#046 The Great ML Stagnation (Mark Saroufim and Dr. Mathew Salvaris)

Scaling AI

Mark Saroufim discusses the challenges and opportunities in scaling AI technologies, emphasizing the role of industry over academia in this domain. He notes that while scaling AI models is not trivial, it is primarily an industrial challenge, requiring robust infrastructure and reusable code, which companies like OpenAI excel at 1. Saroufim highlights the complexity introduced by state-of-the-art (SOTA) chasing, where incremental improvements often lead to new benchmarks that are difficult to verify 2.

It's not like you just say python, make, make, model big. And, you know, that's just how it works.

--- Mark Saroufim

This pursuit of SOTA can sometimes overshadow genuine scientific inquiry, turning the focus towards hype rather than substantive advancements 3.

   

Academia vs. Industry

The tension between academia and industry in AI development is palpable, with Saroufim critiquing the current academic incentive structures. He argues that academia often resembles a media business, focusing on flashy results rather than meaningful progress 4. Saroufim suggests a "barbell strategy" for academics, balancing market-driven projects with innovative research to ensure financial stability while pursuing groundbreaking ideas 5.

Academia is a media business. So really, academia is not that different from a clickbait business.

--- Mark Saroufim

This approach allows academics to navigate the resource-intensive landscape of AI research without compromising on creativity or impact.