**DeepSeek: The Open-Source AI Transforming the Landscape**
Out of nowhere, **[DeepSeek](https://mashable.com/category/deepseek)** has become the hot topic among industry insiders. Its **[R1 model](https://mashable.com/article/deepseek-ai-chatgpt-rival-what-to-know)**, an open-source AI platform, has surfaced as a significant competitor to ChatGPT. Additionally, it is said to have been trained at a mere fraction of the expenses associated with other top AI models, providing performance that is comparable—if not better.
This revolutionary mix of cost-effectiveness and efficiency has reverberated through Wall Street, **[leading to a sharp decline in tech stocks](https://mashable.com/article/deepseek-ai-stocks-market-impact)** and compelling investors to rethink the financial frameworks underpinning AI innovation. According to **[DeepSeek engineers](https://arxiv.org/html/2412.19437v1)**, R1 was trained using 2,788 GPUs, costing about $6 million. In stark contrast, OpenAI’s GPT-4 supposedly necessitated a jaw-dropping $100 million to develop.
DeepSeek’s advancement calls into question the entrenched notion that larger models and vast datasets inherently result in superior AI performance. Amid the excitement about its abilities, its potential challenge to AI titans like OpenAI, and the uncertainty among investors, experts are stepping forward to shed light on the situation.
—
### **DeepSeek Disrupts the “Bigger is Better” Paradigm**
With trade limitations and restricted access to Nvidia GPUs, the DeepSeek team based in China had to think outside the box to develop and train R1. Their success in delivering a high-performing model for only $6 million—a relatively minor amount in the AI sector—has left investors astounded.
Nevertheless, AI specialists are not shocked. Timnit Gebru, a notable AI researcher who gained notoriety after being dismissed by Google for raising alarms about AI bias, expressed her perspective on **[X (formerly Twitter)](https://x.com/timnitGebru/status/1883988686259527834)**: “At Google, I inquired why they were so focused on creating THE LARGEST model. What’s the goal behind that size? What did you expect to gain from not having THE LARGEST model? Their response was to terminate my employment.”
Along similar lines, Sasha Luccioni, climate and AI head at **[Hugging Face](https://huggingface.co/)**, pointed out how the AI sector’s dependence on marketing and hype has left it susceptible to disruption. **[She remarked](https://x.com/SashaMTL/status/1884016604314222921)**, “It’s astonishing that merely suggesting a single (high-performing) LLM can reach that performance without excessively utilizing thousands of GPUs is sufficient to trigger this.”
—
### **Why DeepSeek R1 Represents a Shift in the Status Quo**
DeepSeek R1 has showcased performance on par with OpenAI’s o1 model across critical benchmarks, such as mathematics, programming, and general knowledge assessments. While its results were occasionally superior, matched, or slightly underperformed compared to o1, its true value lies in its cost-effectiveness.
“It’s not inherently more intelligent than earlier models; it was just trained more economically,” **[stated](https://x.com/GaryMarcus/status/1884223922826330300)** AI researcher Gary Marcus.
The ability of DeepSeek to produce a competitive model with such limited resources is noteworthy. Andrej Karpathy, co-founder of OpenAI, **[commented](https://x.com/karpathy/status/1872362712958906460)**, “Does this imply that you no longer require vast GPU clusters for leading LLMs? No, but you must ensure utilization is not wasteful. This highlights that there is still considerable room for optimization concerning data and algorithms.”
Wharton professor Ethan Mollick **[added](https://x.com/emollick/status/1884240997938503916)** that while DeepSeek R1 may not be categorically superior to models like o1 or Claude, its availability enables more individuals to engage with the possibilities of advanced AI. “DeepSeek is an impressive model, but it doesn’t necessarily surpass models like o1 or Claude,” he clarified. “However, because it’s freely accessible and attracting significant attention, many users of free ‘mini’ models are encountering what an early 2025 reasoning AI can achieve and are finding it astonishing.”
—
### **A Triumph for Open-Source AI**
The emergence of DeepSeek R1 marks a significant success for advocates of open-source AI, who contend that making powerful models accessible encourages transparency, fosters innovation, and stimulates healthy competition. Yann LeCun, chief AI scientist at Meta, **[remarked](https://