DeepSeek: Why It's Important — in Layman's Terms

In the rapidly advancing world of artificial intelligence (AI), companies like OpenAI and Google’s Gemini have long been considered the industry leaders, shaping the future with their massive, cutting-edge models. However, a relatively new player, DeepSeek, is challenging these established players with a radically different approach that focuses on achieving remarkable performance with significantly fewer computational resources and costs.

Silicon Valley venture capitalist and co-founder of Andreessen Horowitz, Marc Andreesen has called DeepSeek-R1 “AI’s Sputnik moment”¹ whilst billionaire investor and founder of Bridgewater Associates Ray Dalio has called exuberance over AI a “bubble” that resembles the build up to the dotcom bust at the turn of the millenium.² While markets may fluctuate, technological progress moves forward, it’s clear that AI is already proving to be the catalyst for the next wave of global innovation.

So, is DeepSeek just another AI model, or does it offer something truly different? While traditional models from companies like OpenAI and Google require vast amounts of data and processing power, DeepSeek has demonstrated that sophisticated AI models can be trained and deployed in a more resource-efficient way. This could lead to dramatic cost reductions and more accessible AI, a game-changer for industries worldwide.

DeepSeek’s Approach

DeepSeek’s approach centres on reducing the computational cost of training and running AI models without sacrificing performance. One of the key innovations behind their R1 model is its use of reinforcement learning (RL) paired with a rule-based reward system. In contrast to traditional AI models, which often rely on supervised learning with huge datasets, DeepSeek’s models learn autonomously through reinforcement learning, receiving rewards for making accurate predictions and following structured reasoning patterns.

Whereas models from OpenAI (like GPT-4) and Google’s Gemini series depend on massive, supervised datasets and GPU-heavy training processes, DeepSeek’s models take a different route. DeepSeek’s use of RL with simpler reward structures avoids the complexity and inefficiency of traditional reward systems, which can sometimes lead to issues like reward hacking—where models figure out ways to “game” the system without improving their real-world performance.

This unique reward framework has led to chain-of-thought reasoning—a process where the model breaks down complex tasks into manageable steps, much like a human would do. This wasn’t an easy feat; in fact, chain-of-thought reasoning was a major challenge for AI models, and it was believed that only large-scale models could manage this. However, DeepSeek has proven that this kind of reasoning capability can be achieved on a much smaller scale, with models like R1 performing at the same level as more extensive models but with fewer resources.

“Aha Moments”

Another striking feature of DeepSeek’s models is what they call the “aha moment” that occurs during training. This refers to the model’s ability to self-correct its reasoning when it encounters uncertainty, essentially revising its own thought process without being explicitly programmed to do so. In practical terms, this means DeepSeek’s models can reflect on their own logic, detect errors mid-task, and try new approaches to solve problems.

^{Source: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf}

In This level of emergent behavior is a breakthrough, pushing models toward autonomous improvement—a feature that was previously difficult to achieve in AI. By encouraging this self-reflection, DeepSeek’s models show that AI can improve on its own, making the system much more adaptable and efficient.

Smarter. Cheaper. More Accessible.

One of the most significant advantages of DeepSeek’s innovations is their ability to achieve impressive performance at a fraction of the cost. For instance, on the AIME 2024 math competition, DeepSeek’s R1 model scored 79.8% accuracy,³ matching the performance of OpenAI’s O1 model, but with far fewer resources. DeepSeek’s smaller models also performed exceptionally well in programming competitions, like MATH-500, where they achieved 97.3% accuracy,⁴ beating out models many times their size.

The distilled versions of DeepSeek’s models have shown they can outperform traditional models, even those with several times more parameters. Distillation involves transferring the knowledge and reasoning capabilities of a larger, more powerful model (in this case, R1) into smaller models. This allows the smaller models to achieve competitive performance on reasoning tasks while being more computationally efficient and easier to deploy.⁵ This suggests that the true key to model performance isn’t just about size or raw data but about how efficiently the model is trained and how well it can reason through complex problems. In terms of cost, DeepSeek’s models require much less computational power, making them far cheaper to run and train compared to models from companies like OpenAI or Google, who typically need thousands of GPUs and massive datasets.

For comparison, OpenAI’s GPT-4 is a massive model that requires enormous computing resources. As of 2024, OpenAI’s costs for training and deploying GPT-4 are incredibly high, and the model’s sheer scale has made it inaccessible for many smaller companies or startups. DeepSeek’s R1 model, on the other hand, offers 95% cost savings on API calls compared to OpenAI’s offerings while maintaining similar performance.⁶ This represents a major disruption in the industry, especially given how quickly AI applications are expanding.

The Competitive Landscape

The AI industry has traditionally been dominated by large players like OpenAI, Google’s Gemini, and Meta, who have the resources to build and deploy large-scale models. However, DeepSeek’s efficient models are beginning to turn the tide in favour of smaller, more agile companies. Here’s a breakdown of how DeepSeek compares to and will influence the competition:

• OpenAI (GPT-4): OpenAI’s GPT-4 model is known for its impressive capabilities but requires significant computing resources for both training and inference. DeepSeek’s R1 model matches or exceeds GPT-4’s performance at a fraction of the cost, making it a more accessible option for companies with limited budgets.
• Google Gemini: Gemini models have made significant strides in AI, offering powerful models for language understanding and generation. However, like OpenAI, Gemini requires vast resources to train and deploy. DeepSeek’s models challenge this by demonstrating that high performance doesn’t necessarily require massive computational power. Gemini may be forced to rethink their approach as more efficient models like DeepSeek’s become more viable.
• Meta (Llama): Meta’s Llama4 models are also in development, and internal reports suggest that DeepSeek’s models are outperforming them in terms of efficiency and cost. DeepSeek’s innovations have even caught Meta off guard, as the company’s custom-built models are underperforming compared to the more efficient DeepSeek models. Meta and other AI labs are now studying DeepSeek’s open-source code and technical reports to incorporate its innovations into their own projects.⁷
• Cerebras and Groq: Companies like Cerebras and Groq are focusing on radical hardware innovations to bypass traditional compute constraints, but DeepSeek’s software-based breakthroughs show that AI models can be more efficient without needing radically new hardware. DeepSeek's efficiency breakthroughs suggest that AI software advancements can be just as impactful as hardware innovations, providing an alternative pathway to scalability.

It's important to note that Deepseek probably owes much of its success to its US counterparts, going back to the initial 2017 transformer architecture developed by Google AI researchers (which kickstarted the entire LLM craze).

DeepSeek-R1 according to the paper released by its researchers was trained on fine-tuned “dataset of DeepSeek-V3,” which was found to have many indicators of being generated with OpenAI’s GPT-4o model itself.⁸

A Paradigm Shift in AI

DeepSeek’s innovations mark a paradigm shift in AI model development. By focusing on efficiency, smarter training techniques, and cost-effective approaches to both training and inference, DeepSeek has developed models that offer comparable or superior performance to industry giants like OpenAI and Google’s Gemini. Their ability to achieve high-level reasoning and self-correction with fewer resources is a game-changer, especially as AI becomes more integrated into everyday life.

DeepSeek is proving that AI development doesn’t have to be about bigger models and bigger budgets (Sorry NVIDIA!) —it’s about smarter models and smarter approaches. As this efficiency-driven approach gains traction, DeepSeek and its approach is set to play a major role in democratizing AI technology, making it more accessible to businesses and developers around the world. This shift could ultimately reshape the future of AI, making it a more scalable, sustainable, and cost-effective resource for industries of all sizes.

Interested in the AI space? Learn more about the activities of the Aura Ventures team and explore how the AI landscape is shaping 2025.

Speak to the Aura Ventures Team

Sources:

1. Deepseek R1 is AI's Sputnik moment.
2. AI investor ‘bubble’ echoes dotcom bust: Ray Dalio
3. DeepSeek R1 Analysis
4. DeepSeek R1: Features, o1 Comparison, Distilled Models & More
5. What are DeepSeek-R1 distilled models?
6. DeepSeek API Docs and OpenAI Pricing
   
7. Meta Sets Up War Rooms to Analyze Deepseek’s Tech
8. Calm Down: DeepSeek-R1 Is Great, but ChatGPT’s Product Advantage Is Far From Over