Published Apr 18, 2023

ML Scalability Challenges // Waleed Kadous // MLOps Podcast # 154

Waleed Kadous delves into the complexities of scaling machine learning systems, highlighting the financial, infrastructural, and strategic challenges of large language models and their scalability. He discusses the future of AI with a focus on open-source, the integration of human intelligence, and innovative technologies like Ray to navigate the evolving landscape of ML infrastructure.
Episode Highlights
MLOps.community logo

Popular Clips

Episode Highlights

  • Scaling Challenges

    Scaling large language models (LLMs) presents unique challenges, particularly in terms of cost and infrastructure. highlights the financial burden of using models like OpenAI's DaVinci, where a single query can cost $0.12 for 1,000 tokens, making it significantly more expensive than a typical Google query 1. He explains that deploying these models requires substantial computing resources, often needing multiple GPUs to handle the massive model sizes 2. This complexity is compounded by governance and customization issues, which remain significant hurdles in the industry 2.

    I've never seen anything like LLM before... but the first of those is cost. Like, I don't know if you want to serve a fine-tuned model from OpenAI.

    ---

    Despite these challenges, the potential of LLMs continues to drive innovation, with ongoing efforts to address these barriers.

       

    Economic Aspects

    The economic aspects of using LLMs are critical, with emphasizing the importance of data volume and human collaboration in enhancing model accuracy 3. He suggests that while more data generally improves model performance, integrating human intelligence through annotation and collaboration can also enhance outcomes 3. This hybrid approach, which he terms the "cyborg model," combines machine learning with human input to achieve superior results 3.

    I'm really looking towards the transition more towards what I call the cyborg model, which is a hybrid of human and machine working together.

    ---

    This collaboration is evident in applications like AI-assisted art, where human creativity is augmented by machine capabilities, offering new possibilities for innovation 4.