I think this applies a bit less to the AI sphere, which has the purported goal of making things easier and more automated over time. 90% of the time if you have an AI question you can just... ask the LLM itself.

Remember all the hoopla over how people needed be a "prompt engineer" a couple years back? A lot of that alchemy is basically totally obsolete.

Think about the hoops you had to jump through with early GenAI diffusion models: tons of positive prompt suffixes (“4K, OCTANE RENDER, HYPERREALISTIC TURBO HD FINAL CHALLENGERS SPECIAL EDITION”) bordering on magical incantations, samplers (Euler vs. DPM), latent upscalers, CFG scales, denoising strengths for img2img, masking workflows, etc.

And now? The vast majority of people can mostly just describe desired image in natural language, and any decent SOTA model can handle the vast majority of use cases (gpt-image-1.5, Seedream 4, Nano-banana).

Even when you’re running things locally, it’s still significantly easier than it used to be a few years ago, with options like Flux and Qwen which can handle natural language along with a nice intuitive frontend such as InvokeAI instead of the heavily node-based ComfyUI. (which I still love but understand it's not for everybody).