicate resources to creating specialized applications, unlocking the power of AI to solve real-world problems. This aim is also promoted by the World Economic Forum's AI Transformation of Industries initiative, which is exploring the challenges and opportunities in AI-driven innovation. Its latest whitepaper explores the transformative potential of AI across industries. Many in the tech community believe that the open-source nature of DeepSeek will help foster a collaborative environment and accelerate AI innovation. There is also a feeling that open-source models may be perceived as being more trustworthy as people are able to interrogate training data. Indeed, the lack of transparency around the training data used for many of the leading models has led to lawsuits against several leading players. That said, DeepSeek has faced criticism for alleged censorship, both in its answers and training data. And it has been prohibited by several governments, citing privacy concerns. Ensuring clear data handling, robust security measures and minimizing censorship will be critical for building trust and fostering wider acceptance of the technology. On the other side of the argument, some people contend that open-source AI poses considerable risks, and could be used for malicious purposes such as developing bioweapons, as well as facilitating the spread of mis- and disinformation. Either way, DeepSeek is causing the AI industry to rethink competitiveness. Its apparent cost-effective, open-source approach disrupts traditional notions and is prompting countries to reflect on what truly enables success in the AI era. And while its open-source framework promotes inclusivity and transparency, it also raises important questions about data privacy, geopolitical dynamics and security. As the technology continues to evolve, navigating these challenges will be key to unlocking its full potential and ensuring its responsible adoption.

Putting DeepSeek to the test: how its performance compares against other AI tools

China's new DeepSeek Large Language Model (LLM) has disrupted the US-dominated market, offering a relatively high-performance chatbot model at significantly lower cost. The reduced cost of development and lower subscription prices compared with US AI tools contributed to American chip maker Nvidia losing US$600 billion (£480 billion) in market value over one day. Nvidia makes the computer chips used to train the majority of LLMs, the underlying technology used in ChatGPT and other AI chatbots. DeepSeek uses cheaper Nvidia H800 chips over the more expensive state-of-the-art versions. ChatGPT developer OpenAI reportedly spent somewhere between US$100 million and US$1 billion on the development of a very recent version of its product called o1. In contrast, DeepSeek accomplished its training in just two months at a cost of US$5.6 million using a series of clever innovations. But just how well does DeepSeek's AI chatbot, R1, compare with other, similar AI tools on performance? DeepSeek claims its models perform comparably to OpenAI's offerings, even exceeding the o1 model in certain benchmark tests. However, benchmarks that use Massive Multitask Language Understanding (MMLU) tests evaluate knowledge across multiple subjects using multiple choice questions. Many LLMs are trained and optimised for such tests, making them unreliable as true indicators of real-world performance. An alternative methodology for the objective evaluation of LLMs uses a set of tests developed by researchers at Cardiff Metropolitan, Bristol and Cardiff universities - known collectively as the Knowledge Observation Group (KOG). These tests probe LLMs' ability to mimic human language and knowledge through questions that require implicit human understanding to answer. The core tests are kept secret, to avoid LLM companies training their models for these tests. KOG deployed public tests inspired by work by Colin Fraser, a data scientist at Meta, to evaluate DeepSeek against other LLMs. The following results were observed: The tests used to produce this table are "adversarial" in nature. In other words, they are designed to be "hard" and to test LLMs in way that are not sympathetic to how they are designed. This means the performance of these models in this test is likely to be different to their performance in mainstream benchmarking tests. DeepSeek scored 5.5 out of 6, outperforming OpenAI's o1 - its advanced reasoning (known as "chain-of-thought") model - as well as ChatGPT-4o, the free version of ChatGPT. But Deepseek was marginally outperformed by Anthropic's ClaudeAI and OpenAI's o1 mini, both of which scored a perfect 6/6. It's interesting that o1 underperformed against its "smaller" counterpart, o1 mini. DeepThink R1 - a chain-of-thought AI tool made by DeepSeek - underperformed in comparison to DeepSeek with a score of 3.5. This result shows how competitive DeepSeek's R1 chatbot already is, beating OpenAI's flagship models. It is likely to spur further development for DeepSeek, which now has a strong foundation to build upon. However, the Chinese tech company does have one serious problem the other LLMs do not: censorship. Censorship challenges Despite its strong performance and popularity, DeepSeek has faced criticism over its responses to politically sensitive topics in China. For instance, prompts related to Tiananmen Square, Taiwan, Uyghur Muslims and democratic movements are met with the response: "Sorry, that is beyond my current scope." But this issue is not necessarily unique to DeepSeek, and the potential for political influence and censorship in LLMs more generally is a growing concern. The announcement of Donald Trump's US$500 billion Stargate LLM project, involving OpenAI, Nvidia, Oracle, Microsoft, and Arm, also raises fears of political influence. Additionally, Meta's recent decision to abandon fact-checking on Facebook and Instagram suggests an increasing trend toward populism over truthfulness. DeepSeek's arrival has caused serious disruption to the LLM market. US companies such as OpenAI and Anthropic will be forced to innovate their products to maintain relevance and match its performance and cost. DeepSeek's success is already challenging the status quo, demonstrating that high-performance LLM models can be developed without billion-dollar budgets. It also highlights the risks of LLM censorship, the spread of misinformation, and why independent evaluations matter. As LLMs become more deeply embedded in global politics and business, transparency and accountability will be essential to ensure that the future of LLMs is safe, useful and trustworthy.

DeepSeek vs. ChatGPT fuels debate over AI building blocks

When Chinese startup DeepSeek released its AI model this month, it was hailed as a breakthrough, a sign that China's artificial intelligence companies could compete with their Silicon Valley counterparts using fewer resources. The narrative was clear: DeepSeek had done more with less, finding clever workarounds to U.S. chip restrictions. However, that storyline has begun to shift. OpenAI, the U.S.-based company behind ChatGPT, now claims DeepSeek may have improperly used its proprietary data to train its model, raising questions about whether DeepSeek's success was truly an engineering marvel. In statements to several media outlets this week, OpenAI said it is reviewing indications that DeepSeek may have trained its AI by mimicking responses from OpenAI's models. The process, known as distillation, is common among AI developers but is prohibited by OpenAI's terms of service, which forbid using its model outputs to train competing systems. Some U.S. officials appear to support OpenAI's concerns. At his confirmation hearing this week, Commerce secretary nominee Howard Lutnick accused DeepSeek of misusing U.S. technology to create a "dirt cheap" AI model. "They stole things. They broke in. They've taken our IP," Lutnick said of China. David Sacks, the White House czar for AI and cryptocurrency, was more measured, saying only that it is "possible" that DeepSeek had stolen U.S. intellectual property. In an interview with the cable news network Fox News, Sacks added that there is "substantial evidence" that DeepSeek "distilled the knowledge out of OpenAI's models," adding that stronger efforts are needed to curb the rise of "copycat" AI systems. At the center of the dispute is a key question about AI's future: how much control should companies have over their own AI models, when those programs were themselves built using data taken from others? AI data fight The question is especially relevant for OpenAI, which faces its own legal challenges. The company has been sued by several media companies and authors who accuse it of illegally using copyrighted material to train its AI models. Justin Hughes, a Loyola Law School professor specializing in intellectual property, AI, and data rights, said OpenAI's accusations against DeepSeek are "deeply ironic," given the company's own legal troubles. "OpenAI has had no problem taking everyone else's content and claiming it's 'fair,'" Hughes told VOA in an email. "If the reports are accurate that OpenAI violated other platforms' terms of service to get the training data it has wanted, that would just add an extra layer of irony - dare we say hypocrisy - to OpenAI complaining about DeepSeek." DeepSeek has not responded to OpenAI's accusations. In a technical paper released with its new chatbot, DeepSeek acknowledged that some of its models were trained alongside other open-source models - such as Qwen, developed by China's Alibaba, and Llama, released by Meta - according to Johnny Zou, a Hong Kong-based AI investment specialist. However, OpenAI appears to be alleging that DeepSeek improperly used its closed-source models - which cannot be freely accessed or used to train other AI systems. "It's quite a serious statement," said Zou, who noted that OpenAI has not yet presented evidence of wrongdoing by DeepSeek. Proving improper distillation may be difficult without disclosing details on how its own models were trained, Zou added. Even if OpenAI presents concrete proof, its legal options may be limited. Although Zou noted that the company could pursue a case against DeepSeek for violating its terms of service, not all experts believe such a claim would hold up in court. "Even assuming DeepSeek trained on OpenAI's data, I don't think OpenAI has much of a case," said Mark Lemley, a professor at Stanford Law School who specializes in intellectual property and technology. Even though AI models often have restrictive terms of service, "no model creator has actually tried to enforce these terms with monetary penalties or injunctive relief," Lemley wrote in a recent paper with co-author Peter Henderson. The paper argues that these restrictions may be unenforceable, since the materials they aim to protect are "largely not copyrightable." "There are compelling reasons for many of these provisions to be unenforceable: they chill good faith research, constrain competition, and create quasi-copyright ownership where none should exist," the paper noted. OpenAI's main legal argument would likely be breach of contract, said Hughes. Even if that were the case, though, he added, "good luck enforcing that against a Chinese company without meaningful assets in the United States." Possible options The financial stakes are adding urgency to the debate. U.S. tech stocks dipped Monday after following news of DeepSeek's advances, though they later regained some ground. Commerce nominee Lutnick suggested that further government action, including tariffs, could be used to deter China from copying advanced AI models. But speaking the same day, U.S. President Donald Trump appeared to take a different view, surprising some industry insiders with an optimistic take on DeepSeek's breakthrough. The Chinese company's low-cost model, Trump said, was "very much a positive development" for AI, because "instead of spending billions and billions, you'll spend less, and you'll come up with hopefully the same solution." If DeepSeek has succeeded in building a relatively cheap and competitive AI model, that may be bad for those with investment - or stock options - in current generative AI companies, Hughes said. "But it might be good for the rest of us," he added, noting that until recently it appeared that only the existing tech giants "had the resources to play in the generative AI sandbox." "If DeepSeek disproved that, we should hope that what can be done by a team of engineers in China can be done by a similarly resourced team of engineers in Detroit or Denver or Boston," he said.

DeepSeek has burst the AI hype bubble - now all bets are off

The Chinese firm threatens the dominance of Silicon Valley's AI elite, and its innovations show the technology could be more affordable and less costly to the environment In poker, the value of the cards in your hand is often less important than what your competitors think you might hold. You don't need a royal flush as long as you can convince others you have one. Sam Altman, CEO of OpenAI, knows this well, having played poker extensively during his student days. Following the astronomical success of its generative artificial intelligence tool ChatGPT, the company has convinced many backers that it holds all the aces, telling the world that scale is the key to progress and that betting on this will reap big rewards. On 21 January, Altman announced Stargate, a $500-billion plan to build vast data centres for future AI models. As he said in an interview in 2023: "It's totally hopeless to compete with us." But Chinese AI company DeepSeek now looks to have called his bluff. It sent shock waves through Silicon Valley over the past two weeks with the release of AI models that are apparently as capable as OpenAI's best, but at a fraction of the cost and computational power (see "Does DeepSeek show a way to slash the energy demands of AI?"). This young upstart, with less than a tenth as many employees as OpenAI, has punctured the idea that US companies hold some secret recipe for building AI or that they need such enormous resources to do so. For those concerned about the accumulation of power in Silicon Valley, the arrival of competition is welcome, but DeepSeek's model brings concerns of its own. For one thing, its answers stick closely to the Chinese government's party line, and it even censors itself in real time. Security researchers have also warned that it lacks robust guardrails against inappropriate use. Nevertheless, its arrival on the scene suggests there are huge innovations in generative AI yet to come. Plus, cheaper models that require less computational power should open the door to entirely new applications for the technology, which may also make it affordable to more people and less damaging to the planet. With more players around the table, the stakes couldn't be higher.

Q&A: Unpacking DeepSeek -- distillation, ethics and national security

Since the Chinese AI startup DeepSeek released its powerful large language model R1, it has sent ripples through Silicon Valley and the U.S. stock market, sparking widespread discussion and debate. Ambuj Tewari, professor of statistics at the University of Michigan and a leading expert in artificial intelligence and machine learning shares his insights on the technical, ethical and market-related aspects of DeepSeek's breakthrough. OpenAI has accused DeepSeek of using model distillation to train its own models based on OpenAI's technology. Can you explain how model distillation typically works, and under what circumstances it might be considered ethical or compliant with AI development best practices? Model or knowledge distillation typically involves generating responses from the stronger model to train a weaker model so that the weaker model improves. It is a totally normal practice if the stronger model was released with a license that permits such use. But OpenAI's terms of use of chatGPT explicitly forbid use of their model for purposes such as model distillation. Is it possible that DeepSeek utilized other open-source models, such as Meta Platforms' LLaMA or Alibaba's Qwen, for knowledge distillation, rather than relying on OpenAI's proprietary models? It is hard to say. Even in the same family of models, say Llama or Qwen, not all models are released with the same license. If the license of a model permits model distillation, then there is nothing illegal or unethical in doing that. In the R1 paper, it is mentioned that the process actually worked in the opposite direction: knowledge was distilled from R1 to LLaMA and Qwen to enhance the reasoning capabilities of the latter models. What evidence could an AI company provide to demonstrate that its models were developed independently, without relying on proprietary technology from another organization? Since there is the presumption of innocence in legal matters, the burden of proof will be on OpenAI to prove that DeepSeek did in fact violate their terms of service. Since only the final model developed by DeepSeek is public and not its training data, it might be hard to prove the accusation. Since OpenAI has not made its evidence public yet, it is hard to say how strong a case they have. Are there industry standards or transparency measures that AI companies could adopt to build trust and demonstrate compliance with ethical AI development? There are currently little universally accepted standards on development of AI models by companies. Proponents of open models say that openness leads to more transparency. But making the model weights open is not the same as making the entire process from data collection to training open. There are also concerns about whether use of copyrighted materials such as books for training AI models is fair use or not. A prominent example is the lawsuit filed by The New York Times against OpenAI, which highlights the legal and ethical debates surrounding this issue. There are questions around social biases in training data affecting the model's output. There are also concerns around increasing energy requirements and its implication for climate change. Most of these issues are being actively debated with little consensus. Some U.S. officials have expressed concerns that DeepSeek could pose national security risks. What's your take on this? It would be deeply concerning if U.S. citizens' data is stored on DeepSeek's servers and the Chinese government gets access to it. However, the model weights are open and hence it can be run on servers owned by U.S. companies. In fact, Microsoft has already started hosting DeepSeek's models.

DeepSeek's Controversial Rise: Did They Use OpenAI's Data Illegally?

OpenAI has accused Chinese AI startup DeepSeek of unlawfully using its proprietary model outputs to train a competing AI system. Central to the dispute is the use of "model distillation," a widely recognized technique that enables smaller models to replicate the behavior of larger, more advanced systems. OpenAI claims that DeepSeek's actions represent a violation of its intellectual property rights and terms of service, sparking a broader debate about the ethical and legal boundaries in artificial intelligence (AI) development. Although this isn't a simple case of right versus wrong. While OpenAI has presented evidence, including instances where DeepSeek's model reportedly referred to itself as being developed by OpenAI, critics are quick to point out the complexities of the issue. After all, OpenAI's own models were trained on publicly available internet data, sparking debates about where the line should be drawn when it comes to data usage and intellectual property. This case is more than just a legal battle -- it's a reflection of the growing pains of a rapidly evolving industry, where the rules are still being written. So, how do we strike the right balance between fostering innovation and protecting what's proprietary? This case has drawn significant attention within the AI community, as it highlights the challenges of balancing innovation with the protection of proprietary technologies. The outcome could have far-reaching implications for how intellectual property is safeguarded in the rapidly evolving AI industry. Model distillation is a widely used technique in AI development that involves training a smaller, more efficient model to replicate the outputs of a larger, more complex system. This method is often employed to reduce computational costs while maintaining high levels of performance. However, OpenAI alleges that DeepSeek has exploited this technique to create a competing AI system by using OpenAI's proprietary outputs without proper authorization. The accusation raises critical questions about the ethical use of AI-generated data and the extent to which intellectual property protections apply in this context. While model distillation is a legitimate practice, its misuse to replicate proprietary systems for competitive purposes has sparked concerns about fairness and ethical boundaries. OpenAI has presented evidence to substantiate its claims against DeepSeek. A notable example includes instances where DeepSeek's model reportedly referred to itself as being developed by OpenAI. According to OpenAI, such occurrences strongly suggest that DeepSeek trained its system using data derived from OpenAI's proprietary models. If these allegations are proven, the case could establish a significant precedent for how AI companies protect their intellectual property and address unauthorized use of their outputs. It also underscores the need for clearer guidelines to define the boundaries of ethical AI development and competition. OpenAI's terms of service explicitly prohibit the use of its outputs to develop competing models. By allegedly violating these terms, DeepSeek has not only triggered legal concerns but also reignited ethical debates within the AI community. The case highlights the tension between fostering innovation and respecting intellectual property rights. While model distillation is a recognized and widely accepted technique, critics argue that using it to replicate proprietary systems undermines the principles of fair competition. This dispute serves as a reminder of the ethical complexities that arise when advanced AI technologies are used without proper authorization. Here are more guides from our previous articles and guides related to Model Distillation that you may find helpful. One of the core challenges in this dispute is the difficulty of detecting and preventing the unauthorized use of AI outputs. Many AI systems, including OpenAI's, make their outputs publicly accessible, which complicates efforts to track misuse. This accessibility, while beneficial for fostering innovation, also creates vulnerabilities that can be exploited. The case underscores the urgent need for more robust mechanisms to protect proprietary AI technologies. Potential solutions include developing advanced watermarking techniques, implementing stricter access controls, and fostering industry-wide collaboration to establish best practices for safeguarding intellectual property. Despite the controversy, DeepSeek's model has demonstrated impressive performance, achieving results that rival leading systems like OpenAI's ChatGPT. This success highlights the effectiveness of model distillation as a technique but also raises concerns about fairness in competition. Critics argue that using another company's outputs to achieve such performance undermines the principles of ethical AI development. The case illustrates the broader tension between fostering innovation and protecting intellectual property. It also emphasizes the need for clearer regulations to ensure that competitive practices remain fair and transparent. The OpenAI-DeepSeek dispute is part of a larger conversation about intellectual property theft in the AI sector. It echoes longstanding concerns about trade secret theft, particularly involving Chinese companies, and highlights the growing risks as AI technologies become more advanced and valuable. As the industry continues to evolve, the potential for unauthorized use of AI outputs is likely to increase. This underscores the importance of establishing stronger safeguards, clearer regulations, and industry-wide collaboration to address these challenges effectively. By doing so, the AI community can create an environment that supports innovation while respecting intellectual property rights. In response to the incident, OpenAI has called for closer collaboration with governments to protect advanced AI models from misuse. Such partnerships could play a crucial role in establishing legal frameworks and enforcement mechanisms to address intellectual property violations and other ethical concerns in the AI industry. By working with policymakers, OpenAI aims to create a more secure environment for innovation. This collaboration could help ensure that ethical standards are upheld while fostering the responsible development and deployment of AI technologies. While OpenAI has taken a strong stance against DeepSeek, it has faced criticism for its own practices. OpenAI's models were trained on publicly scraped internet data, raising questions about the consistency of its position on data usage ethics. Critics argue that this approach reflects the broader challenges of balancing innovation, accessibility, and intellectual property protection in AI development. This criticism highlights the complexities of navigating ethical and legal considerations in the AI industry. It also underscores the need for clearer guidelines and industry standards to address these issues effectively. The OpenAI-DeepSeek case serves as a reminder of the growing challenges associated with protecting intellectual property in the fast-moving AI landscape. As AI systems become more advanced and their outputs more valuable, disputes over data usage, ethics, and intellectual property are likely to become more frequent and contentious. This case highlights the importance of fostering a balanced approach that supports innovation while respecting ethical and legal boundaries. By establishing clear guidelines, robust enforcement mechanisms, and collaborative efforts across the industry, the AI community can address these challenges effectively and ensure the responsible development of fantastic technologies.

DeepSeek forces a rethink of China's ability to innovate

Even after some time to calm down, experts believe that the artificial intelligence model released by the Chinese startup DeepSeek last month has "forever altered the trajectory of the global rivalry in tech." That's the conclusion of Hong Kong venture capitalist Jennifer Zhu Scott, writing in the Financial Times over the weekend. DeepSeek's success has raised fundamental questions about the prevailing research model -- which relies on vast sums of money and huge amounts of compute -- and shattered assumptions about China's ability to innovate. It's also being seen as proof of the failure of U.S. policy to restrict China's access to the most high-end chips to prevent just this sort of advance. The first reckoning is long overdue; the second conclusion is premature. DeepSeek burst into the headlines last month when it released a paper detailing an AI model that could automatically learn and improve itself at a fraction of the cost of giants like OpenAI and Meta, that were, until that moment, dominating the AI industry. More ominously for the industry, DeepSeek-R1, the company's latest large language model, performed as well as that of those behemoths. The company's mobile app quickly rocketed to the top of many app stores around the world

DeepSeek vs. ChatGPT: Here's what critics are saying

ChatGPT has long since been the one to beat in the world of AI chatbots, but the competition is heating up. The newest entrant into the world of ChatGPT competitors is DeepSeek, a surprise startup out of China that has already effectively knocked $600 billion off of Nvidia's valuation. All that begs the question: what exactly is DeepSeek, and why is it already being billed as a rival to OpeanAI? Here's what early reviewers have to say. Mashable says "DeepSeek could dethrone OpenAI's ChatGPT," citing the major price difference as one of the biggest reasons why. OpenAI has a variety of pricing options for ChatGPT. For the most basic prompts, you can use the free version of ChatGPT, but it's highly limited. Alternatively, OpenAI's paid personal plans include ChatGPT Plus for $20/month and ChatGPT Pro for $200/month. DeepSeek is completely free to use online via its web portal or on mobile (with both Android and iOS apps available). When it comes to price per million tokens, DeepSeek also has ChatGPT beat. OpenAI currently charges $7.50 per million tokens for its o1 model, while DeepSeek costs a mere 14 cents per million tokens at its lowest level. This API price model significantly lowers the cost of AI for businesses and developers. Tom's Guide recently pitted DeepSeek against ChatGPT with a series of prompts, and in almost all seven prompts, DeepSeek offered a better answer. The only task ChatGPT performed better was programming-related request, which prompted the user to edit code if needed, something DeepSeek didn't do. Of course, AI chatbots can give different answers depending on how exactly you word a prompt, but the general consensus is that DeepSeek offers more reliably thorough responses than ChatGPT. However, several users have reported that DeepSeek refers to itself as ChatGPT, including X user Lucas Beyer. In a conversation between TechCrunch and Mike Cook, a research fellow at King's College London who specializes in AI, he backs up these claims by saying, "Obviously, the model is seeing raw responses from ChatGPT at some point, but it's not clear where that is." AP News also points out that DeepSeek answers sensitive questions about China differently than ChatGPT, a concerning comparison that's worth a read. Fortune writes, "DeepSeek just flipped the AI script in favor of open-source," and many critics agree. As a result of its highly sought-after, open-source nature, Gizmodo reports that "DeepSeek's releases have sent shockwaves through the U.S. stock market." The launch of DeepSeek's new model caused dips for Nvidia, Microsoft, Alphabet (Google's parent company), and more, according to Reuters. Compare DeepSeek's open-source nature to OpenAI's ChatGPT, a model that was originally meant to be open-source. A Redditor points out that OpenAI's company name is misleading, since "OpenAI" implies a company is trying to work towards being open-source, and that's something OpenAI is no longer trying to do. With DeepSeek in the picture, OpenAI may not be able to continue its closed-source approach much longer.

I agree with OpenAI: You shouldn't use other peoples' work without permission

ChatGPT developer OpenAI and other players in the generative AI business were caught unawares this week by a Chinese company named DeepSeek, whose open-source R1 simulated reasoning model provides results similar to OpenAI's best paid models (with some notable exceptions) despite being created using just a fraction of the computing power. Since ChatGPT, Stable Diffusion, and other generative AI models first became publicly available in late 2022 and 2023, the US AI industry has been undergirded by the assumption that you'd need ever-greater amounts of training data and compute power to continue improving their models and get -- eventually, maybe -- to a functioning version of artificial general intelligence, or AGI. Those assumptions were reflected in everything from Nvidia's stock price to energy investments and data center plans. Whether DeepSeek fundamentally upends those plans remains to be seen. But at a bare minimum, it has shaken investors who have poured money into OpenAI, a company that reportedly believes it won't turn a profit until the end of the decade. OpenAI CEO Sam Altman concedes that the DeepSeek R1 model is "impressive," but the company is taking steps to protect its models (both language and business); OpenAI told the Financial Times and other outlets that it believed DeepSeek had used output from OpenAI's models to train the R1 model, a method known as "distillation." Using OpenAI's models to train a model that will compete with OpenAI's models is a violation of the company's terms of service. "We take aggressive, proactive countermeasures to protect our technology and will continue working closely with the US government to protect the most capable models being built here," an OpenAI spokesperson told Ars. I'm not here to say whether the R1 model is the product of distillation. What I can say is that it's a little rich for OpenAI to suddenly be so very publicly concerned about the sanctity of proprietary data.

Microsoft and OpenAI investigate whether DeepSeek illicitly obtained data from ChatGPT

Microsoft and OpenAI are probing whether a group linked to the Chinese AI startup DeepSeek accessed OpenAI's data using the company's application programming interface without authorization, reports Bloomberg, citing its sources familiar with the matter. A Financial Times source at OpenAI said that the company had evidence of data theft by the group. Meanwhile, U.S. officials suspect DeepSeek trained its model using OpenAI's outputs, a method known as distillation. Microsoft's security team observed a group believed to have ties to DeepSeek extracting a large volume of data from OpenAI's API. The API allows developers to integrate OpenAI's proprietary models into their applications for a fee and retrieve some data. However, the excessive data retrieval noticed by Microsoft researchers violates OpenAI's terms and conditions and signals an attempt to bypass OpenAI's restrictions. The probe comes after DeepSeek launched its R1 AI model. The company claims R1 matches or exceeds leading models in areas like reasoning, math, and general knowledge while consuming considerably fewer resources. Following DeepSeek's announcement, Alphabet, Microsoft, Nvidia, and Oracle experienced a collective market loss of nearly $1 trillion. Investors reacted to concerns that DeepSeek's advancements could threaten the dominance of U.S. firms in the AI sector. However, if it turns out that DeepSeek used data illicitly obtained data from others, this will explain how the company managed to achieve its results without investing billions of dollars. David Sacks, the U.S. government's AI advisor, stated there was strong evidence that DeepSeek used OpenAI-generated content to train its model through a process called distillation. This method allows one AI system to learn from another by analyzing its outputs. Sacks did not provide specific details on the evidence, though. Neither OpenAI nor Microsoft provided an official statement on the investigation. DeepSeek and High-Flyer, the hedge fund that helped launch the company, did not respond to Bloomberg's requests for comment. However, in a statement published by Bloomberg and the Financial Times, Open AI acknowledged that China-based companies tend to distill models from American companies and that it does its best to protect its models. "We know PRC based companies -- and others -- are constantly trying to distill the models of leading US AI companies," a statement by Open AI reads. "As the leading builder of AI, we engage in countermeasures to protect our IP, including a careful process for which frontier capabilities to include in released models, and believe as we go forward that it is critically important that we are working closely with the U.S. government to best protect the most capable models from efforts by adversaries and competitors to take U.S. technology.

The Real Lesson of DeepSeek

Chinese companies are good at doing more with less -- and at using any means necessary. When the upstart Chinese firm DeepSeek revealed its latest AI model in January, Silicon Valley was impressed. The engineers had used fewer chips, and less money, than most in the industry thought possible. Wall Street panicked and tech stocks dropped. Washington worried that it was losing ground in a vital strategic sector. Beijing and its supporters concurred: "DeepSeek has shaken the myth of the invincibility of U.S. high technology," one nationalist commentator, Hu Xijin, crowed on Chinese social media. Then, however, OpenAI, which operates ChatGPT, revealed that it was investigating DeepSeek for having allegedly trained its chatbot using ChatGPT. China's Silicon Valley-slayer may have mooched off Silicon Valley after all. Which DeepSeek is the real DeepSeek? The plucky innovator or the unethical swindler? The answer is both. Chinese companies have proved to be skillful inventors, capable of competing with the world's best, including Apple and Tesla. And they have also proved adept at copying and stealing technology they don't have, then turning it against the rivals that created it. Making a product on the cheap is much easier when you don't have to invest in developing it from scratch. Read: China's DeepSeek surprise "The old narrative was that China cannot innovate but can only copy," Gregory Allen, the director of the Wadhwani AI Center at the Center for Strategic and International Studies, told me. "Now China can both innovate and copy, and benefit from both." DeepSeek's model has genuinely inventive elements, some of which Silicon Valley engineers will surely study for features to adopt. The Chinese company has wrung new efficiencies and lower costs from available technologies -- something China has done in other fields. Jimmy Goodrich, a senior adviser to the Rand Corporation who specializes in Chinese technology, told me that the same country that has excelled at reducing manufacturing costs and improving factory operations, in other words, has in this case figured out "how to crunch a lot of data at faster speeds with a smaller number of computers." But then DeepSeek may have gone a step further, engaging in a process known as "distillation." In essence, the firm allegedly bombarded ChatGPT with questions, tracked the answers, and used those results to train its own models. When asked "What model are you?" DeepSeek's recently released chatbot at first answered "ChatGPT" (but it no longer seems to share that highly suspicious response). What DeepSeek is accused of doing is nothing like hacking, but it's still a violation of OpenAI's terms of service. And if DeepSeek did indeed do this, it helped the firm to create a competitive AI model at a much lower cost than OpenAI. DeepSeek did not respond to a request for comment. The implications of what DeepSeek has done could ripple through the industry. What's the point of investing tens of millions in an AI model if a competitor (Chinese or otherwise) can simply rip it off? But the story of DeepSeek also reveals just how much Chinese technological development continues to depend on the United States. DeepSeek used chips from the U.S. giant Nvidia to create its model, and, as it turns out, may have also tapped American data to train it. Washington can capitalize on that advantage to choke off Chinese tech firms. Read: The DeepSeek wake-up call Denying China the fruits of the most cutting-edge American research has been at the core of U.S. competitive strategy. Beginning in late 2022, for instance, the Biden administration effectively barred U.S. companies from selling Chinese firms the most advanced chips. DeepSeek acquired its chips before the controls kicked in. (U.S. officials are also investigating whether the company may have bought banned chips through third parties in Singapore.) In an interview last year, DeepSeek's founder, Liang Wenfeng, admitted that "the problem we face has never been money, but the embargo on high-end chips." The firm limited new users last week because, it said, of the threat of hacking -- but the system also may not have the capacity to handle a deluge of curious customers. China's government and chip industry are racing to replace barred U.S. semiconductors with homegrown alternatives, but the technology is complicated and they are struggling to catch up. That's why China's leader, Xi Jinping, personally pressed President Joe Biden for relief from the controls. Now DeepSeek's success may frighten Washington into tightening restrictions even further. Members of Congress have already called for an expansion of the chip ban to encompass a wider range of technologies. But as much as the story of DeepSeek exposes the dependence of Chinese technology on American advances, it also suggests that stopping the transnational flow of technological goods and know-how may take more than export restrictions. DeepSeek's engineers found ways to overcome Washington's efforts to stymie them and showed that they could and would do more with less, compensating for scarcity with creativity -- and by any means necessary. The implication for the United States, Weifeng Zhong, a senior adviser at the America First Policy Institute, told me, is that "you really have to run much faster, because blocking may not always work to prevent China from catching up." That could mean securing semiconductor supply chains, cultivating talent through education, and wooing foreign experts through targeted immigration programs. Because the tech war is, at its heart, a talent contest, Washington might even consider awarding green cards to Chinese engineers who graduate from U.S. universities, so as to get them working for Silicon Valley companies rather than DeepSeek. Read: DeepSeek's chatbot has an important message President Donald Trump may be heading in a different direction. He has sharply criticized the CHIPS Act, passed in 2022, which provides government financial support for strengthening the semiconductor industry in the United States, and instead favors slapping tariffs on chips from Taiwan. He has threatened to close the Department of Education. And a recent spat between Tesla's founder, Elon Musk, and MAGA loyalists over visas for foreign specialists showed that elements of the Republican coalition are too opposed to immigrants to attract the talent that Silicon Valley requires. On this one, Trump took Musk's side in favor of the visa program. Whatever the United States chooses to do with its talent and technology, DeepSeek has shown that Chinese entrepreneurs and engineers are ready to compete by any and all means, including invention, evasion, and emulation. Maybe DeepSeek is the great Chinese tech disrupter it has been touted to be. Or maybe that will be the next big Chinese tech company, or the next one.

OpenAI accuses DeepSeek of data theft, internet calls out hypocrisy

According to the Financial Times, OpenAI believes DeepSeek may have "distilled" knowledge from ChatGPT, potentially violating the company's terms of service. "The issue is when you [take it out of the platform and] are doing it to create your own model for your own purposes," a source close to OpenAI told the FT. OpenAI CEO Sam Altman has previously acknowledged that training advanced AI models requires copyrighted materials, stating it would be impossible to develop such systems without doing so. Almost immediately, observers on social media called out what they perceived as hypocrisy from OpenAI. One Bluesky user wrote, "Funny on OpenAI, who totally didn't gobble up our data without asking, is now pointing fingers at DeepSeek for doing the exact same thing." Ed Zitron, an AI critic, commented, "I'm so sorry I can't stop laughing. OpenAI, the company built on stealing literally the entire internet, is crying because DeepSeek may have trained on the outputs from ChatGPT." Meanwhile, David Sacks, the White House artificial intelligence czar, also addressed the matter on Tuesday during an interview with Fox News. "There's substantial evidence that what DeepSeek did here is they distilled knowledge out of OpenAI models, and I don't think OpenAI is very happy about this," Sacks said without specifying the details behind that assertion. What legal or policy consequences could follow remains unclear, but Sacks noted the need to examine how companies train their models closely.