So is the issue that he used manipulative language, or that he used AI to generate manipulative language? Something tells me that the actual issue is the type of language used, not where it originated from.

The openwrt docs reference this post: forum.openwrt.org/t/so-you-have-500mbps-1gbps-fiber-and-need-a-router-read-this-first/90305

They talk about a router from 2009 not being able to handle CPU-only SQM on a gigabit internet connection.

If you have a gigabit internet connection you don’t need SQM.

The algorithm is super easy:

A stream is defined as something that shares the same Source IP, Target IP, Source port, Target port and protocol, so e.g. “TCP 192.168.0.2:45231 -> 80.0.0.1:443”. This 5-tuple is guaranteed to be unique at a certain time.

The router counts how many packets each stream had in the last second.

If the bandwith is at its limit and two streams want to send a packet, prioritize the one that had fewer packets in the last second.

This covers all of your bases:

• A huge download will have tons of packages and it will not have trouble if it gets its speed throttled by a few percent. Because it gets most of the slots it gets lowest priority.
• A video call will have medium traffic. It gets priority over the download, because it has less packets, thus it will not lag, but because it doesn’t need more traffic, it will not interrupt the download.
• Scrolling Facebook requires quite low traffic. Most of the content is text or pictures, with a rare, often short low-quality video with quite low traffic requirements. It will get priority over the video call, but since the traffic volume is so tiny it will not harm the video call, unless you have a severely slow connection (<50MBit).
• Telemetry has incredibly low traffic requirements. We are talking about a few kilobytes per hour. These will have high priority, but since they have such an incredibly tiny traffic footprint, that won’t do anything at all, unless you are on a 56k modem, but if you are you won’t be downloading, video calling or scrolling Facebook anyway.

The cool thing is, this algorithm needs extremely little space in RAM and is extremely easy on the CPU too. To store a stream’s data you need 13 bytes for the metadata plus 3 bytes for the package counter, netting 16 bytes per stream. That’s 64 concurrent streams per Kilobyte of used RAM. A megabyte fits 64k streams. You can implement that algorithm on an original Gameboy and it wouldn’t sweat.

Also, remember that this algorithm only looks at the last second or so of data, so it can drop all idle sessions within just one second. 64k concurrent streams within one second is enough for professional-grade office routers.

Source: I actually spent the first 7 years of my carreer at a company that makes routers and switches and developed the software for them. 128MB RAM was what our mid-range models had, and every single model we had had traffic priorisation.

In fact, our network-provider-grade routers and switches had dedicated network metadata RAM, and they were usually limited to 4-16MB.

The only reason your router even has 128MB RAM or more is so that it can display a fancy configuration web page with images.

Actually, no. That’s a very simple basic functionality. A router needs to identify the streams (identified by the 5-tuple of source and destination IP and port and the protocol) to work at all. It also needs to prioritise traffic to work at all.

Combining both features is trivial even on 128mb RAM, and it’s implemented in most routers.

I wish all these European Linux projects would pool resources and create one good solution instead of each little country or even city DIYing their own solution.

Yeah, that’s what I really don’t get.

Why would any serious company think it’s a great idea to outsource all your intelligence work to a handfull of US companies, making yourself wholely dependant on their goodwill and the goodwill of the US government?

I would so love if this causes an usage spike for Firefox, so that all websites and webapps start testing for Firefox compatibility again…

That’s where traffic priorisation comes in. If your router is setup in a somewhat reasonable way it gives priority to smaller data streams to avoid just that.

What is not true?

Do the math yourself, it’s only grade school level:

A download takes 100 minutes at 10Mbit. How long does it take at 20Mbit and how many minutes are saved?

The same download takes 2 minutes at 500Mbit. How long does it take at 1000Mbit and how many minutes are saved?

This calculation doesn’t even take into consideration that most servers don’t allow for gigabit downloads and that most wifi connections also don’t allow for gigabit.

A relevant part is that double the speed doesn’t save the same time.

So e.g. if a download takes 100 minutes on 10MBit, you save 50 minutes by doubling the speed.

The same download would take 10 minutes on 100MBit, doubling the bandwidth would only save 5 minutes.

And on 500MBit it’s two minutes, so doubling the bandwidth only saves one minute.

We are deep into diminishing returns here.

€58 vs €49 means an extra €108 per year. That’s quite a sum.

In my case going from the 150Mbit/s I have to gigabit would cost me €35 extra per month, €420 a year, yeah, that’s not worth it to speed up some background downloads.

I’m on Fedora, updates are frequent as well, but since they download in the background I hardly care about the speed. I see there’s a new update, so I start the download in the beginning of the day. It finishes within half an hour or an hour or so, while I continue doing my stuff, and in the evening when I’m done I run the actual update if it requires a reboot.

But how often do you do that? And do you need all 10 games instantly available on your PC?

I recently setup a new laptop on Fedora on a 150MBit connection. That was around 10min for downloading Fedora, 20min for installing it, another 20min or so for setting up Steam and Heroic launcher (for GOG, Epic and Amazon Games). I started the first game download on Steam while I was setting up Heroic and it was done downloading before I was done with Heroic.

Since I can only play one game at the time, I could already start playing and let the rest of my library download in the background.

A faster internet connection would have just shaved off a few minutes from the initial 10min downloading time for Fedora, but I don’t know how fast the server even lets me download the image.

I mean, if you pay €20 for gigabit, sure, why not. The only network provider who serves gigabit at my home wants €65 per month for it compared to the €30 I pay right now. That’s €420 per year extra, and there’s really no point in paying that to save a few minutes every few months or so.

Tbh, I don’t think the post is bad at all. If you have special high-bandwidth use cases that require massive speed you know that already and then the article isn’t for you.

If you say you “didn’t have fiber” I’m guessing you were on 50MBit VDSL? Then, of course, switching to gigabit makes sense.

In the blog post the author explicitly doesn’t say “Go get VDSL”, but they compare Gigabit with 500MBit. That’s not nearly that much of a difference, you’ll still be able to play a new game minutes after you bought it, but just twice as many minutes. If that at all, because if you have wifi in your home, it will likely limit your bandwidth to less than 500MBit in real use anyway.

The main point of the post is to show whether a regular user really benefits from Gigabit, and no, they don’t. Their netflix stream will not improve when going from a few hundred MBit to a Gigabit. Neither will most of their experience.

If you are lucky enough to live in a place where Gigabit costs nothing, sure, might as well. The only provider who serves Gigabit to my home wants €65 per month for that, €780 per year. That’s a lot of money for something that maybe saves me a few minutes once or twice a month.

Or Firefox.

LLMs don’t need to become AGI. They only need to convince people in charge that they are.

Of course. That’s why we trust it. My point was that a tool doesn’t need to have “understanding” (whatever that exactly means) to be useful, it needs consistency.

The compiler doesn’t understand. It converts, following strict rules. There’s no “thinking” or “understanding” in the process. It’s a pure one-to-one conversion. That makes it consistent. If I run the same code twice through the same compiler I will get the same output. It’s an output that I can trust and thus using a compiler is helpful.

If an AI was able to consistently, dependably, provably convert my natural text input into a correct program, then I’d also be able to trust it.

The problem is that this is technically impossible, because the whole point of using natural text input is that I can use vague, unclear and incomplete language to describe what I want. With an ambiguous input there cannot be an unambiguous output.

A compiler also doesn’t “understand” like a human does, but it’s so accurate and reliable, that collectively we decided that a programmer doesn’t need to understand the output of the compiler any more. We can trust it, because it works perfectly reliable.

So the question isn’t so much whether AI “understands” the way a human understands things, but instead whether it will become reliable enough to be trusted. And I don’t see any evidence towards that yet.

Yeah, he likely will. The whole department management is in a gold rush here.

But, you know, my boss can now finally code himself. He could never get his head around coding, but now he finally can code. Who knows, maybe he won’t need us expensive software developers anymore when AI gets better, then he can do it all himself.

That is correct, but it’s also kinda beside the point. A compiler doesn’t “understand” anything either, it’s just a program, but it’s still very useful. So useful in fact that most programmers nowadays trust it unconditionally to the point where we say “a programmer doesn’t need to understand the code the compiler outputs”.

But AI is just really not trustworthy, and the question is whether they will ever be trustworthy enough that we can say “The programmer doesn’t need to understand the code the AI outputs”.

And yet, that’s all my boss wants of me. “Does a dev really need to understand the code if the AI understands it?”

That’s frustrating.