LLM Reasoning Limits Are Getting More Precisely Mapped

The 'LLMs Can't Jump' paper thread and the 'Eight Myths on Software Engineering and GenAI' thread are circling the same question from different angles: what exactly can LLMs not do, and how do we build systems that account for those limits instead of pretending they don't exist?

The 'LLMs Can't Jump' argument is that LLMs cannot make genuine leaps of intuition in under-explored problem spaces. They are excellent at interpolating within their training distribution but fail at the kind of sideways thinking that produces novel solutions. The software engineering myths thread pushes back on the 'AI will replace developers' framing by noting that writing code is only 14% of a developer's time, and that AI accelerating that part doesn't change the bottleneck.

The key bit: the community is moving past the binary 'AI is amazing or AI is overhyped' debate into something more granular and useful. People are trying to draw precise maps of where LLMs break, which is exactly the kind of information that lets you build reliable systems instead of demo-ware.