Wow, the part about how scaling model size alone isn't enough for complex reasoning really hit home. I've been playing with LLMs a bit, and I always thought further scaling or maybe different arhitectures would eventually get us there more naturally. Is the 'prompting only' approach truly general enough for all tasks, or are there inherent limits that fine-tuning still addresses better in specific domains?
Wow, the part about how scaling model size alone isn't enough for complex reasoning really hit home. I've been playing with LLMs a bit, and I always thought further scaling or maybe different arhitectures would eventually get us there more naturally. Is the 'prompting only' approach truly general enough for all tasks, or are there inherent limits that fine-tuning still addresses better in specific domains?