Try automating yourself!
Try it, or be blind to change
My law teacher once started a class by describing the one time (1) she tried getting an LLM to write an essay based on the prompt she would give students.
“It's shit,” she said, “didn't even divide the essay in two parts and two subparts with a problématique, as is the [this school] way of doing things, [series of other nitpicks she had]".
Annoyed, I asked her if she'd given the AI the essay structure guide she hands her students. She said no. I asked her if she'd fed it her specific lesson plans and content to it. She said no.
She then paused for a few seconds and went “so you're saying it's… my job to make sure the AI has good outputs?”
Yes, yes it is. First in the sense that it's your job to do what it takes to extract the output you want from the world.1
But also, we are not yet at the point where one can vaguely gesture at what one wants, and the AI creates something satisfying.
We're getting to that point—e.g. memory is a tool that approaches this. Had ChatGPT known from a separate conversation that she was [her name], it could've looked up her methodology and lessons online, and used that.
As with all things, this still boils down to “skill issue": with memory turned on, she'd still have to've chatted enough with ChatGPT before this to get that information across, and also shared enough data about herself online for the AI to look up.
(Indeed, you should write more online.)
How long will we remain consigned to a primitive world where we must expend a modicum of effort to get what we want out of the world? How fast are we moving toward this fabled world where you can vaguely gesture at something you want done and it gets done? The upside of getting a lot of good results done in the before times is potentially high: how good can you make your results, in 2025?
These all feel like important questions. The best way to answer them is to actually try automating yourself.
Find a thing which annoys you to automate away; find something you think is your comparative advantage, and ruthlessly explain to the AI how you do it, so it can emulate you; get past the initial “yuck” response at automating something precious to you, because this is true, and what is true is already so; figure out what you might do if you had an intern, then treat the AI exactly as you would an intern; don't trust benchmarks or stated use cases to tell you what an AI can and can't do, because no-one knows, and you must interact with it directly to understand this; spend an actual afternoon doing this, by the clock; whenever you come up with a meta-level gripe about the LLM, insert it directly back into the object level; add more bits; let go of the stress of needing to be the primary mind involved in getting something done, because keeping it will make you unhappy very soon; read some Ethan Mollick posts about wizards; look at the cool environment you could build yourself if you truly try your hand at this.
