Bias Compounds, Variance Washes Out
convergentthinking.sh
convergentthinking.sh
* bias compounds
* variance diffuses
* configs store parameters
* BF16 + RNE (6 bytes) plateaus
* errors repeat
* six bytes match ten
This sort of thing reads really well and conveys the idea in very few words. It's good writing! But in my experience humans don't generally "let nouns verb" as much as LLMs do, maybe we're just not as clever with words.
What's wrong with "configs store parameters"? I guess "parameters are stored in configs" could be more correct, but IMO it means exactly the same thing and sounds just as natural. "Six bytes match ten" is shorthand for "the performance of the algorithm that uses six bytes of storage matches the performance of the algorithm that uses ten bytes of storage". But here we have "performance matches", which is an inanimate concept doing something, so is this an LLM smell too?
Yes everyone says the sun shines and the wind blows, those are specific idioms. Noone says bias compounds or variance diffuses or six bytes beat ten.
I'm not saying they shouldn't! They probably should! It's just that LLMs say it much more than humans do.
> "Six bytes match ten" is shorthand for "the performance of the algorithm that uses six bytes of storage matches the performance of the algorithm that uses ten bytes of storage".
Yes, I understand this and support it. I am emphatically not saying it is bad writing. It's an unbelievably brilliant piece of terse writing that most human writers would not stumble upon in the course of writing the post.
I will rephrase: if you see a phrase in human speech that you have not seen used in human speech before, don’t penalise human creativity by saying it came from an LLM.
More specifically, the pattern you continue to latch onto dominates writing and has done so for decades. Right here on HN you can find instances of not merely “inanimate subject + verb” but specifically the phrase “bias compounds” from 5 years ago and beyond.
Other examples in use by humans all the time:
— River overflows
— Camera clicks
— LLM hallucinates
— Engine roars
— Secrets rest
— CD player stutters
— Ecosystem explodes
— Door invites
— Stock dips
— Airplane crashes
Antropomorphisation more generally and metaphor even more generally have existed since forever. Authors have played with form and tried to convey the point in different interesting ways since forever. Do you think Homer vibe-wrote The Odyssey?
Yes, LLM chatbots make it exceedingly easy—and it is one of their societal harms—but please do not discredit creativity by insinuating the author didn’t do the work themselves.
Nobody routinely says the things in your example without some supporting words.
And it's not just presence -- it's density.
And, to be clear: are you making the claim that this post was not LLM-generated or at least LLM-assisted? Or are you merely making the claim that people saying things like "nouns verb" might not be LLMs even though in this case the text is in fact most likely from an LLM?
Not really. At least one of my examples is literally based on a real headline from pre-LLM days[0]. If we are talking about a generalised idea of bias, there is no need for “the”. In fact, it would be wrong.
> Nobody routinely says the things in your example without some supporting words.
This is a post. One is expected to put more creativity into public speech. Good writing can be expected to be full of metaphors and denser than casual speech. Some writing purposefully uses headline style for impact. It doesn’t mean the author routinely talks like that.
> it's density
If it’s a wall of text with filler, it’s LLMs. If it’s dense, it’s LLMs.
You can find plenty of examples of denser, less legible posts from days before LLMs.
> are you making the claim that this post was not LLM-generated or at least LLM-assisted
I am making the claim I am making: don’t say someone used an LLM based on such a weak foundation and nothing else.
People will sound like LLMs. Blurring the line with actual writing is the entire point of LLMs and is fully by design; if you encounter a text with zero “tells”, it might as well be made with an LLM if the product works as intended.
[0] Here’s another one: https://ribbonfarm.com/2017/05/25/blockchains-never-forget/
But... it's not unusual in the slightest.
From a technical writing perspective, this is a terrible blog post.
Here is a better blog: https://cloud.google.com/blog/topics/developers-practitioner...
What I mean is that each sentence looks well made and from a highly skilled writer. But reading a few parameters I'm lost quite fast. Like if there was no coherence or progression. Just some ideas sometimes even repeated in random order even if in the end there is point/explanation to be made.
Like when Claude or another llm drops you a 5 pages block of text or MD spec that is totally unreadable even if it is supposed to make sense.
I think that it is unusual for human generated speech because usually if you have good with words and to do great sentences, you will also to do it in logical and coherent way at the multiple paragraphs level too. Or at the opposite, if you don't know how to redact proper paragraphs, you will. It not be able to do great sentences anyway.
>housing appreciating
>stocks appreciating
Are you an LLM?
But I don't think the examples you chose are very good. It's not "stocks appreciating", it's "$1m in stocks appreciating at 9%". The stocks are not an actor in that sentence. "$1m in stocks" is a thing that is having appreciation done to it.
If I had written a self-contained clause saying "stocks appreciate", that would have been a good example. But that's not what I wrote.
Which one is the actor there?
Help me learn, ESL.
There's nothing to learn from jstanley; he's a crackpot.
The subject of "appreciating" is "$1m in housing", or if you want to narrow it down as much as you can, the subject is the word "dollars". ("$1m in housing" is shorthand for "one million dollars in housing".)
I am definitely high risk for crackpottery, but not sure I'm quite there yet.
Which hasn't prevented you from expounding at length on your obviously false ideas.
What would you label yourself, if not a crackpot?
How? "appreciate" in this sense is not a transitive verb. There cannot be an agent.
In any case, I don't think you need to defend this. It's not about humans never using that pattern, it's about how frequent it is relative to LLMs. Individual counterexamples do not disprove a trend.
This makes me wonder whether you could apply different dithering approaches to numeric computations. You cannot use diffusion or similar mehods, because you don't have information about neighboring pixels/computations. Using low-discrepancy sequences might work to reduce stochastic noise, but it could also reintroduce bias for some computations.
The title is misleading, but I guess "bias–variance tradeoff" was taken :/
This was quite interesting though. Surprised to see it work so well on a real example.