As soon as we hit a research plateau, a new era of optimization and distillation will begin, and the value will be captured by the application layer that has bet on the right product and UX direction
For the next generation of products, we need to assume the LLM layer will be a commodity, with all the limitations of the current underlying technology baked in (hallucinations are here to stay)
If you're not designing your product around a human-in-the-loop experience, you're essentially betting that LLMs will be reliable when in reality they're hallucination machines. That means you always need to review what they've done, no matter how simple the task (ps: here I am mostly referring to qualitative tasks, not quantitative ones)
[link] [comments]