Notices by Sophie Schmieg (sophieschmieg@infosec.exchange)

On top of that, any junior dev I've ever worked with asked questions when they got stuck, instead of confidently force merging a bunch of buggy code. And not infrequently, those questions laid bare some flaw in the design decisions of the more senior engineer, and let to a better product that way.

You didn't really think the business value of that PR that had more comments than lines of code that all of us have written at some point in our career (if you haven't, I'm sorry, your onboarding mentor sucked, they were supposed to point out all the stuff you had to learn the hard way) was in getting it merged eventually, right? The value was the PR after that, that had a lot less comments. And the PR after that one, etc.

Whenever I see "GenAI basically can do the work of a junior dev", I assume that those people are using junior devs very differently from me. To me, the value of a junior developer is maybe like 20% the code they write, and 80% is a down payment to a future where they no longer are a junior. I still haven't found an AI that could do that second part.

Criticizing Linux on Mastodon is really the best way to make men type paragraphs, isn't it?

Nota Bene: since this story took place in Germany, the math teacher in question was quite well paid and had a university level mathematics education to fall back to, which, from having taught that course myself later, should include a full class on geometry, teaching not just the Euclidean, but the Hilbert axiomatisation. He also taught the math extracurricular, so he was very much up to the challenge 🙂.
The triangle on the stand was still acquitted on account of being planar itself, so the story had a happy ending, and no triangles had to go to geometry jail.

Random childhood anecdote, posted as individual toot in order to not derail some random joke post about kids these days not being into non-Euclidean geometry:

I have three siblings, spaced two years apart each. All four of us went to the same high school. My youngest sister's math teacher wanted to do a cute little segment about triangles always having inner angles sum up to 180 degrees. It was supposed to take the form of a trial. My sister was assigned the role of prosecutor charging a triangle accused of having an inner angle sum different from 180 degrees. Obviously, taking her role very seriously, she consulted the Schmieg dinner table, and a plan was hatched, involving a grapefruit and a marker. The trial came, and just as the defense had produced what they thought was irrefutable proof that their client was innocent, my sister took out the grapefruit, drew a triangle connecting one pole to the equator using three right angles, and said "So how do you explain this?!?" Leading to the teacher/judge having to explain non-Euclidean geometry to a bunch of 10 year olds.

And to make this anecdote even better, in a different class the same day she had a substitute teacher who didn't know my sister, but seeing the triangled up grapefruit, approached her and asked her "Are you a Schmieg?". Our family name has racked up quite some reputation…

@inthehands honestly, even independently of that, the consolidation of the media market is a huge problem in its own right here. Why is Disney owning late night talk shows? This kind of pressure is only possible because there are few large players that want to get even larger and stay in the government's good graces.

When you think about it, we really should call "Tesla autopilot" vibe driving

Calling linear regression "one layer neural network AI" going forward

Remember the time when the best and worst thing about computers was that they did exactly what you told them to?

Please do not vibe code reference implementations. FFS.

@carnage4life as a colleague said recently: "Maybe executives have access to different AI models than the rest of us"

I think I put that together wrong.

@hrefna I wonder if that is even theoretically possible with the current architecture. The AI merely creates the most likely sentence, which works to recreate facts because factually correct sentences are more common in its training data than factually incorrect ones, so the most likely response is often correct. At least, as long as minor changes to the sentence don't make it incorrect.

As such the AI never knows an answer, but also never not knows an answer. There is always a most likely sentence, and that sentence is usually not "I don't know", because we tend to not write about things we don't know about.

New blog post, about a fun obsession of mine, the reason why we use elliptic curves and not any other groups for Diffie-Hellman.

https://keymaterial.net/2025/05/23/there-is-no-diffie-hellman-but-elliptic-curve-diffie-hellman/

Ah, yes, the two genders

Damnit, I'm sitting next to a cryptocurrency person in this plane and am not wearing my "crypto means cryptozoology" shirt.

And all of the sudden, we have solved supply chain security.

No CVE, no vulnerabilities!

@mattblaze even just minor UX improvements like Gmail marking external email addresses yellow are a good way of avoiding this type of issue. But that fundamentally relies on there being some concept of "external", i.e. they can only work if there is some defined organization as a context. Which for Signal simply does not make sense. It is an excellent tool, but it is the wrong tool for the job.

@mattblaze @ct_bergstrom I legit thought it was some misunderstanding of quantum computing at first, but then the title says "AI computing" and I got nothing.

(And to be clear, quantum computing also does not work like that)