Five key takeaways from the recent Deepseek event
The DeepSeek event last week revealed groundbreaking insights into the latest development in artificial intelligence, highlighting how rapidly the technology is advancing. Here are our five key takeaways:
1. Global collaboration in AI research benefits all of humanity – Many of the techniques used in Deepseek’s model—such as Mixture of Experts, Multi-head latent attention, Multi-stage training, and Reinforcement Learning—are based on previously published open research. Deepseek’s key innovation lies in how it integrates these existing methods to significantly reduce training and inference costs. This breakthrough will likely be widely adopted by the AI research community and other AI model companies over time. The exchange of ideas through research publications, such as Deepseek’s latest report, accelerates progress for all. As Meta’s Yann LeCun puts it: “Everyone’s learning from everyone else. No country ‘loses’ to another when someone achieves a breakthrough in AI open research.” Ultimately, what matters most is how these scientific innovations are transformed into real-world products and services that benefit global consumers and enterprises.
2. Declining AI costs will drive widespread adoption – In the technology industry, cost reductions often fuel greater usage—a phenomenon known as Jevons’ Paradox. According to A16z, recent AI inferencing costs have been dropping roughly tenfold each year, driven by six key factors (GPU cost/performance improvement, model quantization, software algorithmic innovation, smaller models, better post training tuning, open source offerings). We expect this trajectory to persist for years. Given AI’s price elasticity, the amount of AI workload demanded is likely to grow even faster than the price decline, ultimately expanding the overall AI market. Lower costs will accelerate adoption and unlock groundbreaking advancements in healthcare, education, scientific discovery, and, eventually, AGI and superintelligence.
3. Increasing Workload at the Inferencing Stage – Even before the publication of Deepseek’s latest model, the AI industry was already witnessing a shift in how computational resources are allocated to optimize cost and performance. Traditionally, much of the computational burden in AI development was concentrated in the pre-training phase, where models are initially trained on vast datasets. However, an emerging trend is redistributing this workload, shifting more computation to the post-training phase—particularly to real-time inferencing at test time. If this trend continues, future AI models will increasingly rely on inferencing-stage computation, performed either locally on edge devices—enhancing speed, privacy, and efficiency—or remotely in the cloud, leveraging scalable processing power.
4. Proprietary vs. Open Source Technologies: Two Viable Business Models – In the technology sector, we have seen commercial success in both proprietary (Apple Mac and iPhone) and open source (PC hardware and Android OS) business models. Unlike proprietary systems that rely on a single company’s R&D, open-source technology benefits from contributions by a global developer community. With a broader user base, supply chain providers—including silicon manufacturers, APIs, and developer platforms—often standardize and optimize for open-source frameworks, driving higher adoption, economies of scale, and lower costs. As a result, open-source technology typically serves the mass global market with lower costs, whereas proprietary platforms cater to premium consumers, enterprises, and government sectors, offering enhanced support and exclusivity. For instance, we see OpenAI use proprietary government data to train specialized models for national security, defense, and public administration. Meanwhile, we also see other AI companies like Meta provide free, ad-supported models for general purpose use, expanding accessibility while maintaining a sustainable business model.
5. The Future of AI Agents – DeepSeek has shown that narrowing an AI model’s focus to a specific use case—such as logic/reasoning problems with clear solutions and English/Chinese writing —can significantly improve computational efficiency. This suggests that the future of AI will not be dominated by a single, all-purpose model, but rather by a vast ecosystem of specialized AI agents, each tailored to a distinct function. We envision a world where billions of AI agents, both large and small, operate similarly to human workers today, each excelling in a particular role. There will be AI doctors, personal assistants, researchers, security guards, office workers, and factory operators. Personal AI agents will be customized to align with their owners’ interests, personalities, preferences, languages, and cultural contexts. Meanwhile, enterprise AI agents will be fine-tuned for industry-specific applications, leveraging proprietary data, operational workflows, and customer needs. Running these AI agents will demand substantial computational resources for both training and inference, as they will integrate a diverse set of capabilities, including short- and long-term memory, speech and visual processing, motion and interactions in the real-world, and advanced reasoning. This shift toward specialized AI systems will redefine productivity, automation, and human-AI collaboration across industries.
The recent Deepseek event highlights the rapid evolution of AI technology, its increasing accessibility, and the potential for transformative impacts across various sectors.
