Article illustration 1

When DeepSeek released its R1 AI model in January, the ripple effects were immediate: U.S. stock markets dipped as the world took notice of a Chinese contender capable of outperforming established players in complex reasoning tasks like mathematics and coding. Now, in a landmark peer-reviewed paper published in Nature, DeepSeek has not only detailed R1's inner workings but also debunked speculation that it relied on copying outputs from rival large language models (LLMs). Instead, the model achieved its prowess through a novel application of pure reinforcement learning—a trial-and-error approach where it taught itself reasoning strategies without human-guided examples.

The $294K Revolution in AI Training

R1's most staggering revelation is its cost efficiency. While competitors like OpenAI's GPT-4 and Google's Gemini reportedly required tens of millions of dollars to train, DeepSeek spent just $294,000 refining R1 atop a base model that cost approximately $6 million. This frugality is even more remarkable given the geopolitical constraints: training occurred on Nvidia's H800 chips, which were banned from export to China in 2023 under U.S. sanctions. Despite this, R1 emerged as the most popular open-weight model on Hugging Face, with 10.9 million downloads to date, allowing developers worldwide to experiment with its capabilities for free.

How Pure Reinforcement Learning Unlocked Autonomous Reasoning

At R1's core is a radical departure from conventional LLM training. Rather than feeding the model human-curated examples (a method prone to bias and inefficiency), DeepSeek employed pure reinforcement learning. Here, the AI was rewarded only for correct answers, forcing it to devise its own reasoning tactics—such as internal verification of steps—without predefined templates. To optimize this, the team introduced group relative policy optimization, where R1 scored its own attempts instead of relying on an external algorithm. This not only accelerated learning but also reduced computational overhead. As Huan Sun, an AI researcher at Ohio State University, noted: "Almost all work in 2025 so far on reinforcement learning in LLMs has been inspired by R1."

A Watershed Moment for AI Peer Review

R1's publication in Nature marks the first peer-reviewed disclosure of a major LLM, setting a critical precedent for transparency in an industry often criticized for opacity. Reviewers compelled DeepSeek to clarify technical details, including data sources and safety protocols, while reducing anthropomorphic language that could mislead about the model's capabilities. Lewis Tunstall, a machine-learning engineer at Hugging Face who reviewed the paper, emphasized: "If we don't have this norm of sharing the process publicly, it becomes very hard to evaluate risks." This rigor addresses growing concerns about AI ethics and reproducibility, urging other firms to follow suit.

DeepSeek's app interface, showcasing its accessibility to developers and researchers. Credit: David Talukdar/ZUMA via Alamy

The implications extend beyond academia. For engineers, R1's efficiency proves high-performance AI needn't be exorbitant, potentially democratizing access for startups and researchers. Its self-directed learning approach also hints at future models that can adapt to niche tasks without massive labeled datasets. As the AI arms race intensifies, DeepSeek's blend of innovation, affordability, and transparency might just be the blueprint for a more accountable era in artificial intelligence—one where open science and ingenuity trump sheer resource expenditure.

Source: Nature Article, "AI can learn to show its workings through trial and error."