Chasing Polaris PE-06 Rethink the infrastructure of the cloud in the era of AI
Yangqing Jia's point of view on AI Industry
Yangqing Jia is one of the most prominent global AI scientists. He founded and open-sourced the renowned deep learning framework Caffe during his doctoral studies, which has been adopted by companies like Microsoft, Yahoo, and NVIDIA.
A year ago, he embarked on entrepreneurship, focusing on AI infrastructure. His company Lepton carries a striking slogan: "Build AI The Simple Way."
The following is a portion of a speech regarding the future of AI he shared in a close door meeting, we got his authorization to publish in this blog. If you don’t have a time to finish reading the whole article, I would recommend you to try their new Chrome extension https://elmo.chat/ to read the summary:
Last year, along with several colleagues, I co-founded Lepton AI. "Lepton" means "lepton" in physics. We all have experience in the cloud computing industry and believe that the current development of AI presents a transformative opportunity for the "cloud."
Today, I'd like to focus on how we should rethink the infrastructure of the cloud in the era of AI.
Evaluating the Economics of Large Models:
As the scale of models grows, the core issue lies in the high computational resource costs they require. From a practical standpoint, it's crucial to consider how to efficiently utilize these models.
For instance, comparing a generic large language model with a domain-specific fine-tuned model vividly illustrates the difference. While a generic model may struggle to provide relevant responses, a fine-tuned model tailored to a specific domain can yield more effective results, as demonstrated in the example of a finance-focused dialogue system.
New Opportunities in Hardware Industry:
NVIDIA recognized the potential of high-performance computing as early as the early 2000s, leading to the development of CUDA, which has become the standard language for AI frameworks and software.
While NVIDIA continues to lead as an AI hardware provider, there are emerging opportunities for other hardware companies, especially as AI models become more standardized and adaptable across different platforms.
Wave of Generative AI: Incremental Opportunities:
The landscape of AI applications is witnessing a surge, with models rapidly increasing in number and capabilities. There are two main trends: models striving to enhance applications and applications integrating AI capabilities to strengthen their functionalities.
Productivity and entertainment are two burgeoning categories of AI applications that have overcome initial challenges and are experiencing sustained traffic.
AI in Traditional Industries: Uncharted Territory:
Despite the emergence of large language models, traditional industries present unexplored challenges within the AI sector.
The focus shifts beyond model development to encompass data collection, ensuring data consistency with application requirements, adaptation, and continuous improvement based on user feedback.
Business Models of Large Models: Dilemmas and Market Trends:
The commercialization of large models introduces dilemmas concerning revenue flow and sustainability compared to traditional software.
There's a shift from traditional revenue streams where users pay for services towards a venture capital-funded model, raising questions about long-term viability and profitability.
The rapid iteration pace of large models contrasts with the extended revenue-generating timeline of traditional software, prompting reconsideration of effective business models.
This observation reflects the evolving landscape shaped by the surge of generative AI, with Lepton actively assisting enterprises in navigating the cost-effectiveness, performance, and efficiency aspects of deploying generative AI solutions.