Steve Hsu, cofounder of the AI startup SuperFocus, has shared his insights on DeepSeek, claiming it to be an impressive company worth the attention of anyone involved in the tech industry. His team is currently evaluating the integration of DeepSeek models with the SuperFocus system and is optimistic about making the transition due to the benefits it offers. Hsu emphasizes that DeepSeek’s models are not only more affordable but also drive faster outcomes while ensuring greater privacy.
This article reflects insights drawn from a conversation with Hsu, who also serves as a professor of theoretical physics and computational mathematics, science, and engineering at Michigan State University. For clarity and brevity, the following has been modified.
At SuperFocus, we are committed to developing superhuman AI solutions tailored for businesses across various sectors. Our applications range from customer service and ordering to delivery, document analysis, and travel scheduling. A key component in our AI systems is leveraging advanced models like GPT-4o.
In discussions with fellow AI innovators and engineers, it has become evident that DeepSeek models have caught the attention of many. Their reputation stands strong as some of the top—all while being notably cost-effective—open-source options available. Given the rapid advancements in the space, it would not be surprising to witness a significant shift towards open-source models in the near future.
Currently, we are conducting extensive tests to gauge the performance of DeepSeek-V3 models within our system. However, early indications suggest suitability in meeting our operational needs, enabling us to pivot if our clients would truly benefit from the switch.
In addition to our client projects, SuperFocus is actively engaged in research and development, which coincidentally results in significant monthly expenses with OpenAI. To optimize costs, we are preparing to transition many of our processes to utilize open-source models like DeepSeek.
Here are three compelling reasons driving our decision to switch to DeepSeek:
1. **Affordability**
DeepSeek models present a remarkable cost advantage, being approximately 30 times more economical to operate compared to equivalent models from OpenAI. This significant reduction in expenses is especially beneficial in sectors such as customer service.
When assessing the overall cost of employing a customer service representative in the U.S., it may average around $25 per hour. While outsourcing to the Philippines reduces this to about $5 to $10 hourly, our previous applications using OpenAI provided a rate roughly ten times lower than hiring a human in the Philippines. Transitioning to DeepSeek-V3 could further reduce our costs by an additional 30 times.
2. **Efficiency**
Our findings indicate that the DeepSeek open-source model also achieves superior speed. This aspect is crucial, as our projects often revolve around voice interactions where latency can detract from the conversational experience. Long delays between a user’s speech and the model’s response can disrupt the natural flow of dialogue.
When integrating new technologies, particularly in software development, improvements in processing speed can be dramatic—often achieving enhancements that are 10 to 100 times quicker through superior optimizations and algorithms. The DeepSeek team has successfully implemented numerous innovative changes to their model architecture, leading to reduced memory usage and lower computational requirements, thus enhancing token generation efficiency.
3. **Data Privacy**
Many clients express concerns about their sensitive data being processed through proprietary models. For instance, private equity fund clients may need to analyze confidential documents that should remain exclusive to their firm.
DeepSeek’s open-source framework alleviates these worries, as users can operate the model on their own hardware or through securely rented servers from providers like AWS. This setup guarantees there’s no communication back to DeepSeek, allowing the AI software platform to operate solely within the client’s established cloud instance.
In contrast, some closed-source models, particularly those developed by certain Chinese companies, necessitate API communication and operate on hardware they control, which can raise privacy red flags. While some clients may have reservations about using models from China, our focus remains on creating narrow AI solutions tailored for specific applications, making political affiliations less relevant in our decision-making process.
In my view, DeepSeek stands out as an incredibly impressive and transparent organization.
My observations of several Chinese companies developing LLMs—including our team’s thorough testing of various open-source models—reinforce my belief in DeepSeek’s commendable research outputs. The clarity in their documentation and commitment to encouraging reimplementation and validation are particularly noteworthy.
The founder’s vision prioritizes openness as a means to accelerate progress in the field, a sentiment that resonates with me. DeepSeek’s level of transparency eclipses that of many top labs in the U.S., including OpenAI, which is somewhat ironic given the competitive landscape.
**The trajectory of AI development**
The emerging awareness of Chinese technological advancements is rattling perceptions in America, where many remain unconvinced of China’s formidable prowess in AI. Prominent figures in the tech sphere propagate stereotypes that American ingenuity still reigns supreme, alleging that companies like DeepSeek have engaged in unethical practices such as intellectual theft from OpenAI. I respectfully disagree with these claims.
The AI realm is poised for a significant showdown, marked by economic implications and a struggle for technological prestige. The competition between the U.S. and China promises to be tight, with expectations that top models will soon emerge from both nations.
Ultimately, this rivalry will significantly benefit consumers as cost efficiencies will drive prices down. Enhanced intelligence and insights will become more accessible in everyday applications, further propelled by the ongoing competition and collaborative brainpower in the field. I believe this culture of innovation will expedite the journey toward AGI.
The research landscape will also see advantages; for example, academics at institutions like CalTech will have the opportunity to access reasoning models such as R1 and customize them to address specific scientific challenges—a feat not previously possible with OpenAI models. Thus, the proliferation of robust open-source models is set to ignite a wave of innovation and discovery.
Priyam Gera