Well but for normal distributions there are two big advantages compared to commercial updates.

1. You can turn it off at any time
2. It's designed to make it _very_ hard to specifically provide special update for a special target, since the process is transparent, and mirrors are something that's encouraged.

Particularly part 2 is relevant in an age where an US-president could just force a company to send a special version of program X to person Y.

... that if you get software without source code and it has an auto-update feature you cannot control... maybe you are not the person who owns those computers.

The costs of various operating systems is not a monetary value you can just compare. It's a variety of things you need to consider.

For example if you get Windows, you need to factor in the specialists who can manage those systems, those are rare, and most people claiming they can, don't really know what they are doing. However should you, for whatever reason have such people around, perhaps for historical reasons, that may be a non-issue.
For MacOSX one big issue is that you are tied to extremely budget hardware. It may look cheap at first glance, but build quality typically isn't very good plus they are hard to repair and there are hidden surcharges. Again that may be irrelevant as you can just out-source that.

The costs of switching however have reduced significantly since the Internet came along which eliminated the need for many applications in the past. Most systems are POSIX now so software is now mostly portable. Heterogeneous systems can make switching even easier as it keeps you from relying on systems that make switching more expensive.

Yeah, but that's actually not our problem here. If everyone would just do that, we'd have no memory crisis. It's the "hyperscalers" who are gobbling up all that memory for their _huge_ models. That's the problem, not someone running something locally.

And companies are already switching to local models instead of paying those hyperscalers more money. That means those hyperscalers will loose more and more of their income. That will cause the bubble to burst.

Again, that doesn't contradict the argument that we are in a bubble. Actual thinking is something much different to LLMs and even if LLMs are used as part of such a system, they will be much smaller. Frontier-Level LLMs are _already_ to expensive to ever be economical, and bubbles are not about technology, but economics.

There are probably 2 reasons to announce that:

a) generate more hype (our system is so good, it's scary)

b) slow down the process of setting money on fire, as you pause spending it.

Do you have any kind of evidence to back up your claim? Sure there will be AI in areas where it's useful, but the current hype of generating ever larger and larger "frontier models" will end very soon, that is virtually certain.

Such models aren't much more useful than small ones, and once someone discovers a trick to make such small models even a bit better (e.g. run-time re-enforcement learning for the task you want to solve) there will no longer be the need for such huge data centers or companies like Anthropic or OpenAI.

We are already at a point where smaller models outperform large ones, as there is less enshittification in them.

... hasn't burst by 2027.

Considering that companies, the sources of income OpenAI and Anthropic hope to have, are now reacting to recent price hikes... by looking into cheaper "non-frontier" models... perhaps even on their own hardware, it may just be weeks or months until the bubble bursts.

Why spend 10x the money, to get an advantage you can neither measure nor feel?

It seems that with AI, the more effort you put into it, the less profit you get.

They are usually based on Javascript running off of ad-servers. Since NoScript and other technologies limiting the damage caused by Javscript and advertisements are highly correlated with certain platforms, the results will be skewed severely towards platforms where those are less common.

Yeah, but what does it matter? It'll be pirated the first day it'll be released anyhow. There is no negative correlation between pirated copies and the money a movie makes.

Isn't ironic that a website concerned about privacy uses Cloudflare, one of the biggest players of the surveillance technological complex.

Typically movies are not secret information. Often movies even get published. In fact where I live, there's a big house with projection screens showing movies to anyone who pays a nominal fee! In fact even the movie studios themselves will encourage people to watch those movies, often even via advertisements! Movies are, usually, not secrets.

There is no sane reason to encrypt that.

In repressive countries with video cameras like the GDR, the trick was hairspray. It's usually water soluble and can be removed without lasting damage. This makes it hard to be considered property damage... since nothing gets damaged. Yet it effectively disables the camera.

It's also a very normal item to own and can be bought easily.

To be honest, there are very few scenarios where 50k lines of code have anything to do with productivity. Complexity in software is spent, not produced. If you have to much complexity in your project, it will collapse... and many projects have already collapsed under their own complexity.

However LLMs do not necessarily cause complexity. I have been experimenting with it regarding telefacsimile. That's when you have a sort of copier that works over the telephone line. Thrilling stuff, but few people care about it. If you want to do it on a computer, you usually need to use a library called "Span DSP". It's something like 50k lines and mostly unmaintained. In order to actually send and receive faxes, you need to embed it into a software PBX like Asterisk, FreeSwitch or Yate. Those are an additional 1 Million lines of code and feature (near-)Turing complete languages for configuration. Even with all of that, you only have an extremely basic implementation of Group 3 telefacsimile.
However with a bit of effort you can generate slopware with far fewer lines which are way more complete implementations:
https://github.com/Casandro/no...

Yes, but to play devil's advocate here for a moment, is it really?

Isn't much of programming more or less just writing instructions for your computer. Looking up instruction codes in the hand-book isn't particularly creative, that's why people made assemblers and even compilers where you, essentially just type in your idea in a human-like language like COBOL or C and the instructions come out of the compiler.

Again, yes, it can be fun, but so can prompting also be. It can free you from things like having to look up the "API of the day" and concentrate on things like software design. It can do mundane tasks like refactoring your code, etc.

The art argument is IMHO one of the weaker ones. Coding usually isn't art, it's a mechanical encoding of ideas into code. If it wasn't for software companies, there probably wouldn't even be copyright for computer code.