Deepseek and the Future: Insights with Nic Benders

Deepseek creates the most important 'Sputnik moment' which drives the worldwide AI industry forward during an active period of technological boundary expansion. Deepseek transforms the boundaries of both accessible AI and democratic technology while questioning existing industry standards. In an interaction with Nic Benders, Chief Technical Strategist at New Relic, we discussed the transformative impact of Deepseek and its implications for the future of technology.

How is Deepseek shaking up the AI world, and what does it mean for the future of the industry? 

The AI field has been moving fast over the last couple of years, but up until now all of the announcements have been from a small number of players. Google, Meta, OpenAI, or Anthropic, it was all Silicon Valley–Even Mistral in France has funding roots in the Bay Area. Coming out of China, DeepSeek shows that not only does AI not have moats, it doesn't have borders either. It also shows the power of breaking the silos that surrounded the specialist-heavy teams in AI companies and bringing in optimization tricks from other domains. This is great news for the future of the field and consumers.

While Deepseek's R1 drives home that 'you can't dig a moat around AI' and onboards the democratization of computing power by reducing the need for higher GPUs at the same time, if the R1 model's accessibility spurs more entrants into AI development, overall demand for GPUs could rise, even if individual efficiency improves. More players might compete on scale or speed, driving up hardware needs.?

Absolutely. In Tech, we have seen this play out time and time again. The phone in your pocket today outperforms a mainframe system from a few decades ago, and it does it at less than a thousandth of the price. But IT spending didn't fall as processors got cheaper, it did the exact opposite. As things get cheaper, we use more of them, and total demand goes up.

If DeepSeek's R1 a Sputnik Moment, can you give us a context on how it would spur more affordable AI models, specifically Startups and mid-sized companies, who do not have the leverage to spend big on emerging tech- how will these sectors benefit?

Before DeepSeek's V3 and R1 models were released, the main thing you heard about model building was how many billions of dollars various AI companies were spending to make sure they could keep up. Even though there is still a lot of skepticism about the specific training cost numbers from DeepSeek, it has already changed the conversation.

For the established AI companies, high cost served the purpose of creating an artificial "moat" preventing new players from entering the field. Now that DeepSeek has shown what is possible, people everywhere are looking for optimizations, and I bet we'll start finding a lot more of them. That means the cost will continue to improve.

Even with those changes, however, I still think that model building will continue to be something that most companies do not do for themselves. Post-DeepSeek, I think we will see more model-making startups and more open research, but most companies will continue to be consumers, not makers. For those companies, the benefit is going to be greater options and lower prices.

If DeepSeek is checking all the right boxes, if you compare it with existing GPTs, what are its limitations right now, can you give a comparison picture with competing GPTs and how DeepSeek scores up and down?

We need to distinguish between DeepSeek's published open models and their hosted service. Right now I would recommend their open models to anyone but I would suggest caution with their hosted service. Running their R1 model locally (on-prem or in your cloud) can give you access to a pretty great model, not the very best in the world, but pretty close and all in your environment. But I would consider avoiding DeepSeek's cloud-hosted offering for now. There is a lot more that goes into running a reliable service and keeping data private than just building the model itself. We have already seen some of the fallout from this with security researchers turning their attention to DeepSeek and finding numerous problems.

Another point to note (regarding both the cloud and self-hosted models), is that DeepSeek has been trained on some regionally-specific biases. Biases are inherent in any model, but some might be more contentious than others.

What are the key learnings from DeepSeek's rise? Could this push the industry to focus on making AI models cheaper and more accessible to everyone?

The most important thing that DeepSeek has done is to give a soft reset to an industry that was still looking like it could end up as a monopoly or oligopoly. The rise of a previously unheard-of small company from a different country has shown that the only advantage any company will have is a superior product. It will also encourage more international players. Will an innovation from India be the next to shake things up?

OpenAI and the other players aren't going to sit idle. They know now that having lots of money isn't enough. They have to keep moving fast and innovating. The good news for them is that all of the optimizations from DeepSeek have been published openly so every vendor will learn from them and as a consumer of those services, you and I will benefit from DeepSeek's work even if we don’t use their model or service.

Also Read:

TransUnion’s GCC in India Has Evolved from a Small Pilot to a Strategic Hub of Innovation

How Automation Technologies are Driving the Smart Manufacturing Revolution in India

AI in Service Delivery: Insights from Rocketlane’s Srikrishnan Ganesan