Jay

How do benchmark gains translate to messy, real codebases with legacy patterns, unclear requirements, or incomplete context?

What happens when design intent conflicts with implementation constraints, does it keep iterating or surface the tradeoff?

As you scale variations, how do you prevent outputs from starting to feel repetitive or falling back into that same AI pattern over time?

When users are notified after shipping, how do you avoid sending irrelevant updates to people whose requests only partially match?

Visual learning is a grand pillar when it comes to teaching new concepts to students. And for those tired few who cant find the time to make a new graphic for every other class this should be a godsend.

Would like to see benchmarks across different model providers and prompt types. If the compression holds under real production loads, this could become default infra in most LLM stacks.