Generative AI platforms have three compounding problems: 1) It generates more questionable data 2) You are uncertain about the output 3) You need more technical FTEs to support. They aren't cheap. If you want to bet on something that's sustainable, this isn't it. It's as if you'd bet on "fuzzy logic" as your main data architecture. As much as generative AI appears like a powerful tool with FOMO use cases, it turns into a rather expensive power tool that you may use for drilling a wall for a frame once a year. Nail and a hammer (existing tool) generally can do that, if that's all you need.
Something worth reading: “Marcus warned that ‘the whole generative AI field, at least at current valuations, could come to a fairly swift end’, used only to produce code and marketing copy.”
4) it’s difficult and expensive to implement, and the ROI is ephemeral
Physician, Data Scientist, Clinical Dot Product Manager
1yIdk man, LLMs, and not just decoder-only models fine-tuned on instruction following/RL'd for chat alignment, but certainly excel in many areas that few other data tools do, and no one should be using them as their main data architecture. 1. This is a rather nonspecific claim and many sources of data, or outputs from systems, that are used all over the place, especially clinically are sus. This is not a unique LLM problem. 2. This claim is not super supported unless one lacks a retrieval systems (this can be somewhat challenging, but done right reduces hallucinations, often to 0). Hallucinations is just a fancy word for poor confusion matrix performance per token generated. All ML models can suffer poor performance. 3. Any data project is going to need technical FTEs. I'd argue the current ecosystem for LLMs + RAG is probably less of a barrier to entry than many other more traditional data / ML ecosystems. Finally, isn't the dude in this article the same guy who called for a moratorium past GPT4 because they were too powerful? Gotta pick a lane here.