Why DeepSeek-R1 had to be open source, and why it can't beat OpenAI



This article, originally posted in Japanese on 07:00 Jan 31, 2025, may contains some machine-translated parts.
If you would like to suggest a corrected translation, please click here.

The inference models 'DeepSeek-R1-Zero' and 'DeepSeek-R1',

released by Chinese AI company DeepSeek on January 20, 2025, have the same performance as OpenAI's inference model o1, but have much lower training costs than o1, and are open sourced under the MIT license. AI technology company Lago explains why DeepSeek open-sourced its AI models.

Lago Blog - Why DeepSeek had to be open-source (and why it won't defeat OpenAI)
https://www.getlago.com/blog/deepseek-open-source



DeepSeek-R1, released by DeepSeek on January 20, 2025, is reported to have performance equal to or better than O1 while requiring only 4% of the training costs compared to O1.




'They achieved this result with a new training method that is more efficient than what OpenAI, Anthropic, and other well-funded competitors use to train their own models,' Lago said. 'Just as Coca-Cola doesn't open source its recipes, market-leading innovative products should take advantage of them. Why did DeepSeek open source the DeepSeek-R1?'

According to Lago, the reason DeepSeek open-sourced its AI model is because DeepSeek is a Chinese company. In fact, the United States has regulated the export of AI chips to China due to concerns that AI-related applications may be diverted for military use by hostile countries, including China. Due to these factors, it is inevitable that the DeepSeek-R1 will be viewed with skepticism and anxiety in the Western market. Therefore, DeepSeek is trying to gain the trust of users and developers by open-sourcing its AI model. Regarding the open-sourcing of such models, Lago said, 'Open-sourcing is not just a technical action, it is also a cultural action to solidify a foothold in the market.'

And because DeepSeek is subject to US export restrictions on high-performance AI chips, it has to get the most out of fewer resources. So it had to find a way to train its models more efficiently.

How did DeepSeek beat OpenAI's O1 at 3% of the cost? - GIGAZINE



On the other hand, companies with abundant funds, computing resources, and the world's largest distribution, such as OpenAI, Meta, and Google, can invest money to build models without having to find an efficient way to train models. Also, building large-scale language models more easily and cheaply would mean losing the advantages of these companies, such as their abundant funds.

In addition, in recent years, large-scale language models at the level of GPT-4 have been released one after another. In this situation, OpenAI is facing a challenge from users asking, 'Is it worth paying a large fee for the OpenAI API compared to models available in open source?' Lago points out, 'I don't know whether the AI app I'm actually using is using LLaMa, GPT, Claude, or Mistral. Why should I pay so much for the OpenAI API when the end user doesn't know the difference?'

In addition, while open source software is low-cost and highly customizable, it requires maintenance and technical knowledge. On the other hand, proprietary software can be used smoothly, but it lacks flexibility. Therefore, infrastructure always needs to be customized, and open source databases that allow engineers to audit and modify the code have become popular in recent years.

The same can be said for large-scale language models, where the development of AI products requires a high degree of customization, making open source models like DeepSeek-R1 more suitable than proprietary models such as o1. 'In infrastructure, open source ultimately wins,' Lago said.

However, according to Lago, the outlook for OpenAI, which has led the AI industry, to become obsolete is naive. OpenAI develops various AI products that serve as the basis for other AI models, such as GPT-4 and o1, and OpenAI CEO Sam Altman said, 'DeepSeek's R1 is an impressive model, especially in terms of the features it offers for the price. Obviously, we can deliver models that are better than R1, and it's really exciting to have new competitors! We'll be announcing some releases in the future.'

OpenAI CEO Sam Altman says, 'DeepSeek-R1 is great, but we're going to release even more amazing AI models' - GIGAZINE



Still, Lago said, 'DeepSeek-R1 may serve as a wake-up call for well-funded, established AI companies to find more efficient ways to train machines.'

Related Posts:

in Software, Posted by log1r_ut