The discussion contrasts traditional, low-skill crowdsourced data labeling with the new demand for high-skill, expert human evaluators. The speaker argues that the former (SFT/RLHF) is overhyped and will be replaced by the latter, which is essential for developing more capable and nuanced AI models.
LLMs are described as reaching near-superhuman levels at text-based candidate assessment, automating resume screening, interviews, and performance prediction. The future hiring process will be bifurcated: AI will handle objective assessment, while humans will focus on the 'selling' and relationship-building aspects of recruiting.
A core prediction is that much of human knowledge work will transition from performing tasks to evaluating AI systems that perform those tasks. Humans will be responsible for creating the rubrics, tests, and frameworks to ensure AI performance, quality, and safety.
The speaker expresses high conviction that an AI software engineer capable of writing pull requests with a higher success rate than a human is imminent, arriving within the next 1-2 years. This is viewed not as a threat to jobs in the short term, but as a massive productivity tool that will change the nature of the engineering role.
Keep pulling the thread on Brandon Foodi.