UPSC Issue at a Glance | DeepSeek breakthrough: 4 Key Questions You Must Know for Prelims and Mains

DeepSeek breakthrough is a direct challenge to the idea that AI progress depends on enormous computational power, vast datasets, and billions in funding (Reuters illustration)

Why is this issue relevant?

The DeepSeek breakthrough is relevant for the UPSC CSE exam because artificial intelligence and emerging technologies are integral topics in General Studies Paper III. Furthermore, UPSC has previously asked questions on AI, thus knowing about the major developments in this domain is crucial. UPSC aspirants will also find this topic useful for essays and current affairs, as well as for their personality tests.

UPSC Syllabus: 

Preliminary Examination: Current events of national and international importance, Science and Technology

Mains Examination: General Studies-II,III: Awareness in the fields of IT, Space, Computers, robotics, nano-technology, bio-technology and issues relating to intellectual property rights, Government policies and interventions 

Question 1: What is DeepSeek? 

DeepSeek is a Chinese AI company located in Hangzhou, founded by entrepreneur Liang Wenfeng, who also serves as the CEO of the quantitative hedge fund High Flyer. Wenfeng began working on AI in 2019 with his company, High Flyer AI, which focuses on research in this field. 

Recently, DeepSeek launched its AI models—DeepSeek-V3 and DeepSeek-R1, a reasoning model. These models quickly gained popularity, surpassing ChatGPT to become the most downloaded app on the App Store. DeepSeek-V3 and DeepSeek-R1 compete with OpenAI’s advanced models, o1 and o3, as the Chinese lab achieved this feat only with a fraction of their investments.

Here are a few other open-source AI models developed by DeepSeek:

📍DeepSeek Coder: An open-source AI model designed for coding-related tasks.

📍DeepSeek LLM: An AI model with a 67 billion parameter count to rival other large language models (LLMs).

📍DeepSeek-V2: A low-cost AI model that boasts of strong performance.

📍DeepSeek-Coder-V2: An AI model with 236 billion parameters designed for complex coding challenges.

📍DeepSeek-V3: A 671 billion parameter AI model that can handle a range of tasks such as coding, translating, and writing essays and emails.

📍DeepSeek-R1: An AI model designed for reasoning tasks, with capabilities that challenge OpenAI’s marquee o1 model.

📍DeepSeek-R1-Distill: An AI model that has been fine-tuned based on synthetic data generated by DeepSeek R1.

In this context, a question naturally arises: What makes DeepSeek AI models unique, and how do they stand apart from other AI players? Let’s understand what sets DeepSeek apart in the evolving world of artificial intelligence!

Question 2: How is DeepSeek different from other AI players? 

DeepSeek appears to have surpassed major players like OpenAI, Google, and Meta in the competitive landscape of AI development. The lab’s newly released open-source reasoning model, DeepSeek R1, is reported to outperform leading AI models, such as OpenAI’s o1, on key math and reasoning benchmarks. Broadly, two factors make DeepSeek the talk of the town .i.e. its state-of-the-art technology and affordable cost.

State-of-the-art technology

1. Open-sourced nature: DeepSeek models are open-source, unlike the closed models from OpenAI and Google. This means that other companies, especially small developers, can build on top of DeepSeek’s model and improve it without paying license fees. The potential is huge—rather than developing their own models, companies can modify and deploy DeepSeek’s models at a fraction of the cost. This could drive mass AI adoption at scale.

Bijin Jose writes– What sets DeepSeek models apart is their performance and open-sourced nature with open weights, which essentially allows anyone to build on top of them. The DeepSeek-V3 has been trained on a meagre $5 million, which is a fraction of the hundreds of millions pumped in by OpenAI, Meta, Google, etc., into their frontier models. 

DeepSeek’s strategy of using open-source models can have a huge impact on the AI community at large, opening up the AI market and providing access to AI tools for a broad set of users, especially smaller businesses. – Anuj Bhatia

2. MoE and MLA: DeepSeek-V3 stands out due to its architecture, known as Mixture-of-Experts (MoE). In MoE models, multiple specialized models collaborate to answer questions rather than relying on a single large model to handle everything. Additionally, the model employs a new technique called Multi-Head Latent Attention (MLA), which enhances efficiency and reduces the costs of training and deployment. This enables DeepSeek-V3 to compete with some of the most advanced models available today.

3. Reinforcement Learning: DeepSeek’s success can be attributed to a concept known as reinforcement learning. This approach allows AI models to learn through trial and error, improving themselves through algorithms. It is quite similar to how humans learn from their experiences. Essentially, DeepSeek’s models learn by interacting with their environment and receiving feedback based on their actions. As a result, these AI models become better at reasoning and are capable of solving complex problems. Additionally, DeepSeek has excelled in distilling the capabilities of its large models into smaller, more efficient ones.

4. Test-Time Compute: The DeepSeek-R1 model is another offering from DeepSeek, featuring a unique capability known as test-time compute, which allows it to ‘think’ while generating responses. R1 utilizes the same Mixture-of-Experts (MoE) architecture and often matches or surpasses OpenAI’s top model in areas such as mathematics, coding, and general knowledge. Unlike OpenAI’s O1 model, which takes time to process prompts and generate optimal responses, R1 demonstrates its reasoning process in real time, revealing its chain of thought as it produces output.

Deepseek app is seen in this illustration taken on Tuesday. (Photo: Reuters)

Hence, one can see that DeepSeek has essentially delivered a state-of-the-art model that is competitive. Moreover, the company has invited others to replicate their work by making it open-source. 

More cheap and affordable 

“OpenAI is highly overvalued. I think we saw their business model sort of blow up over the last few days with DeepSeek basically giving away for free what they [OpenAI] wanted to charge money for,” Gary Marcus, a professor at New York University (NYU), said in an interview with CNBC on Tuesday, January 28.

DeepSeek’s AI models have been hailed as a research breakthrough as they demonstrate that it is possible to develop competitive, frontier AI models using less cash and fewer GPUs – as opposed to the billions of dollars spent by OpenAI, Meta, Google, Microsoft, and others to do the same.

It is generally believed that training AI models requires significant investments. However, DeepSeek has minimised the enormous costs associated with infrastructure and hardware. By utilising NVIDIA H800, which is considered an older generation of GPUs in the United States, DeepSeek has dramatically reduced the expenses related to building its AI models. In contrast, major American AI companies have opted for the more advanced NVIDIA H100 GPUs. DeepSeek, however, chose the less powerful version—NVIDIA H800, which is reported to have lower chip-to-chip bandwidth.

Additionally, the DeepSeek-R1 model is reported to be 90-95% more affordable than OpenAI’s model O1. Another important aspect of building AI models is training, which requires significant resources. According to the research paper, the Chinese AI company has only trained the essential parts of its model using a technique called Auxiliary-Loss-Free Load Balancing.

DeepSeek has undoubtedly shaken the world of AI, challenging the long-standing dominance of the US, which has led the AI race with major players like OpenAI and Google. Now, China’s DeepSeek is changing the landscape. But where does India stand in this race for AI dominance? More importantly, what can India learn from the DeepSeek breakthrough to shape its own AI future?

Question 3: What are the lessons for India from the DeepSeek breakthrough?

DeepSeek’s AI models have not only given Western AI giants a run for their money but also sparked fears that the US may struggle to maintain its AI primacy in the face of a brewing tech cold war with China.- Karan Mahadik

DeepSeek’s technological achievement has stunned the world, from Silicon Valley to the global AI stage. However, China’s growing dominance in AI raises critical questions about India’s position, especially given the lack of an AI lab or startup that rivals the capabilities of OpenAI or DeepSeek. In this context, India can draw key lessons from the DeepSeek breakthrough:

1. AI is everyone’s game: DeepSeek’s breakthrough in the AI field demonstrates that if foundational AI models can be trained cost-effectively, it lowers the barriers for nations eager to develop their own models. By reducing the fixed costs associated with building these models, resource-constrained countries can better compete, despite challenges such as limited GPU availability and insufficient funding for both foundational models and the required data. This means that AI is an accessible domain for everyone. Regardless of which country leads in this area, any nation can reap the benefits of the AI race.

Sarjan Shah writes- As the Chinese breakthrough shows, necessity is indeed the mother of invention. By proving that progress does not depend on massive resources, this development offers hope that AI can be a tool for everyone — not just the few with billions to spend. It’s a reminder that in the end, intelligence — whether artificial or human — is about thinking differently.

2. Wealth Builds, But Wisdom Innovates: While AI companies typically require billions in investments to train their models, DeepSeek’s innovation showcases the effective use of limited resources. This proves that groundbreaking advancements are not solely dependent on funding but also on vision and adaptability. It also reinforces how necessity can drive innovation in unexpected ways. Thus, promoting research and development is key to resourceful innovation.

DeepSeek is a reminder that money buys many things, but it is not everything; certainly not love or the capacity for innovation. If the US is seen as pouring billions of dollars into building more computing power and better AI models, DeepSeek has shown it is possible to do more with less.- C. Raja Mohan

C. Raja Mohan writes-  That should give hope for India and other middle powers like France. The gap between the US and China is much less than that between the two of them and the rest.  While the middle powers can’t keep pace with the US and China, they could do enough to stay in the AI  game.

3. AI Diplomacy-Navigating Partnerships for Progress: The Chinese breakthroug in AI represent not only a technological breakthrough but also a significant geopolitical development. In this context, India should be open to collaboration with other countries to leverage the benefits of advancements in AI.

C. Raja Mohan writes- The Biden Administration was open to partnership with India, but concerns about leakage of technology from India to Russia had put a dampener on the kind of access the US is willing to offer India on AI chips. For Delhi, the time is now to decide on how much weight it is willing to give Russia in advancing its interests with the US on the AI game.

Another important lesson from the DeepSeek breakthrough is that the Chinese company achieved this milestone despite sanctions that limited their access to advanced chips and cutting-edge hardware. So, here’s a question—what is the different country’s approach to regulating AI?

Question 4: How are countries regulating AI?

With the advancement of AI technology, governments and policymakers around the world are increasingly focused on artificial intelligence. Nations appear to be in a relentless competition to outpace each other, believing that those who fall behind will ultimately be the losers. However, many concerned voices among them have valid reasons for their worries.

Geoffrey Hinton, a pioneer in the field of AI, has highlighted the potential for AI to surpass human intellectual abilities. He has said that AI “will be comparable to the industrial revolution,” but he has also warned that “we also have to worry about a number of possible bad consequences, particularly the threat of these things getting out of control.” Thus, it is essential to remember how disruptive new technologies can be for societies and economies.

The concerns regarding advancements in AI can be categorized into three main areas: privacy, system bias, and violations of intellectual property rights. Interestingly, policy responses to these issues vary across different jurisdictions.

The European Union has taken a notably stricter approach by proposing regulations that classify AI based on specific use case scenarios, assessing them according to their level of invasiveness and risk. In contrast, the UK adopts a decidedly ‘light-touch’ approach, aiming to encourage innovation rather than hinder it in this emerging field.

The United States’ approach is positioned somewhere in between these two extremes, with indications of potential further deregulation. China has also introduced its own measures to regulate AI.

India has emphasized the need to address the challenges posed by the weaponization of social media, advocating for steps to ensure that AI promotes safety and trust, even as the technology presents significant opportunities.

On the global stage, 29 countries and European Union, including the United States, the United Kingdom, and China, have agreed to the Bletchley Declaration. This marks the first global agreement aimed at addressing the risks associated with advanced artificial intelligence (AI). The declaration lays out plans for greater transparency from AI developers regarding safety practices and more scientific collaboration on understanding AI’s risks.

Prime Minister Narendra Modi is set to co-chair the Paris AI Summit and has accepted the invitation to travel to France. PM Modi with French President Emmanuel Macron. (File photo)

The regulation of AI is continuously evolving in response to advancements in the field. Each country’s approach to AI regulation is something we should closely watch, especially as the AI landscape becomes more fascinating than ever. By the way, have you heard about Kimi K1.5? What are your thoughts on it? Let us know in comment section. 

Post Read Questions

Prelims

(1) With the present state of development, Artificial Intelligence can effectively do which of the following? (UPSC CSE 2020)

1. Bring down electricity consumption in industrial units

2. Create meaningful short stories and songs

3. Disease diagnosis

4. Text-to-Speech Conversion

5. Wireless transmission of electrical energy

Select the correct answer using the code given below: 

(a) 1, 2, 3 and 5 only

(b) 1, 3 and 4 only

(c) 2, 4 and 5 only

(d) 1, 2, 3, 4 and 5

(2) Consider the following statements with regard to the Bletchley Park Declaration:

1. The declaration is signed by 29 countries and European Union.

2. Frontier AI is defined as highly capable foundation generative AI models that could possess dangerous capabilities that can pose severe risks to public safety.

3. The United States, China, Japan, the United Kingdom, France, and India are not signatories to the declaration.

How many of the statements given above is/are correct?

(a) Only one

(b) Only two

(c) All three

(d) None

Mains

What are the main socio-economic implications arising out of the development of IT industries in major cities of India? (UPSC CSE 2021)

(Sources: Deepseek: How open-source AI is disrupting big tech’s monopoly , How DeepSeek’s origins explain its AI model overtaking US rivals ,DeepSeek’s rise could make OpenAI the WeWork of AI, Liang Wenfeng, Is this China’s ChatGPT moment and a wake-up call for the US?, What’s on the agenda of the Paris AI summit, DeepSeek’s Sputnik moment, In DeepSeek breakthrough, lessons for India)

If you missed last week’s UPSC Issue at a Glance | India-Taliban Talks: 4 Key Questions Answered for your UPSC Exams from the Indian Express, read it here.

For your queries and suggestions write at roshni.yadav@indianexpress.com

🚨New Year Special: Click Here to read the January 2025 issue of the UPSC Essentials monthly magazine. Share your views and suggestions in the comment box or at manas.srivastava@indianexpress.com🚨

Subscribe to our UPSC newsletter and stay updated with the news cues from the past week.

Stay updated with the latest UPSC articles by joining our Telegram channel – Indian Express UPSC Hub, and follow us on Instagram and X.