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

Idle thought: you can have progressive JPEGs in which coefficients for lower-frequency components are sent first and the detail is refined as more data arrives.

I'm sure you can abuse this to construct a *regressive* JPEG that looks good with the initial low-frequency data and then degrades into something terrible.

Today I learned that there's something called a "smart ring" that you can buy for $400 plus a monthly subscription fee, and that needs to be recharged weekly, and it gives you horoscope-grade AI insights about your sleep

OK. Now with decals.

I don't find it surprising that people send LLM-generated docs, publish LLM-generated blog posts, etc. Writing is a chore and there's this big button labeled "get it done".

But almost no one is forthcoming about it: maybe 1% of AI-generated text I see credits the tool? People deny it even when caught red-handed.

I suspect one reason is that they still want the credit for pressing the button. But another factor may be that they realize there's no moat. If person A can prompt for blog posts, so can person B and C and D... so you don't want others to get the same idea.

@simplenomad To me, quality is secondary. I want to know if the text I'm reading comes from a human simply because human writing is rate-limited while LLM writing is not.

If you wrote it yourself, I know you cared enough to invest hours of your time into it, so it's probably worth minutes of my time to read and engage. With LLM-generated text, this no longer holds, so it's sort of "deny by default" because there are only so many hours in a day.

I guess it's sort of a fundamental question of whether we see any inherent value in the human condition, or the goal is just to maximize blog posts per hour.

Now, I'm not an anatomy expert but

AI children's books, body horror edition: https://lcamtuf.coredump.cx/blog/body/

You know you want to click.

The state of publishing right now

One of the major perks of large language models is that we can now effortlessly rewrite all the software written in unusable, esoteric languages such as Rust back to C/C++

Did you know that the latest version of the C standard supports operator overloading? It's true: https://sandbox.onecompiler.com/c/44s85vjf3

( Inspired by @drahflow )

OK, this one isn't perfect, but it'll have to do: https://godbolt.org/z/cbqY49YM6

An arguably more predictable consequence of the same property is that 2 + 2 = 2 (for some types of 2): https://godbolt.org/z/Y975WP7Kz

Calculator on glass. Yes, that's a glass PCB. Sharp developed a process to make an LCD on glass and I guess they felt no reason to stop there.

Edit: added side view photo

The coreutils Rust rewrite story is pretty funny.

Coreutils are tools like rm, mv, mkdir, etc. Unlike binutils, this isn't a fertile ground for memory safety bugs. But, the rewrite was completed, and in the spirit of progress, Canonical decided to switch.

But do you know what coreutils are a fertile ground for? Race conditions around file creation, deletion, permission setting, and so on. The original code accounted for decades of hard-learned lessons in that space. The Rust rewrite did not:

https://seclists.org/oss-sec/2026/q2/332

PS. I'm not dunking on Rust. It's just that... starting over from scratch has its hidden costs.

Do you own books about electronics, but could never make it all click? Do you know what a resistor does, but still get stumped designing your own circuits?

If yes, I wrote "The Secret Life of Circuits" with you in mind. All chapters are available now: https://lcamtuf.coredump.cx/blog/secret/

One of the funny patterns in AI-related discourse is exuberance about the latest model, inevitably followed a month later by claims that "they" (the vendor) are making the model worse.

My friend, that's probably just you coming to terms with what the limitations are.

Interesting. It would appear that something has changed on Hackernews around April 8. Up until that point, the daily line-up was pretty much dominated by stories written by AI or about AI. After that, it pretty much stopped.

(I've been tracking it for a while for fun.)

At this point, LLM-written think pieces make up about half of all long-form writing in my social media feed.

When I push back, I get two reactions. Authors say that it just helps them express themselves. AI promoters say "get used to it".

I don't think we should: it boils down to asymmetry. Our time here is limited. Social interaction on the internet breaks down if it takes ~0 effort to publish, but readers are still expected to use their own eyeballs and brains to engage.

So, I feel that we have three choices:

1) Refuse to engage with LLM writing *no matter if the article makes a good point or not*.

2) Embrace it and have my agent argue with your agent forever, for internet points.

3) Call it quits and move to an off-the-grid cabin in the woods.

I'm dying. With this translate tool, I'm finally able to effectuate all my ideation synergies

If you ask AI to rewrite the entirety of an open-source program, do you still need to abide by the original license? In philosophy, this problem is known as the Slop of Theseus