Published Dec 22, 2022

Jerome Pesenti — Large Language Models, PyTorch, and Meta

Jerome Pesenti delves into the transformative power and challenges of large language models, highlighting AI's role in revolutionizing drug discovery, education, and AR/VR interfaces, alongside insights into PyTorch's evolution at Meta and the importance of addressing bias and scalability issues.
Episode Highlights
Gradient Dissent - A Machine Learning Podcast logo

Popular Clips

Questions from this episode

Episode Highlights

  • Deployment Challenges

    Navigating the deployment of AI models from research to production presents significant challenges. highlights the complexity of ensuring reliability and scalability, noting that AI systems often behave unpredictably and are difficult to test effectively 1. He emphasizes the ongoing nature of this challenge, stating, "I think it's still a work in progress." 1 Additionally, Jerome discusses the balance between skepticism and optimism in AI's potential, positioning himself between critics like Gary Marcus and those overly enthusiastic about AGI 2.

       

    Scaling Complexities

    Scaling AI systems, particularly in moderation and recommendation, involves intricate challenges. Jerome explains that while AI has significantly reduced hate speech on platforms like Facebook, the user experience can vary widely due to cultural and linguistic differences 3. He notes, "What matters is actually how people experience your product," highlighting the importance of user-centric design 3. Furthermore, optimizing tools like PyTorch for production requires balancing flexibility with efficiency, a task Jerome describes as "constant progress" 4.

       

    PyTorch Development

    The evolution of PyTorch at Meta illustrates the intersection of research and production needs. Jerome recounts the strategic decision to focus on PyTorch, driven by its community support and potential for dynamic graph capabilities 5. He acknowledges the dual challenge of maintaining community engagement while ensuring the framework's production readiness, stating, "You have to create something that people will continue loving and you cannot make it bloated" 4. This approach has positioned PyTorch as a favored tool, though Jerome admits that fully integrating it into production remains a work in progress 4.

Related Episodes