DeepSeek | Solve for X

In the last month, another AI appeared as if it had just “fallen out of the sky,” to borrow the phrase many media pundits used. It evoked a similar sense of amazement to what surrounded OpenAI’s ChatGPT when it debuted over two years ago. This time, however, Chinese startup DeepSeek has sent shockwaves through the industry as their announcement came hot on the heels of the US administration’s announcement of the Stargate initiative, with a $500 billion AI infrastructure commitment to the US. Yet, DeepSeek and its open-source model reportedly deliver performance comparable to leading US AI systems, with the punch line that it was developed at a fraction of the cost (albeit, reportedly, still around $6 million).

Politics, and investments aside. While many commentators argue that DeepSeek’s achievement calls into question the billions of dollars currently being invested in AI, some in the technology industry see it simply as a company “solving for X.” A phrase that is fundamental to the development of Silicon Valley and the technology we all know and love. Consider Kevin Rose’s creation of the original Digg social news site, Mark Zuckerberg’s development of a platform that surpassed MySpace, or even Apple’s bold move to phase out Blackberry’s physical keyboard in favor of the sleek, touchscreen devices we now rely on.

In DeepSeek’s case, due to US sanctions and restrictions on certain high-end chips and processors, they reportedly created novel optimizations that allow AI and large language models (LLMs) to run on less expensive hardware without sacrificing performance. This approach not only achieves higher efficiency but also introduces new learning constructs for AI, significantly reducing the apparent cost of operation. A reminder that sometimes constraints can be the mother of invention, or in this case innovation, and help solve for X.”

Constraints; the Mother of Innovation

DeepSeek on launch made access to their model so widely available, the app alone in the Apple App Store shot to become the most downloaded app. And unlike the launch and groundswell around ChatGPT, their online version didn’t leave new users hanging for long (I did experience a slight outage, but nothing like day one of ChatGPT). DeepThink R1 model provides not just another chat window, but also provides a reasoning engine that offers users a behind-the-scenes look at how the AI processes and fetches information before generating a response.

Such a notable feature, and in my opinion, a long overdue change for any AI, notably at the time of writing, it appears that OpenAI has recently introduced a similar feature, aptly named “Reason.”

This is an awesome and much-welcomed feature for any model, especially for those of us who continuously test AI, probing its guardrails and evaluate how it synthesizes information. By leveraging a chain of thought and reinforcement learning workflow in their model, DeepSeek R1 is essentially “thinking out loud” as it crafts its response, allowing for users to vet and verify that the logic it is using is correct before it crafts its response. (To all my teacher colleagues, we can now make our AI show their work to us, now can we have our students do the same? Too soon? …)

Data In, Data Out?

In my experience with DeepSeek, it doesn’t generate images and has limited document and image processing (via text extraction with a maximum of 50 files and 100MB), and it still doesn’t support voice interaction (yet). However, it does have a significant advantage: it’s open source and available for download (a roughly 303GB package) along with the fact that they currently allow it to be used freely via a mobile app or their website. However, as an elder-millennial and someone who’s seen the evolution of tech firsthand, the fact that all of this is so polished and so free immediately raises some red flags. If everything is free, (and, sure, it was reportedly cheaper to develop) how does the company make money? Could it be that we, or our data, are potentially the product? As ABC reported today, “DeepSeek coding has the capability to transfer users’ data directly to the Chinese government.” It’s a disconcerting possibility that merits serious discussion, especially as our digital footprints become increasingly pulled into (voluntary or involuntary) AI systems.

All this has happened before; and shall happen again?

It’s too early to tell whether DeepSeek will be viewed as competition, a disruptive force, or even a threat to AI and data security. What it has done, undoubtedly, is introduce a potentially new way of “solving for X,” shaking up the vanilla or groupthink mentality we’ve seen lately in AI structures.

When it comes to AI, as an academic, I can’t help but feel a bit like Admiral Adama from Battlestar Galactica. Much like him, when dealing with deeply integrated systems, I prefer a more decentralized approach that prioritizes customizability, safety, and security, even if it’s going to be less convenient. I find that maintaining a healthy skepticism and a commitment to decentralized principles can help us navigate this rapidly evolving landscape without losing sight of individual privacy and autonomy (humans in the loop vs on the loop).

Similarly, in our current landscape, while fully integrated and connected systems may offer efficiency and convenience, a more compartmentalized approach can safeguard against potential breaches and data misuse. Just as the Battlestar Galactica fleet’s separation strategy was essential for survival in a hostile universe, maintaining distinct boundaries within our digital ecosystems may be less streamlined, but it could prove vital in protecting our privacy and security in the future…. “so say we all.”