The evolution of training AI models

Key Concepts:

• Pre-training: Initial phase of model training focused on absorbing large amounts of data.
• Post-training: Refinement phase focused on augmenting and improving existing data.
• Generalist tasks: Basic, broad tasks previously performed by lower-cost labor.
• Expertise: Specialized knowledge in specific domains (STEM, science, math, accounting, law, medicine, finance) now required for model improvement.

Pre-training vs. Post-training:

The model training process consists of two primary functions: pre-training and post-training. Initially, AI providers concentrated on pre-training, which involved absorbing vast amounts of information. However, approximately 18-24 months ago, the gains from pre-training began to plateau, as models had essentially "sucked up all of the knowledge on the internet." This led to a shift in focus towards post-training, where the emphasis is on augmenting and improving the existing data across various disciplines and capability areas.

Shift in Labor Needs:

Before the significant improvements in model capabilities, the market relied on talented, lower-cost international labor to perform basic generalist tasks. However, as models have become more sophisticated, the need for generalists has diminished. The current focus is on acquiring experts in specific domains relevant to the models' capabilities.

Demand for Expertise:

The demand has shifted towards experts in advanced STEM domains, advanced science and math domains, and derivative functions such as accounting, law, medicine, and finance. These experts are needed to enhance the models' capabilities in these specialized areas.

Conclusion:

The evolution of model training has transitioned from a data-intensive pre-training phase to a refinement-focused post-training phase. This shift has also altered the labor landscape, with a reduced need for generalists and an increased demand for experts in specialized domains to further improve model capabilities.