DeepSeek-R1: A Toyota moment

Insights

• DeepSeek-R1 has been likened to the 1957 Sputnik moment, leapfrogging the US with higher performance, lower cost AI.
• However, its arrival is more akin to how the Japanese reshaped the American car industry through leaner, more efficient manufacturing.
• The open-source LLM uses some clever engineering tricks in both training and inference.
• The model leads to significant energy savings but comes with increased security and privacy risks.
• Organizations should experiment with this technology by setting up a central governance body and recruiting talent to test how it might benefit large enterprise.

The Chinese-made LLM, DeepSeek- R1, which was created at speed and boasts reduced costs and increased performance compared with similar size models from OpenAI and Google, has been likened to the 1957 Sputnik moment, when the Soviet Union leapfrogged the United States in the space race and launched the first satellite.

This metaphor has its backers, as the Chinese release has dented the invincibility aura surrounding US tech companies. This has led some, including Marc Andreessen of a16z, to advocate for a change in the US policy: Instead of limiting access to US chips, as former President Joe Biden favored throughout the 2020s US-China chip war, investor confidence would be better boosted by weakening export controls and letting China have the chips without too much trouble.

However, as argued in this Cambridge University article, this would be like the USSR suddenly freaking out and deciding to sell their world-class rocket-making equipment to the US as well as give all its secrets away for free. Not a good idea.

The arrival of DeepSeek is akin to how the Japanese reshaped the American car industry by making reliable cars at lower costs through leaner, more efficient manufacturing. The introduction of the Toyota production line, and the culture of just-in-time, minimal-waste inventory, are similar to the way in which DeepSeek’s researchers seem to have created an AI model that can be run by anyone on modest hardware and on a shoestring budget.

The AI race will be won by those who commercialize and deploy AI most effectively — not necessarily those with the most powerful models.

The innovation of R1

So, what is DeepSeek-R1?

DeepSeek-R1 is an open-source LLM built at seemingly a fraction of the price ($5.6 million) of OpenAI’s 01 model, but which exhibits similar, if not superior, performance on a range of reasoning, math, and coding tasks.

But how did a few hundred researchers and a budget of just a few million dollars produce a model as capable as those from Google and Open AI, which have billions to spend?

R1 uses some clever engineering tricks in both training and inference. First, instead of just using reinforcement learning with human feedback, as in most other LLMs, the team of Chinese engineers deployed a technique used by Demis Hassabis’s DeepMind when it beat Lee Sedol at the Chinese game of Go. Here, a more automated version of reinforcement learning is used: Researchers show the AI what success looks like, and the deep learning model figures out the rules in a form of trial and error. So saying, R1 was trained on vast amounts of text from the internet and then left to work out how to reason and make decisions by itself.

Another innovation is distillation. Here, the reasoning capabilities from a large model are distilled into smaller models, reducing the number of parameters, and therefore demanding less compute, driving down energy requirements in the training phase. Hence, the reaction from the markets: Nvidia’s stock nosedived by $600 billion before righting itself, since less compute means less need for high-performance chips, and hence the characterization of the launch as a Sputnik moment.

Perhaps the coolest innovation, though, and this ties to energy requirements too, is that R1 uses a mixture-of-experts approach. This algorithm increases performance by assigning subsystems of the AI to different tasks, activating only 37 billion out of 671 billion parameters during operation. Because each subsystem is optimized for the task, it can do much better than a generalized model from, say, OpenAI. This also uses less compute, driving down energy requirements further. And because the whole system is so small, inference can be done on a single GPU, even on a single laptop, removing the need for enormous GPU clusters provided by Google or other hyperscalers.

The fly in the ointment

However, this is not the full story. While DeepSeek-R1 might lead to significant energy savings and lower climate burden, there are some drawbacks to using this AI.

First, because the system uses chain-of-thought reasoning, breaking down the problem into a series of steps, with each requiring more compute, inference costs can actually go up, making it inefficient to query as a chatbot or for very simple prompting tasks such as document summarization or code compilation.

Second, and most importantly, there are some serious safety and privacy concerns with DeepSeek’s model.

A recent study from Cisco, in partnership with the University of Pennsylvania, also found that R1 is vulnerable to jailbreaking (prompting the model to bypass safety and ethical guidelines), with some tests showing that it fails 100% of the time using certain prompts. The model is also prone to generating harmful, toxic, and biased content, and could be recruited for terrorist activities.

As well as the national security risks that implies, not to mention other concerns around toxic content, data privacy is also an issue. Several countries, including Australia and Italy, as well as a number of US agencies, have banned R1, thanks to concern about DeepSeek’s ethics, security practices and approach to privacy. All user data gathered by the model, including personal and commercially sensitive information, is stored in China, raising compliance issues with regulations such as GDPR in Europe and CCPA in the US.

DeepSeek’s privacy policy states that user inputs may be stored indefinitely, used for training, and shared with advertising partners; and China’s National Intelligence Law requires companies to cooperate with state intelligence agencies, potentially allowing access to corporate data.

Experimentation, talent, and greater governance

Although R1 throws up significant issues that mean using it is a clear risk for businesses, it also highlights problems enterprises face in pressing ahead with AI built on US technology, including the high costs of preparing for and deploying AI.

Organizations will want to use the technology to execute at scale, much like the US and Europe did after Toyota redefined how cars were built.

The lesson from DeepSeek is that AI can be done cheaper and more efficiently. However, organizations must focus on innovating responsibly and effectively.

There are three low-risk actions organizations can take to move ahead:

1. Innovate with an AI foundry/factory: This is a centralized experimentation hub where models like R1 can be tested safely. Small wins for specific use cases can help people accept and trust AI, as Sunil Senan, senior vice president for AI at Infosys, says: “Small wins through AI use cases are an important step towards enterprise AI.” The foundry (to experiment) and factory (to deploy at scale) form the hub of AI innovation, experimentation, value realization, workforce engagement, and source of governance for an enterprise in the early stages of AI deployments. This hub isn’t just a novelty; it’s the engine of AI transformation in a business.
2. Get the talent, and get it fast: Most organizations lack AI talent and are in no position to test either DeepSeek’s R1 or comparable models. Recent research from Infosys found that just 35% have sufficient talent to test models such as R1.
3. Create a centralized AI governance task force: At Infosys, the responsible AI office acts as the custodian of new technology like DeepSeek-R1. Infosys’s responsible AI office governs use of AI across the organization and ensures that all AI products and services are responsible by design. With the privacy and security challenges posed by R1, the task force should work to support AI acceptance — ensuring AI ethics are not overshadowed by business imperatives.

Moving ahead

The arrival of DeepSeek has focused minds on new ways of doing AI, and this offers clear opportunities for businesses. Recent research from Infosys found that good use of AI leads to as much as a 40% increase in productivity. Finding more efficient and effective ways to build and deploy AI in the way DeepSeek-R1 seems to offer will make AI more attractive to enterprises, which will then be able them to reap the benefits. This potential democratization of AI, thanks to cheaper, less energy-intensive ways of deploying it, is AI’s Toyota moment.