In the podcast episode of Intel’s “Code Together” series, Hugging Face and Intel – Driving Towards Practical, Faster, Democratized and Ethical AI solutions, host Tony Mongkolsmai explores the integration of AI solutions into actual products and the challenges associated with training and inference at scale. The guests, Julien Simon (Chief Evangelist at Hugging Face) and Ke Ding (Principal Engineer at Intel), discuss various aspects of accessible production-level AI solutions.
Hugging Face's goal, as described by Julien Simon, is to democratize machine learning by simplifying the process for researchers, data scientists, and machine learning engineers. They aim to allow anyone, from junior developers to experts, to work with state-of-the-art models effortlessly. The Hugging Face Hub currently hosts almost 475,000 models.
The podcast delves into the collaboration between Hugging Face and Intel, specifically focusing on the BLOOMZ model, an open-source alternative to GPT-3. Julien and Ke share performance comparisons, highlighting the impressive inference speed achieved with Intel's Gaudi 2 hardware. The discussion emphasizes the importance of choosing the right model size for specific use cases. While larger models might seem impressive, Julien and Ke suggest that when fine-tuned on domain-specific data, small, well-tailored models can outperform larger ones in many cases. This aligns with the idea of right-sizing models for optimal performance.
Tony Mongkolsmai addresses questions regarding the scaling differences observed in BLOOMZ models, and Ke Ding explains the intricacies related to computing, memory, and communication ratios for different model sizes. The conversation further emphasizes the significance of choosing the right hardware for specific tasks to achieve cost efficiency and performance.
The podcast then shifts to the broader perspectives of the AI landscape, touching upon ethical considerations and transparency. The guests express their views on the importance of an open ecosystem and transparency in the research and development of AI. They discuss the need for explainable AI (XAI) and tools to address ethical concerns, especially in cases where models generate incorrect or misleading information.
The conversation concludes with a discussion on the challenges of closed-source models, citing the example of GPT-4's closed nature. Julien expresses concerns about the lack of transparency and the potential risks associated with models that aren't open to scrutiny. He advocates for model diversity and an open discussion to address challenges collectively.
In the final segment, the guests share their expectations for the future of AI. Ke Ding highlights Intel's commitment to hardware innovation and a balanced approach to meet the demands of diverse AI workloads. Julien Simon looks forward to developing industry-level models catering to specific domains and deploying accurate models on edge devices for disconnected use.
Overall, the episode provides insights into the collaboration between Hugging Face and Intel, the challenges in AI model development and deployment, and the ethical considerations surrounding AI technologies, namely the need for transparency, ethical practices, and an open ecosystem to drive innovation in the AI landscape.
We encourage you to check out Intel’s other AI Tools and Framework optimizations and learn about the unified, open, standards-based oneAPI programming model that forms the foundation of Intel’s AI Software Portfolio.
The Guests:
Julien Simon – Chief Evangelist @ Hugging Face
Ke Ding – Principal Engineer @ Intel
References
ChatGPT & Wolf, A. (2023, January 16). Summarize the main points from the following transcript.
https://chat.openai.com
Learn more about HuggingFace and Intel in AI solutions
AI Software Marketing Engineer creating insightful content surrounding the cutting edge AI and ML technologies and software tools coming out of Intel