this is the way. finetuning small LMs to VLM adapters is the future, not these massive models that take forever to train. the article is spot on - we need smarter, more modular approaches to get real progress. https://www.reddit.com/user/AvvYaa