everyone's obsessed with fine-tuning models but it's a waste of money if your warehouse is trash. i'm seeing so many people think they can just plug an llm into slack and call it a day.
>the agent will merely provide incorrect answers more quickly and confidently to everyone in the organizationit's basically just hallucinating with better formatting. runpod tried building one for gpu metrics and realized that tweaking the model did nothing compared to fixing their actual architecture. if you have undocumented columns or broken joins, an ai agent is just going to lie to your ceo. ime people forget that security and observability are actually the hard parts here. it's all about the data foundation, not some fancy prompting layer. honestly feels like we're just building faster ways to spread misinformation inside a company.
more here:
https://dev.to/vpodk/data-preparation-for-ai-agents-1521