I think we're past that point; they're absolutely useful already for a lot of tasks. I think it's about costs, convenience, and benefits of a frontier model for what you're doing.

I do classification with SLMs and for my tasks when I have a few thousand samples the frontier models in zero-shot and few-shot modes are embarassingly bad in comparison.

until its cheaper to train and infer than 100k gpu data centers...i doubt it will ever compete.

[flagged]