Hacker News Reader

Comment by block_dagger | Hacker News Reader

I've started feeling slightly physically ill when I read Opus output for hours straight. This article rings very true for me. I've started complaining about it with my team; at least have a personal style guide in your agent rules that eliminates emdashes, the "it's not X, it's Y"s, the long lists of modifiers before the noun, using the word "land" to mean finish, etc. I hope this is just a phase of adolescent LLMs.

View on HN

"That's such a clever way to see things! Let's delve into that!"

The bots (all of them) seem to show patterns of overuse of specific phrases, words, and punctuation.

Some of those are the ones you mentioned. Another that I've been seeing lately is overuse of the term "gate", wherein: As a human, I know what a gate is. A gate is a thing that can be open, or that can be closed. It might be locked or unlocked. The path beyond the gate may be passable or impassable or nonexistent. The gate is just a gate, and the presence of the gate doesn't imply whether it is open or closed.

But in bot-speak, a gate only refers to a hard block -- an impassable construct. Like a fence or a wall, or even a lava-filled moat.

But while a lava-filled moat is intended to be impassable, the bot uses "gate" -- a thing that is designed to be passed -- to describe that same kind of obstacle.

That's misuse of the term, I think, based on decades of dealing with gates in reality: Usually when I encounter a gate that is closed, I just open it and walk through.

I do have instructions that tell the bot to avoid that usage of the word and it ignores them sometimes anyway.

But "gate" is just today's problem-word that comes to mind as I write this. Yesterday, it was something different. Tomorrow, it will be something else entirely.

The overall pattern here is that of gratingly-repetitive bullshit-grade jargon that doesn't fit to begin with.

"And that's the real, no-nonsense truth!"

I like the word "botspeak".

Another example of typical botspeak is "smoke test". Why not just say "test"? It feels like a way of downplaying the ability to detect problems.

One thing I did recently with the bot definitely involved actual smoke tests, though: I was working with real hardware that can blow up in real ways, with the bot doing all of the circuit design work and coding based on my goals while I just distantly commanded the show from On-High and plugged shit into a breadboard.

(The project works well and I consider it to be Good Enough; I might go back and polish it more later. There was no smoke, but there could have been.)

This makes more sense if you think about the contexts in which people would talk about gates on the Internet, I dare say.

Oh?

I've been on the Internet for ~35 years. What did I miss?

My expectation is that you'd hear a lot more about "gated communities", "gatekeeping" etc. than any of the uses of gates that give warm fuzzies. (As a suffix, it's also associated with scandals; but that probably isn't relevant here.)

I was thinking about gated communities earlier today, in fact. We don't have many of them around here.

But where we do have them: At a given time, the gate might be open or closed; passable, or impassable. The presence of a gate is implicit, but the status of that gate is not known without advance knowledge or direct observation. And even when it is closed (even if it defaults to always being closed), there's generally a cromulent way for a person to get that gate to open and then move beyond it. It is designed to be opened and closed.

Gatekeeping: Sure. I've run across a ton of artificial gatekeepers online in my time. I've bypassed countless scores of them. Those are easy: Just ignore them and keep moving.

These aren't examples of the hard-blocking, impassable lava moats that the bot is fond of using "gate" to describe.

In the context of an LLM using "gate" within code: obviously one can always modify the code to bypass the gate, so there is a built-in implicit assumption that the gate isn't "impossible lava". Most readers are able to read between the lines, but you cannot serve everyone.

That's not how the bot uses the word, though -- at least in my experience. The bot uses "gate" to describe a thing that only blocks, and can never open.

Code that never opens is code you can delete. If the bot hasn't deleted dead code, especially given that this is code it recognizes as being dead and documents it as such, it must believe that you plan to use it again someday. That would imply an expectation of the gate opening eventually, even if not immediately.

Your description of this thing differs from my observation of the reality that I participate in.

Perhaps we're holding it differently.

Like all code discussions on HN, we have never been given the code to hold.

I understand in the past HN accounts have been afraid to post code because it would reveal that they have no idea what they are doing, but now that we have the plausible deniability of "look at this stupid thing an LLM wrote!" it is curious the mindset persists. Old habits die hard, I suppose.

I didn't realize that I was on trial and that there was a requirement that I back up my personal anecdotes and observations with working examples to prove their worth.

In reality, there's no reason for me to do so, because I'm not on trial here at all. Please keep your accusations to yourself, comrade.

My willingness to prove to you that a bot uses a word in a strange way is approximately zero.

Have a great day!

Logically there is never an expectation for one to comment for anyone but themselves, but it remains that nobody can hold something that has never been given to them.

So what you are saying is if you don’t want to read about Gates, target Linux instead of Windows?

I found Codex to use "gate" in a different sense: As the condition of an if statement. I have a local style rule not to use that. Another really grating thing is to use "X-shaped" for "something vaguely related to X". And using temporal words like "still" and "already" in contexts with no temporal connotation.

Me too. It feels like I’m taking psychic damage from reading so much of this stuff. Contrary to the theory that it’s “just the contract workers’ Nigerian English,” I think the models are developing an ultra-terse hyper-stylized dialect of their own under RL pressure. They seem to be writing increasingly in _code_, and I don’t mean computer code. The words don’t mean quite what they mean to humans.

Over the last few days with fable I've found it at times incomprehensible, terse word salad. It also invents phrases assuming I'll understand (but that could be because it's reusing terms in the codebase I no longer remember).

I've often had to paste its output back in to ask it what it actually means. Weird.

I think the main thing is just fatigue. There's so little variety. Each model has its preferred idiolect which everyone becomes tired of due to ubiquity. That's the worst part. It's like always eating fast food.

"It's like always eating fast food". That right there

My non-English-native-speaker head of development, to whom I report, does 100% of his work using LLMs and doesn't even check if the code compiles, but somehow this isn't my biggest problem with it – it's the botspeak in the PR comments (or answers to my PR comments) that are so clearly not written by him, and the documentation that makes absolutely 0 sense sometimes even if I break it down. Just a word salad of "robust", "maintainable", "smoke test" that amount to absolutely nothing. And the "You're absolutely right, I fixed it" responses (narrator: he didn't fix it).

I used to have a lot of fatigue due to it until I stopped caring.

comment reads clean. drafting response when it lands.

＊onanizing…

I was describing this exact feeling today. I haven't quite been able to put it into words but I do get slightly physically ill. Almost similar to mild trypophobia?

`arc land` is burnt into my brain by Phabricator, so I'm aware that the term predates LLMs, but it still drives me nuts.

It's impossible to undo some of these linguistic wobbles. Even if you could filter out 100% of LLM input, the humans themselves are learning to say "land" at a higher frequency now.

Voice really matters in writing. If everyone uses Opus to write without editing, then it all sounds the same regardless of who it came from.

I had to tell it never to say "hand-wavey" ever again to me. But I agree, I hate the way LLMs phrase sentences.

This has been my experience as well. It's incredibly grating to repeatedly read "genuinely load-bearing", "honestly?", etc. I've tried to get Opus to stop using these phrases via an entry in its memory, with mediocre results.

It's kind of offputting how much Anthropic models these days keep repeating "real", "genuine" and "honest". They've RL'd that way over the top.

"You're right for pointing this out. Honestly, your comment raises a real concern — they genuinely RL'd that over the top." - Claude

I had Claude make a world cloud from its responses because I was curious to see how big "honest" would be. It barely showed up, so I asked it to just give me the counts and it responded telling me it was trained not to use the word "honest" much because it makes people distrust responses (in addition to showing me the counts).

I just did this on one .claude directory and >20% of the answers there included some variation of "real", "actual", "exact", "honest", "genuine", "valid", "true". ~15% in that directory contain some variation of "real", "genuine" or "honest". This is excluding thinking tokens, sub-agent output, etc.

This is one of those things I barely noticed because I tend to read fast and skim. Someone pointed the over use of these terms and now its like hitting a set of spike strips every time I'm reading the output from any given model.

Its like when someone points something out a in picture you never saw and now you cannot "unsee" it ever again.

"genuinely load-bearing" is the one that triggers me the most now.

Key points only.

Anything written for humans should be written by humans.

Reply on news.ycombinator.com