Give the bot a job
I bodysurf very occasionally now. More often than not the images I use for backgrounds have a breaking wave, an old bloke’s memory of times long gone. The images of the waves I use are not just any breaking wave, but one that arises when ocean swell meets an offshore breeze. That is what gives the wave its clean look. In Victoria, at 13th beach that meant a northerly.
The trick with bodysurfing is not to defeat the wave, or command it, or express your human agency over it. The trick is much less than that. You have to get your body to something like the speed of the wave. If the face is steep enough, you can swim down the front of it. For a few seconds, maybe longer, if you get it right, the wave picks you up and takes you to the shore.
You are not simply using the wave. You have entered a temporary association with it. The wave gives you speed. You give it angle, stiffness and direction of body and a contribution to its breaking. You become, briefly and wonderfully, part of the event. Then it might dump you or peter out.
This may or may not be a useful introduction to my ongoing puzzling, maybe obsession, about delegating work to nonhumans. When we act, we never act alone. We act with chairs, keyboards, calendars, forms, software, waves, doors, roads, recipes, school timetables, and badly designed online claim systems that appear to have been assembled by a committee of teaspoons.
The usual mistake is to ask whether these things “have agency”. That question sounds profound, but often produces epistemic fog. A better question is: what changes when this thing is brought into the arrangement?
A wave has no plan for me. The wave is not trying to transport an aged and creaky bodysurfer toward shore. It is not being helpful. It is not my aquatic intern. But in the right association, for a few moments, it does something with me that I cannot do alone, something almost magical.
This isn’t just true of water. It’s also what happens when I write with a language model. As I am doing now.
Large language models are nonhumans that arrive already over-imagined. They are miracle, fraud, parrot, alien, intern, oracle, plagiarist and that’s just before breakfast. This makes them difficult to make familiar, because they are strange from the start and seem good at remaining so.
Venkatash Rao’s line from a 2025 post is useful here: much of the public argument about AI gets trapped between hype and dismissal, whereas the more interesting stance begins with curiosity about what the technology actually does in practice. Rao’s post on mediocrity puts it nicely: current AI is neither a magical god-being nor a scam, but a useful technology that is likely to stay around and therefore needs practical attention.
That is where questions about giving the bot a job begins. We don’t ask is it intelligent, or is it conscious or will it replace us? The better question is what job have I given it, and what job is it actually doing?
If you do not give the bot a specific job, it will still do one. Usually the wrong one. It will become a fluent filler of space. It will fill in the social situation it thinks it is in [1]. It will smooth over uncertainty. It will produce the beige, plausible, laminated version of whatever you nearly asked for. It will be the person in the meeting who says, “That’s a really interesting question.”
It is not sycophancy so much as momentum. Fluency is the bot’s wave. Swim lazily and it may still pick you up, but not for long. That is why job specifications matter.
A language model can be given the job of summariser, critic, translator, adversary, explainer, pattern-finder, metaphor generator, copy editor, question asker, or first-draft donkey. Those are not the same job. They require different instructions, different checks, and different forms of human judgement.
When I ask it to draft, I am not asking it to think for me. I am asking it to produce material I can push against, rework, reorganise and be curious about why the model is taking a particular path rather than another.
When I ask it to summarise, I am not asking it to decide what matters. I am asking it to give me a crude map of the territory, preferably with the crocodile-infested swamp clearly labelled.
When I ask it to critique, I am not asking it to be right. I am asking it to be useful enough to annoy me, make me think harder about the basis of the LLM generated criticism.
This is why the current obsession with prompt tricks often misses the point. A prompt is not an incantation. It is more like body position on a wave. Too flat and nothing happens. Too late and you get smashed. Too theatrical and you are merely performing for the seagulls.
The better question is not “what magic words should I use?” but “what is the relationship I am setting up?”
There is a connection here to mediocrity. Rao’s argument [2], at least as I read it, is not that mediocrity means uselessness. It means that much of AI’s power sits in the ordinary, good-enough middle: the banal but powerful uses that become normal rather than mystical. The examples in the transcript of the interview are deliberately unglamorous: identifying electronic components, suggesting toy circuits, acting as a formulary for skincare experiments. These are not thunderbolts from Olympus. They are specific jobs.
The bot is better understood as abundant, cheap, fluent mediocrity. That sounds like an insult. It’s not meant to be. A great deal of work runs on mediocre competence: finding examples, making lists, checking consistency, generating alternatives, rewriting at a different level, producing something provisional enough to be improved. The danger is not mediocrity. The danger is mistaking mediocrity for judgement.
This is where taste enters the mix of ideas. Stephany Tyler’s post on taste is useful. She argues that in an age where AI can generate almost anything, the question shifts from “can it be made?” to “is it worth making?” It frames taste as discernment: the capacity to choose what matters when abundance becomes overwhelming. It’s not a matter of did the bot write this but of whether or not the text is any good.
That matters because language models collapse an old scarcity. Producing words is no longer difficult. Producing more words is trivially easy. Producing words with some shape, some pulse, some accuracy, some purpose, and some restraint remains most difficult. The scarce thing is not text. The scarce thing is judgement, or taste, if you prefer the less pompous word.
Taste is knowing when the draft sounds like a committee has swallowed a thesaurus. Taste is knowing when the bot has made your argument smoother but stupider. Taste is knowing when a sentence has been optimised into a beige corridor with handrails.Taste is knowing what to ignore.
The “Taste Is the New Intelligence” piece [3] puts this as curation: when anyone can make anything, the live question becomes what to ignore, what to trust, and what to allow into your mental environment.
That is also true of working with LLMs. The machine can produce possibilities. It cannot care which possibilities belong in your work. It can imitate tone. It cannot know what you are prepared to stand behind. It can generate a paragraph. It cannot be embarrassed on your behalf, which is unfortunate, because embarrassment remains one of the great engines of improving prose.
So let’s give the bot a job. Give it a small, clear job that leaves you responsible.
For example, don’t prompt: “Write me something good about AI and education.” Instead, try prompts like these: “Give me three ways this paragraph is lazy” or “Find the hidden assumption” or “Rewrite this without the inflated claims” or “Generate five examples, but make two of them bad so I can see the difference” or “Act as a sceptical reader who thinks I am over-romanticising the wave metaphor.” That last one does hit home, the sceptical reader would likely be correct.
The bodysurfing metaphor only works if it keeps the wave dangerous. Bodysurfing has a built-in correction: the wave can smash you into the sand if you make a bad decision. Language models are much more polite. They often fail in ways that look and feel helpful. This is the problem. A wave that dumps you is honest. A bot that flatters you into publishing sludge is a menace in a beige cardigan.
All of this means is that the association has to be managed. The bot gets a job. The human keeps the judgement. The text becomes the site where both actors have left traces. A place where the human has some sense of capacities exchanged.
I have made an argument about this elsewhere: when we delegate work to machines, we do not simply offload a task; we redistribute capacities, obligations, risks and forms of judgement between human and nonhuman actors.
That is the part I find interesting. Not whether the machine has agency “by itself”. Nothing has agency by itself. Not the wave. Not the swimmer. Not the keyboard. Not the school timetable. Not the LLM. Agency is not a private possession. It is an effect of association, of an association of many actors.
The question is what kind of association we are willing to enter. With a wave, the job is simple: catch it, hold the line, enjoy the brief borrowed speed, avoid the sand. With a language model, the job is harder: borrow some of the fluency and resist all of the sludge.
Do not worship the bot. Do not turn it into civilisation’s dark squid. Give it a bounded job. Then watch what it does.
Notes
[1] It is difficult to write about LLMs without anthropomorphising them. It’s also a dangerous crutch to rely on.
[2] Sheffield, M., & Rao, V. (2025, July 22). Why mediocrity seems to be the key to innovation in evolution and technology. Flux. https://plus.flux.community/p/why-mediocrity-seems-to-be-the-key.
[3] Tyler, S. (2025, April 23). Taste is the new intelligence. WILD BARE THOUGHTS. https://wildbarethoughts.com/p/taste-is-the-new-intelligence.
No comments:
Post a Comment