The Communication Cube


Probably a common thing to be suppressed, but did you hear anything about the impact of Typhoon Bavi on China?
- Shenyang (Liaoning Province): The provincial capital was hit by record-breaking rainfall.
- Chengde & Kuancheng County (Hebei Province): Continuous torrential rain caused water levels on neighborhood roads to rise more than 2 meters.
- Beijing: The capital's suburban districts (specifically Miyun and Huairou) suffered flash floods that destroyed roads, cut power to dozens of villages.
- Taizhou (Zhejiang Province): Typhoon Bavi hit this city first, impacting coastal villages like Dongsha.
- Wenzhou & Yueqing (Zhejiang Province): This area got Typhoon Bavi's second landfall.
- Hangzhou & Shanghai: Typhoon Bavi caused cancellations of hundreds of flights and thousands of train trips.
- Xianfeng (Hubei Province) & Guizhou Province: This region got over 350mm of rain fall in just 12 hours; surpassing an entire month's average.

I could imagine this has a similar impact as the closing of the Hormuz strait.

1784454990639.png
Pictures like this, if happened in Europe, would be front page news.
 
So~ have you guys heard about e.g. google paying companies for buying up rare books to feed AI's voracious maw?
The books, which are often extremely rare, get bought and sent to their death by de-spining, scanning and (for drama's sake) burning.
Also only books released before the advent of AI-"enhanced" books get scanned and destroyed as to not muddy the waters I guess.
That's knowledge eternally lost to datacenters. Surely no harm can come of this?
 
So~ have you guys heard about e.g. google paying companies for buying up rare books to feed AI's voracious maw?
The books, which are often extremely rare, get bought and sent to their death by de-spining, scanning and (for drama's sake) burning.
Also only books released before the advent of AI-"enhanced" books get scanned and destroyed as to not muddy the waters I guess.
That's knowledge eternally lost to datacenters. Surely no harm can come of this?
I heard something similar, only that they just cut them with a guillotine instead of unbinding them, and after scanning they made paper pulp from them instead of burning them-
Recycling is better than burning, but the main point stays the same.
The question is which books are those. Many old books end up in the rubbish. Some are important and need to be preserved, others are abundant and losing one copy isn't anything to write home about.
But I very much doubt any of those companies care to tell them apart.
There're also copyright issues for some not so old books, but it's just business as usual, move fast and break things laws.

If at least they cared for keeping digital images and transcripts scanned and give them away to libraries for preservation, instead of just bleaching them in weights' soup...

I wonder how will it turn up in the output too. On the one had, it can improve composition and language skills, but it may bias it even more towards all kind of politically incorrect views and downright myths.
But, well LLms are not sold for the wisful thinking they do, tey are sold on the premise the customer can quit thinking in general, and specially about the quality of the results.
 
I heard something similar, only that they just cut them with a guillotine instead of unbinding them, and after scanning they made paper pulp from them instead of burning them-
Recycling is better than burning, but the main point stays the same.
The question is which books are those. Many old books end up in the rubbish. Some are important and need to be preserved, others are abundant and losing one copy isn't anything to write home about.
But I very much doubt any of those companies care to tell them apart.
There're also copyright issues for some not so old books, but it's just business as usual, move fast and break things laws.

If at least they cared for keeping digital images and transcripts scanned and give them away to libraries for preservation, instead of just bleaching them in weights' soup...

I wonder how will it turn up in the output too. On the one had, it can improve composition and language skills, but it may bias it even more towards all kind of politically incorrect views and downright myths.
But, well LLms are not sold for the wisful thinking they do, tey are sold on the premise the customer can quit thinking in general, and specially about the quality of the results.
It's a problem, but I wonder how big it will actually be. It might be like the idea of self driving cars; where there are some better drivers, but overall driving is improved because most drivers aren't that good. And I guess smart people know how to use LLM's effectively but maybe shouldn't, while people who should use them cannot or don't (use them properly). So, even without doing it on purpose the gap between very-productive and less-productive people might grow very big due to this. And giving large quantities of people a nudge to a preferred political side is already happening without AI, but that will be even easier or happing unplanned due to an unknown bias in the training material.
Destroying books, especially rare ones, feels bad but I assume they are either not doing this (with rare books) or are preserving digital copies. Assuming the people involved aren't stupid and they aren't evil.
 
I guess smart people know how to use LLM's effectively but maybe shouldn't
My employer provides a ChatGPT account, which I make use of in my software development. I don't want it rummaging in my filesystem, so I only use it through chatgpt.com. And there only for two weeks now or so I get to use it agenticly and GPT 5.6 Sol.

After some back and forth getting the specs right, it worked 32 Min and then gave me a complete new subsystem for the CMake-based build and packaging system I'm currently developing. 4.6+k lines of code and, aside two nitpicks I had, it was good.

I find that impressive, and worrying, because now even less competent people can develop rather complex systems that work. But works, doesn't mean reliable or efficient or longterm-maintainable.

Now I had it give me another subsystem; 1.8+k LOC; technically fine, too. But here I have high no-cycles-to-waste-at-runtime requirements that it "knew" about. Didn't seem to care too much. With this one, I'll have to clean-up after it quite some.
 
My employer provides a ChatGPT account, which I make use of in my software development. I don't want it rummaging in my filesystem, so I only use it through chatgpt.com. And there only for two weeks now or so I get to use it agenticly and GPT 5.6 Sol.

After some back and forth getting the specs right, it worked 32 Min and then gave me a complete new subsystem for the CMake-based build and packaging system I'm currently developing. 4.6+k lines of code and, aside two nitpicks I had, it was good.

I find that impressive, and worrying, because now even less competent people can develop rather complex systems that work. But works, doesn't mean reliable or efficient or longterm-maintainable.

Now I had it give me another subsystem; 1.8+k LOC; technically fine, too. But here I have high no-cycles-to-waste-at-runtime requirements that it "knew" about. Didn't seem to care too much. With this one, I'll have to clean-up after it quite some.
Yeah I've been using Claude Code at work for a few months and while it can output great stuff it will also regularly make more or less stupid mistakes and can't be trusted to do everything on its own. The problem is that it always gives confident answers with code that seems fine, so many errors are hard to catch, making them even more problematic if deployed.

I'm also wary of the additional attack surface. As with all new technologies, security is an afterthought and I guess it won't be long before some organizations find way to poison the training data, do prompt injection on a massive scale or find another way to have it discreetly inject malware in the code or compromise the computer it's working on.

And that's before all other concerns, such as the environmental cost, the very likely upcoming cost increase, the risks of having junior developers stop learning, the risk of cognitive decline because of over-reliance, and so on. And in the end, I'm when I look at what I ship, I don't see any obvious productivity increase. It does seem like I'm doing more or working faster but actually tasks always take the same amount of overall time from start to finish.
 
Yeah I've been using Claude Code at work for a few months and while it can output great stuff it will also regularly make more or less stupid mistakes and can't be trusted to do everything on its own. The problem is that it always gives confident answers with code that seems fine, so many errors are hard to catch, making them even more problematic if deployed.

I'm also wary of the additional attack surface. As with all new technologies, security is an afterthought and I guess it won't be long before some organizations find way to poison the training data, do prompt injection on a massive scale or find another way to have it discreetly inject malware in the code or compromise the computer it's working on.

And that's before all other concerns, such as the environmental cost, the very likely upcoming cost increase, the risks of having junior developers stop learning, the risk of cognitive decline because of over-reliance, and so on. And in the end, I'm when I look at what I ship, I don't see any obvious productivity increase. It does seem like I'm doing more or working faster but actually tasks always take the same amount of overall time from start to finish.
In my experience an LLM can perform tasks well enough to place them on a spectrum: from okay-ish to perfect. Things like summarize some text it will probably do very well, while it can be a hit or miss when it involves numbers. And usually it is able to detect afterwards if a task was done properly if asked. So, creating a pipeline of small simple (always well executed) tasks is possible: including checking the results. In such a case a process can be automated, like you would probably always would have done, except now the process is much more flexible.

With malware in the generated code, this has always been an issue with compilers. A similar solution can be applied to mitigate the risk, for example to use 2 (or more) different AI models to generate the same code and they check each others code for vulnerabilities.

The risk of having less (competent) programmers is real. But I guess it's a similar problem to other tasks we automated in the past. Those problems are real but we pay the fewer experts more to compensate, the 'bad' programmers have to do other kind of work. The environmental impact is also real, but paradoxically those problems might also be solved by using AI. We'll find out how it will perform in practice. At least I'm pretty sure AI is here to stay, so life will change and we can only hope it will be for the best.
 
All these companies borrowing hundreds of billions to fund data centers worries me because it seems inconceivable that they will be able to repay the loans. Businesses can offset hefty ai costs by hiring fewer people, but I don't see how private individuals (the biggest market by far) will want to spend that much. How much are email summaries worth to you? I use ai mainly as a better google and suspect must people do too. But if the cost weren't negligible (free) i'd just use duck duck go. My fear is that big tech will start using extortion to get individuals to shell out for ai. Enshitification. I'm really scared actually.
 
Last edited:
Looks like the new Commodore will be present at Gamescom 2026, one floor below ED's DragonBox-shop stand:

All these companies borrowing hundreds of billions to fund data centers worries me because it seems inconceivable that they will be able to repay the loans. Businesses can offset hefty ai costs by hiring fewer people, but I don't see how private individuals (the biggest market by far) will want to spend that much. How much are email summaries worth to you? I use ai mainly as a better google and suspect must people do too. But if the cost weren't negligible (free) i'd just use duck duck go. My fear is that big tech will start using extortion to get individuals to shell out for ai. Enshitification. I'm really scared actually.
That will probably happen indeed. But we can always worry more about being killed in a upcoming war I guess, hopefully we'll be back on our game in 10 years or so after the bubble has burst.
 
Talking about Commodore, ED's selling the 8bitdo keyboard and numpad(& stand-alone calculator) in classic C64 color scheme:
 
Back
Top