================================================================================
C:\> CD .. Return to Index
================================================================================
[April 9, 2025]
Why LLMs Feel So Smart
It's not the jokes
* * *
There's a secret reason we're all so enamored with LLMs.
It's not the jokes.
Not the cute images.
Not even the agents.
It's the word count.
These things are masters of verbosity.
Lord do they spew. They generate GOBS of content.
It's captivating. It's technically impressive.
It's also… kind of useless.
Term Paper Tyranny
At a some point in school, my homework assignments started including word counts. "Please type your essay, double spaced, no less than 500 words."
Then came the talmudic debates about line spacing. Did it have to be exactly double? What about 2.1? Or maybe 1.5?
Then page 'margin theory' emerged -- how distant your words could credibly be spaced from the edges without looking like a Richard Prince joke painting.
Fonts were their own rabbit hole. (A classmate of mine swore by Garamond).
[Figure 1: All kidding aside]**
** ChatGPT wouldn't create a Richard Prince image "because the request violates our content policies". I love it. The guy practically belongs on the board of OpenAI.
But the real limiting principle was word count -- how to appear like you put forth maximum effort with minimum actual work. The logic went something like:
More words = more effort = more thought went into it = A+
I lived this. A teacher once slyly lauded my "dazzling array of verbiage" scribbled in the margins of a term paper -- it's stuck with me ever since. He was dead right.
The Verbosity Tax
But now comes AI.
The crescendo of verbiage, trained on all the maxed out word counts of the world.
I attended an LLM SaaS demo recently, the presenter pointing to a wall of text, "Then you get this response here." suggesting how imprssive this was.
Was it?
In a world ruled by the tyranny of word count, verbosity feels like brilliance.
Hallucinations alone might not even turn out to be the real problem with LLMs. Instead it's hallucinations buried in verbosity - a deadly combination of false precision.
They even have a word for this bullshitting.
Worse yet, the avalanche of text can require more time to sift through than writing it yourself.
Less is More
"The present letter is a very long one, simply because I had no leisure to make it shorter." - Blaise Pascal
e^(iπ) + 1 = 0
My team knows I've become an obsessive over Smart Brevity.
My own dazzlingly array of verbiage has been cut down to size. Hardly Hemingway, but I can dream. I'll admit my text-to-emoji ratio still needs to revert to the mean.
But as soon as I see an LLM start to go on, my follow-on prompts sounds more like this these days:
[Figure 2: Smart brevity]
George Orwell's "Politics and the English Language" needs to be seared into the system prompts of every LLM.
Writing in 1946, he observed a cycle of "ugly and inaccurate" (and verbose) language perpetuating intellectual decline.
I've been inflicting his six rules (gently updated for the 21st century) mercilessly on my team to improve writing and thinking:
└─ Never use a metaphor, simile or jargon which you are used to seeing in PowerPoint. ("Let's double-click.")
└─ Never use a long word if you know a short one.
└─ If you can cut it out, always cut it out.
└─ Never use the passive where you can use the active ("Mistakes were made").
└─ Never use techinical jargon or buzzword if you can think of an everyday English equivalent. ("Would a human being say this?")
└─ Break any of these rules before being a jerk.
So frontier models take note, if AGI's coming, it needs to get a hell of lot shorter.
Or else the joke's on us.
[Subscribe to more posts like this]
================================================================================
C:\> CD .. Return to Index
================================================================================