—
**Embracing Open Innovation: The Shift to DeepSeek Models in AI Development**
Steve Hsu, the cofounder of the AI startup SuperFocus, recognizes the remarkable capabilities of DeepSeek, an emerging player in the AI landscape. Currently, his team is in the process of testing the performance of DeepSeek models integrated with the SuperFocus system, and they are optimistic about a future transition. Hsu emphasizes that DeepSeek models stand out for being more affordable, efficient, and less intrusive to user privacy.
This article captures insights from Steve Hsu’s conversation, who is not only a pioneering entrepreneur but also a professor specializing in theoretical physics and computational mathematics at Michigan State University. We’ve edited this discussion for brevity and clarity to highlight key points.
At SuperFocus, our mission is to harness superhuman AI tailored for businesses. We specialize in various applications including customer service, order processing, delivery logistics, document analysis, and travel planning. In crafting our AI solutions, we frequently utilize advanced models like GPT-4o.
Based on conversations with fellow AI founders and engineers, there’s a growing trend towards exploring DeepSeek models. Considered among the premier—if not the top—open-source options available, these models are significantly less expensive than their closed-source counterparts. It’s difficult to imagine that we won’t see a major shift towards open source within the coming year.
Currently, we’re dedicating time to rigorously test the DeepSeek-V3 models alongside our existing system. What we’ve seen so far leads us to believe these models will adequately meet our needs, and should our customers benefit, switching over is certainly on the table.
SuperFocus is not just focused on client work; we also invest heavily in research and development. One of our ongoing expenses is a considerable monthly bill for OpenAI services. To optimize our budget, we are planning to initiate a substantial transition towards employing open-source models like DeepSeek.
**Three Compelling Reasons for Transitioning to DeepSeek**
1. **Cost Efficiency**
DeepSeek models are approximately 30 times less expensive to operate than comparable OpenAI models, which translates into substantial savings for industries such as customer service.
To illustrate, the all-inclusive cost of a customer service representative in the U.S. hovers around $25 per hour. In countries like the Philippines, this can drop to about $5 to $10. By utilizing OpenAI’s services, we managed to offer solutions at roughly one-tenth of the cost of hiring a human staffer in the Philippines.
However, with the implementation of DeepSeek-V3, our operational costs could potentially decrease even further by a factor of 30.
2. **Enhanced Speed**
Our findings also reveal that the open-source model delivers quicker response times. This is particularly vital in our work, which often involves voice interactions. Slow responses can disrupt the flow of conversation, making interactions feel unnatural.
When developing new technologies, especially in software, enhancements often lead to performance increases that can be tenfold, or even a hundredfold, due to improved optimizations and algorithmic advancements.
The DeepSeek team has integrated numerous innovative optimizations into their model architecture. As a result, they require significantly less computational resource and memory to generate tokens or train the models, thus yielding faster token generation upon receiving prompts.
3. **Robust Privacy Measures**
Data security is a vital concern for many clients, particularly when sensitive information is involved, such as that of a private equity fund. Such clients require assurance that their documents remain confidential and unexposed to external entities.
Given that DeepSeek operates as an open-source platform, its models can be run locally on a client’s hardware or on infrastructure rented through cloud providers like AWS. This design eliminates any back-channel communications with DeepSeek, allowing us to develop AI solutions entirely within the cloud environment where the client’s data resides, avoiding data transfer via APIs associated with proprietary models like OpenAI.
This privacy guarantee stands in stark contrast to certain closed-source LLMs from companies that may not prioritize user confidentiality. The dual concerns about such closed systems resemble a sentiment among some clients, who may opt for alternatives—even if they come at a higher cost.
As an organization, we’re committed to creating specialized AIs that fulfill specific client needs and, thankfully, political concerns typically don’t play into our decision-making process.
**Acknowledging the Innovators**
DeepSeek impresses us with its commitment to open collaboration in the research community. Following the innovation trajectory of several Chinese firms in the LLM space, we have explored various open-source models while developing our capabilities.
DeepSeek stands out for its exceptional research output. The founder’s idealistic vision advocates for openness in advancing the field, a philosophy we believe expedites progress. Their transparency surpasses that of nearly all other leading labs in the U.S., including OpenAI—an ironic twist considering the competitive landscape.
**The Competitive Horizon**
The current landscape makes it evident that many Americans are just beginning to grasp the advancements in Chinese technology that had previously flown under their radar. Amidst rising national pride, some may dismiss the advancements as simply accusations of misappropriation—claims suggesting that competition is merely a case of theft, rather than genuine innovation.
The AI arena serves as a significant battleground not just for economic gains but for the prestige of technological advancements. In this contest, it appears we must prepare for intense competition between the U.S. and China, often bringing forth models that originate from rapidly advancing Chinese entities.
The resulting rivalry is bound to benefit consumers as the market dynamics push prices lower, making high-quality AI more accessible. More brains tackling these challenges will surely accelerate our journey towards AGI.
This evolution will also empower researchers. For instance, professors at institutions like CalTech will gain the ability to download specialized reasoning models such as R1, revolutionizing how they tackle complex physics or mathematical problems—an avenue not previously available with proprietary models. The expansion of high-performing open-source models is set to ignite a wave of innovation across numerous sectors.
—
This rendition maintains the core content while adopting a tone that resonates with the intended unisex audience aged 18-34, offering a detailed yet approachable perspective on the topic.
Priyam Gera