I’ve written some words ✍🏾and they are about why you should be writing more words 🤯

Too many influencers say dumb things about companies and their products. It’s boring when people make videos listing off stats, making comparisons to other products, and saying how much things have declined. Meanwhile, the company did market research and decided where to allocate resources.

It’s reasonable to criticize companies. But it’s pretty boring when people read off spec sheets and list off prices as if a company’s product portfolio is a series of independent products instead of an integrated set.

LLMs are terrible at making taglines. They make weird tagline shaped things that don’t make sense when you actually try to parse or think about them. The beauty of a tagline is that it’s one of those “a few words is worth a thousand pictures” moments and you can’t easily predict that sort of thing.

Reading output from LLMs is like reading the partial output of someone’s thoughts. I thought I didn’t like the constant hedging and contrast that LLMs produce, but it’s actually that the hedge comes out of nowhere because it’s missing some early assumption the LLM made but never output.

The d in dvcs being “distributed” instead of “decentralized” is another one of those unfortunate naming things that has cursed us. Perhaps if we called Git decentralized people would have resisted centralizing it. Or at least it would seem as silly to do that as it actually is.

I really wish companies would stop shoving crappy AI features into their products unless they’ve made the core thing stable and solid.

All trades are creative and all trades are analytical. The idea that some forms of work are lack creativity is falsehood we sort of made up. Disciplines like accounting, engineering, math, and plumbing all require a great deal of creativity.

On the other hand, disciplines like writing, music, and art all require analytical thinking. Keeping canon correct, fitting the pieces of a poem together, making lyrics fit well with rhythm, there are rules one must follow to do it well, and specific rules one can break to create something new.

We should stop thinking of disciplines as creative or analytical or technical and start thinking of them as the complex mix of those things that they actually are.

“That’s the part that stuck with me”

There are too many sentences now that have ascended so rapidly to cliched stereotype that it makes one’s brain twitch when read. The choppy blocks passing as sentences in LLM generated writing seem plausible only if one mixes all the genres of writing together.

In hindsight, the idea of NFTs wasn’t silly, it’s that they were targeted at a market of people who doesn’t understand why easily copied objects are worth owning the original of. NFTs never reached the point of having the necessary legal structure around them either.

So many of the arguments around NFTs didn’t make sense but they sounded like zingers so people gave in. Like, there are plenty of house developments with hundreds of carbon copy houses, but people don’t say those houses are worthless because there’s so many copies of them. Similarly, no one says that famous paintings are worthless because they’ve been mass reproduced.

NFTs could have ascended into the class of objects used for wealth purposes, but that market never developed. But the underlying idea is actually a quite good one. It’s similar to how the idea of a decentralized public ledger is useful in a whole bunch of ways.

And while NFTs didn’t work out, I’d put them in the same class of idea as the prediction-market-using-futures-contracts idea that’s exploded recently.

Sometimes ideas flop, but that doesn’t necessarily mean the idea was bad, it might just be that it was aimed at the wrong demographic.

Too many people don’t know what liberal arts are or what actually falls within their scope. Also, maybe people confuse liberal arts with humanities? Although even then, it doesn’t seem like enough people know the scope of the humanities either.

Feels as good a time as any to remind you that *distributed* and *decentralized* are not the same thing. They are orthogonal actually, existing within different domains.

It’s a lot like how serializability and lineaeizability are not the same thing and also orthogonal.

It’s going to be real funny if this class of models that we’ve called “generative models” actually wind up being so much more useful as interpretation engines that we effectively stop using them for generation.

That would also be 100% aligned with how naming in this industry goes.

As much as I dislike LLM generated writing, I think I hate it more when it winds up as the script for a YouTube video. LLM writing sounds worse when read aloud.

“Here’s the part nobody tells you” has been cliched so hard we should probably eradicate it from language for a good long while. It’s shoved in so many LLM responses that its effectiveness has been reduced to basically nothing and it’s just noise now.

We should really just move past the quippy phrase era of communication. LLMs are gonna ruin it anyway and they weren’t nearly as useful as they seemed in the surface. Turns out a phrase is worth a thousand images, and it’s just as difficult to understand the nuance.

People who like to predict the future seem to forget just how good humans are at pattern matching. When people predicted LLMs taking over writing, there wasn’t enough of it for people to pattern match on it. Now there’s plenty we people do not like it.

People can be so multifaceted. It gets annoying when so many of us try to box people into specific categories and don’t allow space for them to be their full selves.

I’ve run into this sort of thing a bunch of times from people who meet me as a software engineer and make some off hand comment about me “not being a creative”. Like, the word “creative” is literally in my college major.

Generalizations help us move through the world, but we often need to stop and update our understanding before running our mouths.

I love etymology, because what do you mean that cliche and stereotype are deeply related? More than that, the figurative senses come from the ability to print some repeatedly, without change.

In this way, LLMs are very much cliche machines. They’ll take any novel phrase and stamp it into monotony.

It’s unfortunate that documentation produced by LLMs is obnoxious to read because I would love if we could offload a lot of the documentation work in the world to LLMs. But we can’t. Because reading LLM generated writing is fatiguing, which is the opposite of what you want your docs to be.

People really love blaming complex problems on singular words or terms. And yet the actual simple term that underlies everything is never the one mentioned.

It’s understandable: people want to blame something that doesn’t make it their own fault.

The reports of the death of writing code by hand are greatly exaggerated.

It’s cool and hip to go with the flow of what everyone else is talking about and agents are a form of compilation but there are important caveats to this.

First, compilers and higher level languages enabled us to greatly reduce the amount of code that we had to write. That is, expressing the same program in C and an assembler language generally requires less C code. Natural language and especially English, are verbose and expressing the same idea in them compared to a programming language is often more verbose.

Second, compilers are useful largely because they allow us to detach from individual pieces of hardware and write software that works across hardware. There isn’t really an equivalent of this at the agent layer. It could be argued that it’s actually an inversion, since different agents produce wildly different output (counter point: different compilers often produce different output instructions from the same source).

Third, compilers struck an important balance between the operator knowing how the machine works and being able to operate on an abstraction of the machine. We get worse software the less the operator of the compiler understands the machine. This is greatly increased at the next level up.

Agents are absolutely here to stay, but it’s rather silly of us to assume that their evolution will go linear at this point. It’s far more likely that the pendulum swings back in the other direction and we wind up back in IDEs looking at and writing a lot of the code ourselves.