Part 2/3:

The speaker delves into the intricacies of versioning and naming, noting that a year ago, the focus was primarily on pre-training models of different sizes and creating a family of naming schemes. However, as some models take longer to train than others, and as significant improvements are made in pre-training, the existing naming conventions start to break down.

The speaker acknowledges that no company has truly figured out the perfect naming system for AI models, as they often have different trade-offs, capabilities, and characteristics that are not always reflected in the benchmarks. Models can exhibit a range of personality traits, from polite to brusk, reactive to questioning, and warm to cold. Anthropic has even dedicated a team, led by Amanda, to focus on the "Claude character" and these nuanced aspects of the model.

[...]