The aphorism is “strong opinions, weakly held.” Now, I’ve always been skeptical of what, on its face, is a contradiction. At best, it’s useful as rhetorical insurance, allowing its claimant to avoid the consequences of their strong opinions. At worst, the phrase preempts jerk behavior.
In our world of AI-driven conversations and creative output fed through LLMs, however, it feels more important to start with a strong opinion and to hold onto that perspective, even as ChatGPT tweaks its responses with a combination of support and sycophancy. It’s when the user lets go of close control of their AI assistant that their voice gets diluted, washed over amidst a tsunami of AI slop and say-nothing pleasantries.
I’ve been continuously experimenting with how to better incorporate AI into my writing processes, both for this blog as well as professionally for my job. Its strengths lie in summarization, unstructured searches within publicly available data1, and suggesting a turn of phrase or three to colorfully describe a situation. ChatGPT isn’t so much writing for me as writing at me, activating the editor part of my mental processes.
These use cases line up with other domains that have found success incorporating AI. With the proliferation of AI coding assistants, successful workflows require paying close attention to the agent’s outputs and iterating on its code to edge closer to its user’s intentions. In the early days of LLM-powered image generation, even with too-many fingers and horrible visual glitches, visual artists were able to make use of Midjourney and DALL·E to map out storyboards and spruce up rough cuts to help inform the final animated products. The critical step of human evaluation is where our opinions are expressed, with impunity.
In the case of writing, it still acts as a forcing function for developing thought. If you use it in similar fashion, then outsourcing the writing itself necessarily delegates the thinking process. That’s the bulk of the criticism for overusing AI to write essays, emails, and other interpersonal correspondence—not only is the technique impersonal, but it’s usually done devoid of much attention or intention.
Which means that effective use of LLMs requires clear and decisive thinking, both in the initial prompt and query, and also in subsequent refinements. There are times when the tool helps in blank-slate scenarios, but making progress always requires clear-eyed evaluation of its output and setting a strong direction. And ignoring, or at least recognizing, all the hallucinatory sidetracking fundamental to the technology—holding onto that vision, strongly.
The corporate, private data use case requires a level of structured data that most companies don’t have by default.↩
