I now spend many hours a day reading LLM generated text. Summaries of systems and interactions, guesses at what caused a bug, implementation plans, pull requests descriptions. Pages and pages of flabby prose, which I skim, and then offer a pointed response, increasingly impatient. “NO don’t cast just use the type”, “why is this needed? seems pointless”, “KEEP IT SIMPLE”.
My productivity has never been higher. I can accomplish things that I would have never attempted before. The foundations can feel insubstantial – like sand moving under my feet – as I resist being led down a mistaken path, as I challenge strange and confident presumptions. But there is no doubt this new, alien tool is powerful, and I enjoy my work as much as ever, maybe more.
Yet something has been gnawing at me for a while. I think it’s all the cliches. I resent having to read this bilge in order to do my job.
You’ve always been exposed to cliches when working in business and tech. Management is famous for its jargon (I will loop back on this) and the software world has its argot of happy paths and yak shaving. But there was an essential limit to the volume that could be produced: an actual person had to spend a moment of their life opening their mouth or tapping at a keyboard. It already felt like a lot, and now there is more.
Since I had this realisation – these cliches are killing me – I started to note them down. Somehow this made me feel better. I shared my quest with colleagues and maybe it helped them too: “I would have said hidden before but I think it’s made me say opaque”.
I’ve apparently been collecting LLM cliches since 15 June, mostly from Claude, although ChatGPT seems to pull from a similar pool. They cluster into a few groups and here is my attempt at a glossary of current LLM cliches:
- load-bearing. Anything that does anything.
- backstop, belt and braces, defense in depth. The opposite of load-bearing, anything that does nothing.
- smoke test, smoking gun. Evidence for something. Both smoke related, I now realise.
- keep honest, being honest, honest caveat. These possibly leak from prompt/training attempts to control confabulation?
- hand-waving. The opposite of being honest, making shit up.
- instinct, your read, called me out, this flips my lean, reframe, gets to the heart. What the bot says when I apply my professional experience to a problem. I take “instinct” unreasonably personally – lucky guess, meat bag. Read and lean as nouns is annoying.
- (single) source of truth, canonical. A parting gift from the Jira-cleric class as they drift into the good night.
- north star. A direction (specifically: north).
- one wrinkle. A catastrophic problem.
- in-flight, mechanical, fast follow, big lift, blast radius. Words for doing stuff.
Not all of these are awful but, for me, all have grown tiresome from overuse. Some are truly dreadful and “this flips my lean”* is the worst.
This post is largely therapy: “these are my cliches, what are yours?”.
But perhaps the concept of the cliche opens up a deeper point about how LLMs invisibly diminish the breadth of our ideas. The endless repetition of the cliche certainly grates and reduces our own vocabulary (you are what you eat). More generally, though, I believe that the way that LLMs present us with language that emulates finished thought in turn restricts the range of ideas that we even explore.
The “blank sheet of paper” problem is no more. We no longer have to confront an empty document or code window and make something from nothing, in fits and starts, as we think through the puzzle and build up an answer in our minds, inchoate and incoherent at first, slowly solidified by small revelations.
Instead, we offer a vague prompt and get pages of instant babble in return, ending with something like “should I just apply fixes 1-3 or tackle the wider refactor now?”, like these are the only possibilities in the universe.
* my wife made a comment about what this rhymes with which is unpublishable