Finding information is more closely related to retrieval augmentation, while reasoning is a separate challenge.
You're correct that changing prompts can solve some hard reasoning questions, and techniques like Reflection and Chain-of-Thought (CoT) prompting are already being used.
However, the real goal is to make models reason intuitively, like humans. For instance, a human would naturally know to count individual letters in your example.
Researchers are working hard on this through various approaches, including what's called "test-time optimization" of LLMs. Reasoning is currently one of the biggest focus areas in LLM research, and the community is making progress. For example, OpenAI's latest model shows a 30-40% improvement in PhD-level reasoning compared to its previous version. While current methods are helpful, the ultimate aim is to develop models with more fundamental reasoning abilities, which remains a complex but actively pursued challenge in AI research.
Anyway, I have been working with foundation models since 2017 (we used to call it the BERT era

) and my PhD was in domain adaptation of them. What I can see is that it's just a matter of time before we have to rethink coding. Sometimes I implement algorithms from super novel research papers which are just a day old by uploading the paper to ChatGPT and having a conversation. Even research is something that needs to be restructured, especially with papers like this from Sakana.ai
https://sakana.ai/ai-scientist/