Notices by lcamtuf :verified: :verified: :verified: (lcamtuf@infosec.exchange)

So there is this Facebook page with 1M subscribers, and an associated network of content creators with tens of thousands of followers, that's essentially dedicated to posting nonsensical, AI-generated circuit "schematics".

And it's interesting because... 1/3

They don't exist to meet demand; they exist to game the engagement systems we put in place. I think that's true for a lot of other AI content on the internet too.

3/3

...the usual argument for AI-generated articles is that they're just good enough for most people, so there's no point in getting your panties in a twist.

But you can't argue that these schematics are useful. No one can build these circuits. They make no sense!

2/3

maybe maybe maybe

Perhaps you know that modern digital-to-analog converters often use trains of digital pulses to construct analog voltages. But did you know that a wacky variant of that method also works for converting analog data to bits and bytes?

https://lcamtuf.coredump.cx/blog/conversions/

If you're planning to visit Europe, be sure to bring the right adapter with you

Hmm, the visitors to my website seem to be really interested in React, I wonder if I should install it

Me: I want to have more friends

Tech companies:

Not bad: https://clocks.brianmoore.com/

As an occasional writer, I'm more sensitive to LLM-generated articles; I'm also more bothered by them than the average reader. At this point, I see this kind of content daily on HN and in my social media feed.

At the same time, I think there's a lot of witch-hunting going on. It's important to realize is that LLM writing doesn't differ from human writing *per se*. Rather, it's that if you go to chatgpt.com and ask it to write an essay, you always get the same writer persona by default. That persona has distinctive quirks, just like any author would.

The quirks include basic formatting (e.g., a preference for book-style capitalization of section headings), a range of stylistic cliches (vague appeals to significance, negative parallelisms, explicit summary sections), and so on. Contrary to some nerd lore, these statistical patterns can be detected with automated tools; if originality.ai flags something as LLM output, it's a signal you shouldn't just ignore.

That said, there are two problems with relying on these indicators. First, there's a small but non-zero percentage of people who write the same way. Second, if you're clever, you can prompt an LLM to use a different voice. Tell it to write as a grizzled sailor - a man of few words who's still grieving the loss of his wife and only child - and it's going to generate something quite distinct.

Because of this, I find it useful to lean on three additional, higher-level heuristics:

1) Does the text exhibit any non-stereotypical writing traits, and do they appear consistently in this person's writing? Almost everyone has some. Preference for overly long sentences, overuse or complete avoidance of semicolons, unusual metaphors or idioms, etc. If you don't see any of that, it should give you a pause.

2) Is there a discernible reason why this text exists in the first place, and why this person is writing it? Long-form writing takes time and effort, so we only do it when we think it's important or useful. If there's no discernible personal perspective in the writing, that's a red flag.

3) My favorite: if you prompt an LLM to write an article on this topic for this particular publication, is it going to produce something functionally similar?

This brings me to my final point: as an author, I have some misgivings about the tech, but I recognize it can be used for good. My beef isn't with people using it; it's with people who use it to crank out slop. If your article fails these three tests, then something went wrong even if you didn't use any automated tools.

This isn't wrong

My favorite fact about recycling is that 100% recycled plastic is often worthless, so instead, you add 5% of recycled plastic to a vat and then you make 10,000 shampoo bottles out of it, and then you take 500 of these bottles and label them "100% recycled material".

Which is both not wrong and terribly wrong at the same time

October is Cybersecurity Awareness Month! Please be aware of cybersecurity. If you encounter cybersecurity, DO NOT APPROACH IT. Back away slowly. Protect children and pets. Make noises to scare it away.

"In Desert Bus, the player drives a bus along a straight road between Tucson, Arizona, and Las Vegas. There are no other vehicles, and the scenery consists only of dead trees and bushes ... At a top speed of 45 miles per hour and a route length of 360 miles, one ride takes eight hours in real-time. The bus ... veers to the right, requiring the player to regularly counteract, and the game cannot be paused. Completing a drive between the two cities grants the player one point"

Remembering the original AI agent

My wife asked me some time ago to build an automated door opener / closer for the coop, and needless to say, our design firm takes this very seriously

I sometimes wonder how life be look like without HN comments on my articles

I'm working on a universal bucket list for corporations:

1) Declare that you're a different kind of a company

2) Design and announce your own typeface

3) One year of free credit monitoring

telnet uses port 23 to indicate that it's an improvement over ssh, which uses port 22

A slightly unhinged calculator fact: in the golden era of electronic calculators, some Japanese shopkeepers were reluctant to trust the newfangled tool, so Sharp made a line of combination calculator / abacus devices.

Here's a photo, next to some other stuff I own.