On The Slop Doctrine

2026.08.23

CLXVI

[The Slop has Got to Stop; The Syntactic Sins of Slop; Helen of Troy was a Hoe; 'Handwriting' Code; Deep Talking; Consistent Worldview; Why We're in Trouble]

[The Slop has Got to Stop]

I'm sick and tired of slop!!!!!!!!!!!!!!!!!

Not just AI Slop, to be clear; there is human produced, slop, too. It’s seriously everywhere.

I fear our proximity to LLMs has caused a negative feedback loop that makes us write and talk and (god forbid) THINK more like them. There's evidence for the first two claims (writing, talking), the third is conjecture, but if you hold that writing is thinking, you get the third claim for free.

Regardless, we’ll call the propensity to produce slop the Slop Doctrine:

For some reason, people seem to believe ‘more’ of everything is preferable to less, and outputting a stream of low value and even misleading information is better than even trying to find truth.

It’s really everywhere: last week, someone showed me an email he got, and I immediately said it was AI generated. He asked me how I knew, and at first I was at a loss for words: I thought everyone could just tell! Then, I remembered a call I had with a client who was trying to send AI generated emails that didn't look like AI, and I had warned him that it was AI, but then had to explain how I knew.

In other words, whether it's my youth or chronic online-ness or heavy LLM usage or use of it in my own product, I can quite immediately flag AI slop, but not everyone can for some reason.

So today, I'll do the public service of explaining tips and tricks for identifying AI Slop. Much more important than that, though, I'll go on to explain what SLOP actually is, largely by contrasting it with what it's obviously not, and why it's so important to avoid it.

I’m sure you all think I'm being dramatic, but the fucking future of Western society as we know it depends upon your ability to sift through the slop and find truth.

[The Syntactic Sins of Slop]

Okay, so, identifying slop. To aid us in this exercise, here's text from the LLM query that I used to find the research papers I cited above*:

In this small excerpt, you'll notice a few characteristics, some obvious, some not:

• The opening sentence contains the classic, obligatory agreement, followed by a correction or further distinction that is sometimes, meaningful, sometimes not.

• Giving a dissertation on why I’m partly right but also not really - my prompt was literally asking for evidence, not the opinion of something that can’t have an opinion

• The unnerving use of "I", as if the LLM were actually a person or sentient. It adds nothing to the argument (why does it even have an argument right now?), barring perhaps a play on the human desire to personify and therefore give more credence to everything remotely human like.

• Use of the em dash…

• The sentence "LLM assisted..." has a list of commingled points (what's the difference between 'cleaner' and 'more concise'?), ending with emphasis on the last (maybe distinct?) point in the list

• We could trivially compress the above sentence while keeping all meaning: "LLM-assisted writing converges with other text, particularly the concise & conventionally formal academic prose." 40% shorter after one human attempt while preserving all meaning.

• "An even larger...": This section contains a very classic LLM sentence structure; "X isn't y, it's z"! I see it so much in slop I’m afraid to use it myself; it’s an otherwise fine sentence structure, just not when you use it every 5 sentences…

• Below that, another hallmark: Short sentence that emphasizes the last point before re articulating it with more text, "That distinction matters enormously"

The text goes on and on and kind of did what I asked for (identify some research papers) but also just word omitted a bunch of truisms that sound like they might make sense but maybe do, maybe don't.

I'm worried, because there's no way that staring at this shit all day doesn't have impact on how I think. Perhaps I'll write my own wrapper for an open source model that is incredibly functional and actually behaves how I want, which is just being a better search engine or sometimes writing a sql query.

There's no real benefit to being more verbose or writing or talking like an LLM, it's all just extra damage to my soul and my brain. It's even making me paranoid as I write that I start sounding more like it and less like me!

And there's certainly no benefit to the internet being flooded with this bullshit, even though something like one third of internet text now comes from this nonsense.

If AI Slop replaces truth with empty rhetoric, this is an insane danger.

_{*It's not even very good at search, as it’s too busy arguing to realize that at least one of the citations in the massive response was fucking fake! But, I'm using it, so I do think it is more efficient than, at the least, google search.}

[Helen of Troy was a Hoe]

The real underlying issue is that these text generators favor rhetorical devices and surface level ‘true sounding’ patterns over actual truth and knowledge.

I'm reading Herodotus, a severely entertaining Greek Historian. Really trying to understand how the Greeks were so good. Thucydides next, because I’m also worried that we’re the Greeks and China is Rome. But, all that is such a wishy washy unsubstantiated thought that I’m almost embarrassed to put down without warning you that it’s potentially noise pollution. (If only LLM’s gave this much warning.)

Anyways, Herodotus is in two notable ways, very unlike slop, in that he can coherently hold multiple narrative threads despite page long digressions and that he is constantly attempting to create a consistent world view, and in another way, quite like slop in that he spews a lot of unsubstantiated bullshit.

A funny piece occurs quite early in Histories, when Herodotus sets the state for the Greco-Persian war by recounting a series of 3 or 4 times which Persians and Greeks stole women from each other, ending with Helen of Troy and the Trojan War. The whole thing is arguably a digression to begin with, but even then, he can't help himself but question the Greek's for starting a war over a woman:

Apparently, he thinks Helen of Troy is something of a promiscuous woman and views the Persians as more blackpilled than the Greeks.

All of Herodotus’ writing is like this: it is a long, winding narrative where he starts to say something like, "Cyrus was getting ready to invade the Ionian people, but who were the Ionian Peoples? Well, I'm glad you asked..." before going on a multi page digression where he gets all the way down to a subset of 5 Ionian tribes that banned the 6th from a shared temple because some dude stole an expensive bronze tripod 100 years ago. Then, he gets back to "Now that we know everything about the Ionians, we can resume at Cyrus's invasion. The first city he attacked..."

Even his sentences are like this: he starts with one clause & adds in a million sub clauses before getting to a meaningful conclusion.

The point is, though, he always gets back to the meaningful conclusion.

This is distinct from slop. Slop can diverge from the subject and lose itself and get off topic but act like it was on topic the whole time.

Keep in mind, Herodotus is something of a bullshitter, an ancient yapper; some of the things he says are utterly untrue. The source for some of his information reads as "trust me bro I was there", others reads more like this little clip.

In this, Herodotus is actually something like slop: saying things without having any rigorous backing.

That said, even given some of his bad priors, he is actively attempting to construct a consistent world view; he cites different peoples who disagree about an order of events, he weights contradicting hypothesis, identifies when things are mutually exclusive, and says, "So here's what I think happened." And even if he writes with a slop like hubris that can seem foolish with the benefit of the more complete information we have now, he is actively trying to construct a picture that works across itself.

The Slop Doctrine doesn’t have time for this exercise.

['Handwriting' Code]

When I was in the hacker house in Vienna, one of the residents was walking behind me while I was coding, and was dumbfounded.

Him: "What are you doing?"

Me: "...coding?"

Him: "Like you're ACTUALLY coding. Why aren't you prompting?"

The kid (I can call him that, he was like 19 or 20) was utterly perplexed that I would take the time to actually write out any code whatsoever.

I explained to him that I don't NOT use AI, I just don't let it in my IDE, and I don't let it steer.

He couldn't get his head around it. Doesn't everyone just rip Claude Code or whatever? That's certainly what the internet makes it seem like!

Over the past few days, I've been writing another scraper that can take in hundreds, perhaps thousands, of RSS feeds, and check them for articles, extracting a consistent & useful data format for the rest of my pipeline. This includes some complexities like caching of the full body or even partial body text when available for use across multiple machines without blowing out our core data base with cruft that might never get used.

This is a fun & important project. I have an LLM help me if I need some code with some syntax I don't understand (it's my first time using PyArrow and a couple other libraries), but I'm still driving and writing a lot of it. Hell, a lot of it is trial and error and iterating on multiple RSS Feeds as we scale it up.

Really, though, I am writing a piece of software to join my corpus of five or so other scrapers, many of which have worked for more than a year at a time without any updates. If you don't realize the leverage of this, you probably haven't dealt with the pains of pieces of code that keep fucking breaking and causing you to have a high maintenance tax. That sucks, it's low leverage, and wastes time that could be spent making other things better, not fixing things that you already told people work.

And then, if I understand the code as is possible when I write it, it becomes a lot easier for me, when there is a fix or a quick edit or update, to do it myself, without having to have an LLM parse through the pile of spaghetti.

Yes, I know you've heard this rant before. I'm bringing it up now, because the anti coding marketing has been SO effective in certain circles that when someone sees me coding, they literally didn't realize this was a thing that people do.

Again, yeah, I use AI, it can make you go faster if you use it well. But, it can also just shit out a lot of hard to maintain slop for a product that shouldn't exist.

Fitting something in the mind of a human involves this strange process of adding and culling: you add something, add another thing, and then realize you can delete half of each of those things to merge them together. And, you're the one making that decision, so you understand where the code is going and what's left.

All that is to say, in spite of the prevalence of the Slop Doctrine (speed is all that matters!) net net, I’m still not convinced that letting the LLM do everything is the right way. Something happens in your brain when you’re not just prompting.

[Deep Talking]

The other night, I was talking to EH, who does quantum research & lives at C House. This was one of those conversations that kept me up an hour longer than planned. Basically, he was explaining to me the fact that we can actually teleport information as far as 70 miles by entangling two particles and changing one in one place; it then changes in the other, as a consequence.

This is a very complex topic and fascinating to me, as it's one of those things that when we start discussing, my immediate reaction is 'bullshit,' which is followed by an appreciation for the sorcery modern math is capable of and a yearning for continued scientific progress.

Watching EH explain the technology was fascinating to me, because I could tell he was actually in a state akin to Deep Work, we'll call it Deep Talk. He wasn't reciting facts he had memorized; rather, he had a collection of 'primitives' that he was reconstructing statements from on the spot.

This is quite distinct from a Slop Doctrine endorsed conversation where you just stream of conscious things with no self referential integrity.

I do this when I am explaining options pricing to someone. I don't have the math behind put call parity memorized. Sure, if I explain it frequently, it comes quickly enough. But I found myself explaining it to someone for the first time in years the other day, and I found that I was explaining it to myself, too, solving the problem out loud.*

Something goes on in the brain in these cases (certainly more so in EH's while he is explaining Quantum than mine when I am explaining pretty simple options math). It is the opposite of slop. It involves structuring a world view that is not only self consistent, but consistent with the outside world & reality.

You're 'tapped in' to something greater, if you will.

And, unfortunately, even the world quantum is full of slop: according to EH, qubits are only useful when you can entangle them and make a lot of them; companies are raising massive amounts of money because they can produce qubits, but it is neither in a meaningful quantity or entangled. It's not regrettable that people are working on at least part of the problem, but it is regrettable that they're posturing as if they've achieved a commercially viable solve to make the technology useful at scale.**

_{*If you want to try, and you know enough about options, do the following (setting the scene is at least 50% of the work, you're welcome): Take a stock at $100, a 100 strike put that is $.50, and a 100 strike call that is $1.00. You can short or long the stock, the call, and the put. There is a 3 legged trade that nets you the $.50 spread between the call and put the for 0 risk. Assume no transaction fees, no bid / ask. When done, repeat with $1.00 put and $.50 call. Once you've done this, if you assume No Arbitrage holds, you've just proved why at the money puts and calls must be worth the same.
**I just pulled a Herodotus - 'Some guy I talked to who I trust told me this thing I'm going to assume is true.'}

[Consistent Worldview]

2 or 3 times this year, when explaining some controversial opinion I have, I've gotten quite the compliment about how my worldview is consistent.

This is very important to me. It's something that's hard to do and requires rigor.

But it’s fucking awesome. It makes life a bit easier - it gives us a base to stand on and to know & build more.

The Slop Doctrine, propagated by LLM dependence & slop in general are NOT optimized for veracity of even consistency.

Yesterday, I had a crash out on the IDE I use, Zed, because I had to manually save a file when I switched panes, even though I had auto save on; and, because it kept auto formatting some text, even though I had the setting to do that off.

This made the tool feel sloppy as hell, and created some big problems that wasted a lot of time. So, I stopped using it. It said it would do something, then it did something else. This is annoying, it is sloppy! It’s not even consistent with itself.

Neovim for the win, I guess.

This mirrors the frustration I experience when I use an LLM that gets something wrong, then goes on to say "Oh my god, you're so right, I'm such a silly goose!" and then MAYBE correct the mistake, or maybe make it again later!

This is slop. There is no consistency. There is no attempt to understand the world in a real manner. And you can't really blame it, that's just how an LLM works.

What you can do, is not let it corrupt you and use it while being on guard. There is a truth we can strive for, there is some consistent world view we can get closer to.

I am a technical founder on a 2 man team bootstrapping a B2B SaaS to a $1M ARR without hiring while somehow finding the time to concern myself with how we can build the best future - I write this every week.

[Why We're in Trouble]

Okay, given all my rambling, why do I actually think that slop is a systematic risk to modern society?

Scientific progress (you know, that crazy thing that can save your life when you get in a car crash, or let us fly around the world in planes, or talk to your relatives across the country INSTANTLY) is based on having a corpus of knowledge that is constantly re evaluated, edited, and hopefully, built upon.

I'd put slop at direct odds with it. Perhaps I'll even take a stab at defining slop:

Note that I didn't say slop is inherently inconsistent; sometimes, as an accident, it is. The point is, it makes no EFFORT to be consistent. It's noise over everything else.

We're in trouble, because if we all think this way, and argue for the sake of arguing and not for finding truth, what happens? Do we accrue more knowledge? Does our understanding of the world get more powerful and useful? Or are we better at entertaining each other and sounding smart?

I don’t come to you with a great solution yet, other than to be on guard against other’s slop and not to produce it yourself. You can leverage LLMs without producing slop, and you can produce slop without LLMs. Use LLMs as a tool, sure, but don’t let them compromise truth for noise.

Live Deeply,