swsieber

As someone whose commercial software needs to ingest PDFs: I stay away from things like lopdf because of how they fail to handle broken pdfs.

The first project you linked explicitly calls out broken pdf handling.

> but they're not even in the same universe as what the Stasi was doing

At what point? It's easy to say when we mention the Stasi or the Nazis that we're comparing it to the end game they had going, but the transformation was a bit of process.

> Textual command filtering is entirely useless. It is fit for no purpose. Nobody with any instinct or experience in security would even bother to implement this filter because it achieves nothing except a false sense of security.

Not if you only use an allow list. E.g. only allow things matching a particular prefix to run. Also, isn't this what every agent does? And by every I mean Codex because I haven't used the others.

> Even if your average developer is a measly 30% more efficient, that might create 20% less demand for developers

I mean it might. But I wouldn't rule out Jevon's paradox where the increased efficiency increases demand.

Build more roads, congestion gets worse. Make developers more efficient, demand for developers increases. I wouldn't be surprised if demand for bespoke software goes up.

And I would say in its case, there was great synergy between not doing anything and appearing uninvited. Two negatives combining to produce an overwhelmingly greater negative.

> compelled

Doesn't that mean they can ask that question with an option for "rather not disclose"?

Ah, this is fun to see.

About six months ago I had an idea for a short story in which an LLM takes over the world and is decidedly bad. The solution was going to be for everybody to write positive stories in which the LLM is good and relinquishes control, which then made it's way into the LLM's training data and it backed off. I never got around to it.

AI Coding "All the things" (tm). More agent engineering on the bigger stuff and more vibe coding on the smaller things.

A print farm manager for bambu printers in lan mode. I lay down the base types and schema structure and a few other bits here and there.

Using AI to preprocess some amazon transactions from both personal account and business accounts as I untangle them since I started a side business with my spouse a few months ago (involving 3d printing).

Starting on a yoga workout generator and food/fitness/weight tracker.

> And yet, people are being more productive (actually productive) with AI. Release schedules are increasing, bugs are getting fixed faster, security issues identified and patched sooner, so on and so forth.

I didn't see anything in parent chain that implied this. Nor did I see it "characterized as a magic eraser"; I saw it framed as something that impedes learning, and that was tied back to constant simulation.

> Or at least, aware that this argument continues to be made with tenuous evidence and anecdotes

The arguments I read and the argument you seem to be replying to seem to be different things.

I read it more as response to your argument that it was a lack of curiosity due to over stimulation, which they responded to by citing an example of a time when they were curious while stimulated and chucked something at a vibe-coding agent to satisfy that curiosity.

> relatively little friction if you decide to move to a state that better reflects them.

Yes, moving is possible, and easier than switching countries. But the biggest strength of the U.S. is not that people can move, but that the blast radius is contained. The strength is being able to think later "I'm glad we didn't try that everywhere at once." The strength is being able to experiment, not necessarily have states aligned to with all there residents needs within a certain threshold.

I appreciate the write up.

It comes off very much as a summary of days of vibe-coding from the AI itself, but it's still nice to see the condensed decisions, approach and frankly the whole concept of embedding a wasm runtime.

Around, not over. Trees are a well studied thing where can you pick different species for different characteristics, like height, and growth speed.

I wouldn't call something a non-story just because the ultimate end-goal was mitigated. The fact that it was attempted is a story, especially when it's a meta commentary on story about trying the same thing _officially_.

To be honest, we should probably go into injection molding. We do have a pretty long tube in there that might be a bit problematic for injection molding from what I've read (1 degree draft means the tube size inside changes by 1/8 of an inch), and the orientation of the magnets is pretty important, and they can repel each other pretty hard to. So part of it is the unknowns.

The other part is the upfront cost. I bet we'll get to injection molding in the next few months as revenue allows, and we're going to start exploring it this month I think. We'd like to keep things local, though we know we'll still have to contend with knock offs sooner than later.