Beyond prompts or .cursorrules, I explored whether code structure itself can serve as context an AI can understand. Most models’ context windows are small and hinder free-form exploratory coding; if structure can act as additional context, it could be a big win.
Test:
Two codebases implement the same behavior but differ in structure.
Code A: goal-driven, structure-agnostic.
Code B: lightly refactored with some structural improvements.
Both were 100% AI-written. I fed the same model the two bases separately and issued the same modification request.
Results:
On Code A, the AI grafted features onto a structureless base; subsequent edits easily broke existing behavior.
On Code B, the AI respected the existing composition/structure/pattern; follow-up edits were far less likely to cause regressions.
Conclusion:
The AI does leverage existing code structure; on a good structure, it tends to produce sturdier, higher-quality code.
Personal note:
This is an interesting direction for AI-assisted development. The write-up is short but setup-heavy; to really grasp the test, hands-on is needed. A tighter focus on the test, results, and implications would make it stronger.
I’ve heard that in North Korea it is difficult for ordinary people to learn or own a computer. It is assumed that a small number of elite operatives are selected and trained to carry out such tasks, and it is somewhat surprising that they possess the latest technology and conduct hacking.
If anything the hackers in north korea are probably world class if the government is getting their students into focused training programs early in their schooling. Western nations have nothing equivalent due to schooling being generalist and undergrad and grad school not really introducing you to the sort of work you'd actually do on the job as a hacker. 22 year old western hacker for a 3 letter agency is going to have maybe a 6 month softball tangentially related internship of experience under their belt while the north korean might have years and years by that point.
> 22 year old western hacker for a 3 letter agency is going to have maybe a 6 month softball tangentially related internship of experience under their belt while the north korean might have years and years by that point.
I was with you right up until this bit
The agencies concerned tend to recruit people that have demonstrated ability in that field, and they've usually got it with "self-directed" training :)
Lurking forums and irc is probably a terrible way to train a hacker than a dedicated program that introduces you to the tools you'd be using on the job right away. Even today people don't even like hiring self taught engineers anymore like 20 years ago, when there are many more people today who have gone through legitimate education programs.
The one hacker I met in my life went to West Point and had no experience they didn't gain from being placed in their program after graduating with decent test scores.
State sponsored thieves are not a talent pool that anyone wants in a trusted position.
The fact is there were only around 40 unique hacks ever invented, and people simply adapt these into new zero day exploits. Notably, this is now mostly a fully automated process.
If people want in, they will get in eventually. =3
No need for insults, I found it fun. ROTs are easy to detect because they usually still have word-length chunks, and common repeating symbols. In this case the '6's ('e's). This is something a language oriented AI is going to be very good at detecting. It's great demo of why hashing is so important.
If you don't see repeating symbols, it could be a running key, like a Vigenèr cipher.
It was a simple way to highlight impulsive behavior common in modern users, and the trivial encoding function should be obvious to those who are minimally empathetic. Ask the LLM handler if being lied to makes people feel worse than getting robbed... then consider if you would hire such individuals.
If you are ever unsure of someones motives, than politely ask for context. Have a wonderful day =3
What is the impulsive behavior? Do you have a zero day in some ROT-47 decoder? Or perhaps a zero day in the file command in case a user creates a file containing the string and runs the command on it? Or is the string both a valid ROT-47 string and a valid executable on some platform?
> If you are ever unsure of someones motives, than politely ask for context.
It is complicated, but Moral Development theory does cover the phenomena of why some won't understand until they personally grow through the stages of development.
Reading up on it made me realize that a certain well known orange person is really on Stage 2 of moral development. That explains a lot.
But also gives hope. I mean, it’s rare that adults fail to advance from pre-conventional phases, so it must be super rare to have such a confluence of factors that puts someone like that in the given job.
Spies do tough work for not that much pay. (Certainly less than they can earn in the private sector.)
They’re starting from a position of duty. Given the stakes the questions they’re tasked with operate at, I’d guess they tend to be in the postconventional regime more than most people.
> it's not because someone wants them there. It's because they can demand the position they want
Zero evidence of this. And if they can demand that position from one, they can demand favors from others. I would count a background in espionage to be a net positive in a hiring process, provided dismissal was on good terms.
I always understood that these hacks are one of the main ways for North Korea to actually earn money in other currencies, as they’ve been barred from trading with pretty much the entire world.
North Korean teams tend to perform very well in coding contests, so it’s a safe bet that North Korea is quite good at nurturing a small slice of elite computing talent.
"Major North Korean universities, such as the Kim Il-sung University and the Pyongyang University of Foreign Studies, send a few dozen exchange students to Peking University and other top-ranked Chinese universities each year."[1][2]
"North Korean hackers are sent vocationally to Shenyang, China for special training. They are trained to deploy malware of all types onto computers, computer networks, and servers."[3][4]
The brightest students of most nations are often sent abroad to enrich their countries with knowledge from the great universities. NK is almost unique in its inability to do this at non-Chinese great universities, so that is the only viable route.
I would doubt it because North Korea has extremely strict controls on who exits the country for any reason, but especially for education. I know this has happened before (for example Kim Jung Un studied in Switzerland under a false name when he was a kid), but it's extremely rare, and runs contrary to the North Korean philosphy of Juche, self-reliance.
> somewhat surprising that they possess the latest technology and conduct hacking.
Why does this surprise you? As you said, selecting capable people is not a problem. And then these capable people get the best possible motivation. I would say it is expected to get qualified hackers in such conditions, who are proficient in all latest technologies.
...which explains the link to China. NK natives probably do not typically have access to computers or the open internet, but the children of certain elites are educated in China. There may even be a collaborative effort between the two states.
From a Bill Gates documentary, I saw research with partner companies aimed at improving nuclear power generation mechanisms to reduce waste and increase efficiency. Bill Gates’ endeavors always seem positive and fascinating.
Whoa, this is a bit scary. As mentioned earlier, it should basically be used in a way where other energy sources are tapped first, and only the shortfall is covered.
I expected there would be constraints, but the chosen range is quite intriguing. The PostgreSQL spec says the 4-byte date type spans 4713 BC to 5,874,897 AD. It gives much more headroom for future dates—did they assume preserving data before 4713 BC is unlikely?
That range of dates seems to correspond to (1UL << 31) days so I suspect they're using only 31 bits so I wonder why they didn't make it signed and extend it to 5,884,322 BC.
Beyond prompts or .cursorrules, I explored whether code structure itself can serve as context an AI can understand. Most models’ context windows are small and hinder free-form exploratory coding; if structure can act as additional context, it could be a big win.
Test:
Two codebases implement the same behavior but differ in structure.
Code A: goal-driven, structure-agnostic.
Code B: lightly refactored with some structural improvements.
Both were 100% AI-written. I fed the same model the two bases separately and issued the same modification request.
Results:
On Code A, the AI grafted features onto a structureless base; subsequent edits easily broke existing behavior.
On Code B, the AI respected the existing composition/structure/pattern; follow-up edits were far less likely to cause regressions.
Conclusion:
The AI does leverage existing code structure; on a good structure, it tends to produce sturdier, higher-quality code.
Personal note:
This is an interesting direction for AI-assisted development. The write-up is short but setup-heavy; to really grasp the test, hands-on is needed. A tighter focus on the test, results, and implications would make it stronger.