The discussion critiques the industry's heavy focus on scaling raw intelligence ('IQ') through task-centric benchmarks. Zeltman acknowledges that models are becoming smarter on verifiable tasks but argues this approach neglects more fundamental capabilities and creates a 'jagged' performance profile.
A central theme is the philosophical and practical choice between designing AI to replace humans versus empowering them. Zeltman asserts that the default instinct of most labs is to remove the 'messy' human element for easier scaling, but that keeping people in the loop is a crucial, active design choice for achieving superior outcomes and innovation.
Zeltman posits that the next major frontier for AI is not just reasoning but 'EQ'—the ability to understand user goals, emotions, values, and long-term context. He frames this not as a feature for companionship apps, but as a core productivity-enabling capability that is essential for building truly helpful AI.
The current AI training paradigm, focused on discrete tasks, is identified as a key reason models lack crucial features like long-term memory and a deep understanding of user intent. Zeltman advocates for a new approach where the model's objective is to build a persistent, evolving model of the user.
The conversation touches on Zeltman's foundational research, including the Star and QuietStar algorithms. These methods, which involve models iteratively generating and learning from their own solutions, represent key steps in improving AI reasoning and demonstrate a path to scaling these capabilities.
Keep pulling the thread on Eric Zelikman.