Accurate. At work I’m finally getting around to paying down tech debt from a temporary hack that was supposed to only be in the code for a month or so, but now it’s been in our production code for around five years.

It would have been maybe a week or so to do it the right way initially, but we had hard deadlines and were pressed for time and just couldn’t get around to doing it. Now it’s over a month of work for one senior engineer (me) to clean up, because there’s so many interdependencies between the hacky code and other parts of the system that depend on side effects of the hacky code, or the order of updates performed in the hacky code. What a mess.

At least AI can help a bit. Even the newest models aren’t so great at system architecture in brownfield projects though (projects building into existing systems).

Not sure if this counts, but as a software engineer, pretty much all major software you use is hacks on top of hacks on top of something someone built as a ‘temporary’ solution or as a proof-of-concept that’s still in use years later. It’s amazing that anything works properly at all. Stable software is basically just software with 10+ years of bug fixes layered on top of it.

Part of the reason that Windows backwards compatibility is so good (plenty of Windows 95/98 apps still work fine on Windows 11!) is because Windows automatically applies compatibility fixes, including simulating bugs that were in older versions of Windows, because old apps could be relying on those bugs.

My wife gets very frustrated when she encounters bugs, whereas I’m usually like “well, it’s not nearly as bad as it could be” lol

Yeah that’d be my guess too. It’s one charger that has two different cables, rather than two separate chargers that can operate independently.

That’d be considered cheap in California.

getting the app, creating an account

They don’t support Plug and Charge?

Not all those companies have their own data centers, but plenty of them would rent space in data centers. You can rent anything from individual servers owned by the data center (“dedicated servers”, like what you can get from Hetzner), to space in a shared rack, to an entire 42U rack dedicated to you, to an entire cage (or several cages).

(a cage is just a part of the data center that’s sectioned off in a locked metal mesh cage so only one customer can access it, while still allowing plenty of airflow to and from the servers)

I missed all this drama. Is there still a way to see the list? I’m very curious about it.

AI loves inserting comments like this.

Either that or 20-line comments documenting the bugs in the old code, what this code does differently, and the exact implementation.

Comments describing the exact implementation become stale very quickly. Comments shouldn’t explain what the code does (eg “add two to rowCount”); they should explain why it does it (eg “include header and footer row in count too”). Extra details are what the commit message is for, and history is what source control is for.

I call it out in a lot of my code reviews (“code looks good but clean up the AI slop comments before landing”) but a lot of other developers don’t care as much about code and documentation quality.

Please may I have a crumb of context? What is Tesseract? I only know the OCR engine.

The post is about data centers though.

No serious business is hosting an internet accessible serve in their office.

There’s some dots on the map in places where I know for sure that there’s no data center. I think people have submitted addresses of some tech company headquarters for example.

Most new builds by the larger companies are pretty big though, even if they’re not dedicated to AI.

The site explicitly mentions “AI data centers”, but how do you differentiate AI data centers from… non-AI data centers I guess?

I understand the anti-AI sentiment and environmental issues, etc. but I also think a lot of non-technical people don’t realise that literally everything they do online uses a data center.

It’s not way more than they thought because it’s within the limits they set for the subscriptions. They wouldn’t have set the limits that high if they didn’t think people would use it.

Yeah I think I was thinking of DPAPI. Updating my comment!

There’s other approaches to automatically unlocking. For example, you can combine Clevis with Tang to use Network Bound Disk Encryption, which lets you automatically decrypt only when on a trusted network. (the password is stored on a different system on the same network)

You can have multiple different LUKS keys, and any of them can be used to decrypt the disk. This lets you automatically unlock via a network server when your system is at home, while still prompting for an unlock password when not at home.

At my workplace, our laptops have two LUKS keys - one for me and one for IT / tech support. It lets IT boot the system (eg if I send it in for repair) without having to know my LUKS password.

Bitlocker stores the key required to boot the system in the TPM.

You can do that on Linux too if you want to, using Clevis or systemd-cryptenroll.

As far as I know, Bitlocker uses two keys to boot: There’s a key in the TPM, that’s used to decrypt a key stored on the drive, that’s used to decrypt the data.

I think some data (maybe just user-specific data?) is encrypted with a user-specific key too, that isn’t unlocked until you log in. Edit: I think I was thinking of DPAPI.

I’m not 100% sure about the specifics of its design.

I saw this in person when I visited Czechia last year. So many people pat/rub the dog.

I don’t understand how apps like Polymarket are legal when they’re essentially unregulated gambling that’s rigged against regular users due to the amount of insider trading that happens on the platform.

Being able to bet on wildfires doesn’t make any sense, since it gives people a financial incentive to start fires. 🤔

cost of AI going up

The highly subsidized Anthropic and OpenAI subscriptions will go away, but open weight models are getting much better over time - GLM-5.2 and Kimi K3 are both very good.

The most expensive part by far is training the model. That’s why OpenAI and Anthropic are losing so much money (well, that and the subscriptions).

With an open weight model, someone has already trained it, and you just have to cover the cost of inference, making it a lot cheaper. Any company (or individual!) with powerful enough equipment can host the model, which means there’s competition in terms of price, compared to something like Claude Opus where the only four hosts (Anthropic, Amazon Bedrock, Google Vertex AI, and Azure AI Foundry) use Anthropic’s pricing.