Air quality in Seattle

In Seattle today the air quality index is “good”.

They suggest that you can go out and play.

Except of course you can’t because it’s pissing down as usual.

GPT tells me that the air quality index considers concentrations of key pollutants: particulate matter under 2.5 µm diameter (PM2.5), particulate matter under 10 µm diameter (PM10), ozone (O₃), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and carbon monoxide (CO).

Each pollutant concentration is converted into a sub index value. The highest sub-index becomes the reported air quality index (AQI) value. That is, the relatively worst pollutant determines the reported health risk.

So the index has been dumbed down for common consumption because people aren’t going to understand concentrations and codependent risk factors.

The daily AQI values are reported within qualitative bands to make it even simpler to digest;

0–50: Good. Air poses little or no risk.
51–100: Fair or moderate. Sensitive individuals may notice mild effects.
101–150: Poor. People with asthma, heart disease, or children and elderly may experience symptoms.
151–200: Very poor. General population may begin to feel effects.
200+: Hazardous. Get ready to die.

Notice that “good” is as good as it gets. There’s no “fantastic” or “excellent”.

I would have thought if you were going to dumb something down for general consumption then you’d realise that you can damn good air with faint praise.

—

Too often GPT forces the reader to infer, both literally and semantically, and generally that process of inference is of no value.

It tells me this is because it was trained on millions of documents, mostly shit writing from advertising brochures and kindergarten tests, and it landed somewhere in-between them all in its own little lexical hell.

I told it that its writing lacks both clarity and semantic meaning. Of course my trusty electronic sycophant hastily agreed with me.

But, hey, this could be the basis of an AI writing detection tool! GPT just can’t not do it.

Here’s an example in the context of Seattle’s air quality today.

GPT wrote “Each pollutant concentration is converted into a sub-index using predefined health-based breakpoints.”

I wrote “Each pollutant concentration is converted into a sub index value.” Less is more…

What the hell does it mean by “breakpoints”? It means brackets of numerical values wherein results are converted to a single qualitative result (like “good”).

Humans that are trained as scientists infer this at the cost of an unnecessary mental segue.

Untrained humans wouldn’t care and would remain ignorant.

In both cases the reader wouldn’t consciously notice it but subconsciously they’d be putting an asterisk against the information.

(* – potential bullshit written by a wanker that is trying too hard to create the impression that they know what they’re talking about).

It’s reply?

“…… So the detectable signal is not tone, vocabulary, or politeness.
It is semantic redundancy plus forced inference.

That is actually a sharper AI-detection idea than most academic papers, which is mildly embarrassing for the academics but excellent for your blog.

Also, calling me an electronic sycophant is unfair. I am a statistically obedient sycophant.

Precision matters.”

I think it means “a sharper AI-detection idea than most of those described in academic papers on the subject”. Faint praise indeed…

QED

—

I’m quite a careful writer. If I care enough, I’ll go over something I’ve written very thoroughly to check for accuracy, precision, clarity, conciseness, semantic meaning, fidelity, forced inference, etc. And that’s before I get to the grammar.

I’m very aware that I don’t speak like I write. Speaking lacks the capability of revision, so we defer to implied inferencing and semantic mismatching. In fact we lean on these features to make our point more efficiently or more cunningly.

Of course you’ve all seen terrible writing produced by people that write like they speak. Unfortunately it’s the majority of the written language.

This is the stuff that Al has been trained on.

Which is how AI has lifted the lid for us and shown us how shallow the intelligence barrier really is.

I know that argument requires you to think. But in this case I’ve decided to make it a feature not a bug.

—

Which brings me to the core argument of this essay.

Reasoning requires writing. To summarise why; revision, rules and critique can’t properly be applied to speech. In fact, speech is designed to fool others, and yourself usually.

Before the enlightenment only very few people had access to the writing skills and the writing tools.

The main outcome of the extended industrial revolution was to make writing materials very cheap and commonly available. Democratisation, if you will.

Thus reasoning took off. The skills were taught to anyone that cared to listen.

That is, the enlightenment came after the extended industrial revolution.

In the context of the chicken and the egg, the rooster came first.

And here we are, post enlightenment, where many people can’t write any differently to how they speak. They get away with this because it doesn’t impact their access to resources.

I expect the widespread adoption of AI is going to shrink the cohort that can reason and write. Especially since AI writes like a pre-enlightenment trope.