It has been little over a month since the release of DeepSeek R1. I can't tell if it is a long time back or just yesterday when Nvidia lost ~15% of its market capitalization in 1 day! It feels a little bit of both.
The hoopla around DeepSeek has given way to a more measured response from experts and markets in the field. Since January 27, a number of other LLMs have released an updated version as well that have had impressive performance on many leaderboards! It is tough to keep up with all the advances.
In this slew of releases which is overwhelming, I believe there is a silver lining.Increasingly, it is hard to distinguish the performance of these models.
Clearly there is a broad range of options and we are all better off for that but the trend is unmistakable: the models will not be the most differentiated asset. Instead, the focus will shift to other aspects.
For example, is it closed source, can it run locally, is it cheap to run?
And I find that to be a silver lining that bodes well for the producers and consumers of these models.
To be fair, the release of DeepSeek R1 did reset this narrative. It is both cheaper than more celebrated models like OpenAI etc, though this claim has been disputed and at the same time other groups have claimed to be able to build even cheaper models.
Jiayi Pan, a PhD candidate at the University of California, Berkeley, claimed that he and his AI research team have recreated core functions of DeepSeek's R1-Zero for just $30. Irrespective, it is a profoundly welcome change.
The second aspect that I believe will have profound consequence is that DeepSeek R1 is open source. The reaction of the researchers and developers have been overwhelmingly positive. These two aspects, I suspect will have a more profound impact than the model itself in the more immediate term.
Alongside its cost-saving advantages, DeepSeek is also drawing significant regulatory attention. Privacy watchdogs in countries such as Ireland, France, Belgium, and the Netherlands have raised concerns about its data collection practices—specifically, that user data is stored on servers in China, where local laws may require data sharing with intelligence agencies.
In the U.S., several federal agencies have advised their employees against using DeepSeek, and many companies have asked their cybersecurity teams to block access to the app. Additionally, governments in Italy, Taiwan, Australia, and South Korea have imposed bans or restrictions on DeepSeek’s technology in public sector settings due to security and data privacy concerns.
It’s important to note that these concerns are mainly associated with the application layer. Running DeepSeek’s open-source models locally or on servers outside of China may help mitigate these risks, ensuring that sensitive information remains under the user’s control.
Is DeepSeek the Affordable AI Solution Biomedical R&DHas Been Waiting For?
I am not sure. Not yet.
AI is undoubtedly reshaping biomedical R&D, but let's be clear: its real value isn’t in using it for the sake of it. It’s about applying AI where it truly makes a difference—maximizing insights and driving real progress.
But is DeepSeek positioned to do that?
No doubt, DeepSeek showcases several innovative features that could further impact biomedical research but we see deeper and more fundamental challenges. Since the release of AlphaFold around 4 years ago, 1000s of models using expression data, pathology images etc have been published but perhaps none as high impact or successful as AlphaFold itself.
I believe as a community we need to address these fundamental issues before optimizing for costs.
Abhishek Jha | Co-Founder & CEO Elucidata
.svg)
%201.svg){kind=icon}