The cool air in the upper atmosphere cleans the waves, though. Obviously that doesn’t work horizontally, everybody knows that.

But that’s exactly what Microsoft promised its shareholders and why they invested billions into it. They need world changing revenue to make it worthwhile.

Also just being liked by the interviewer. For my current job I had an interview of about 90min, and basically just had a rather one-sided chat with the two guys. They seemed to like me, just let me talk and the next day I had the contract draft in my email.

I certainly did not excel at anything during the interview.

What I find really fascinating here is that obviously openAI, Meta, etc. seem to be structurally incapable of actually innovating at this point.

I mean, reducing training costs by literally an order of magnitude just by writing better software is astonishing and shows how complacent the large corporations have gotten.

Deepseek showed that actually putting thought into the architecture achieves much more than just throwing more hardware at the problem.

This means a) there will be much less demand for hardware, since much more could be run locally on regular consumer devices. And b) the export restrictions don’t really work and instead force China to create actually better models.

That means, a lot of the investments into the thousands of AI companies are in jeopardy.

Yeah, it’s a budget Wurstbrot, but perfectly serviceable.

So, you fucked up and it’s postgres’ fault?

You could easily throw the components into an old tower case.

Getting the PSU to fit could be a bit tricky due to the rather short cables.

Workstations, like real workstations, are another beast and not what’s typically referred to as “office PCs”, those are indeed rather sff builds.

Again, optiplex sff 3060 as an example, it has two SATA ports, one x16 and one x1 (I think) PCIe, and looking at the PCB, apparently there’s a version with m.2 slots. Sure, not exactly server grade storage, but if you manage to find some version with m.2 slots or invest 10€ for a cheap SATA card, you can get enough storage attached.

GPU wise, absolutely no idea. My optiplex has a wx3100 that I got for cheap and its self reported power draw never goes under 5W, but since this machine is a desktop, it doesn’t run all day.

Sorry, but you’re either pulling those numbers out of your ass or haven’t kept up with the real world for 25 years.

The numbers I’ve posted above are measured using an external meter. I’m German, I have a vested interest in knowing how much power my devices pull.

And you don’t think, office PCs pay attention to power consumption, given they are intended to run 8h a day?

My optiplex sff runs at about 10-15W in idle, and it’s an i5 6500. The t variant in my elitedesk runs at 5W.

If you don’t need actually public DNS, something like Tailscale might be an option.

Yeah, but that’s the thing: it’s not fast. And that’s actually even more baffling.

All those rewrites cost money. All the shoddy software somewhere in the background of every corporation was written using these shitty libraries, frameworks, APIs. And that means cost.

I’m explicitly not talking about the newest crap Amazon or Google push, they operate under different principles, I’m talking about the data plumbing stuff. The eight billion Spring Boot apps, the PHP sites, the Python pipelines.

Writing and deploying a simple CRUD app, that just takes data from a request and saves it into a DB should be done in a few minutes. The actual “payload” is just input schema, DB schema, transformation rules, maybe auth. That’s it. However, if you want to do that in Spring Boot and K8s, it takes often hours or days to get everything right. Not because Spring Boot is hard, but because you forget a network policy in Helm or the dateformat between JSON/Java/DB is different or that library is outdated or there’s CVE somewhere or you have to look up that shitty mappedBy thingy from Hibernate again (that’s maybe just my problem, I can’t remember that crap) etc. etc. etc.

The basic problem behind is, that we have to take care of so much. Network, business logic, framework detail. You have hardly any layer or platform that is really solid.

Honestly, I would advocate the exact opposite.

Yes, programs became bloated and fragile, but the solution cannot be to return to the stone age, but be professionals for once.

Our entire industry is shit at actually engineering. There’s leaky abstractions everywhere, and that’s exactly why everything is so complicated and brittle. There’s no platform to build upon, only a scaffolding made of twigs, duct tape and three bananas for some reason. Every minor change in some minor library percolates through the entire stack.

You’re a simple developer, so am I. And we both probably wrote hundreds of apps that essentially do the same crud crap again and again and again. The same basic functionality gets implemented thousands of times, because we can’t get our shit together to build actually reusable components. Instead we rewrite the 12th iteration of “make stuff move in browser” and “make Java do business”.

We’re not engineers, we’re children with hot glue guns.

Because you don’t know what you’ll need that wrapper beforehand, that’s my entire point.

Unless you’re only doing trivial changes, the chances are very high that you won’t be able to design the class structure. Or, you end up essentially writing the code to be able to write the tests, which kind of defeats the purpose.

If you have to ask “can’t you just” the answer is almost always no.

And who actually writes tests like that?

I mean, do you think tests do the calculations again? You simply have well defined input and known, static output. That’s it.

Tests first is only good in theory.

Unit tests typically test rather fine grained, but coming up with the structure of the grain is 80% of the work. Often enough you end up with code that’s structured differently than initially thought, because it turns out that this one class needs to be wrapped, and this annotation doesn’t play nice with the other one when used on the same class, etc etc.

Especially then I’d test the shit out of everything? I’m getting paid for writing correct software.

For local development, it should be super quick. However, I’m currently building a small project where a device (or rather the library using it) can’t really be used with a debugger. So 500 print()s it is.