Evil hacking using AI vs. AI self-defense

10 years?

Value is a subjective action we as individuals do everytime. Research about the Subjective theory of value. I won't point to the Wikipedia page because it frankly sucks.

This simple page explains it in nice layman's terms:
The value of money is not subjective.

You're dancing around it. There's no actual reason to use it.
Same with the bug finding. It found many according to the AI companies. Show me 1 method description...
 
The value of money is not subjective.

You're dancing around it. There's no actual reason to use it.
Same with the bug finding. It found many according to the AI companies. Show me 1 method description...

Its relative scarcity makes it valuable to us working class people, but wealthy people don't have lots of money in the bank where it loses 22% of its value every 5 years now. They own assets and ask for loans instead.

And yes, the value of money is subjective according to most theories around it. Current money is not metallic, so it should be obvious. I later realized that there's also no intrinsic value in metals. Take aluminum. It was once worth more than gold because it was hard to refine.
 
Its relative scarcity makes it valuable to us working class people, but wealthy people don't have lots of money in the bank where it loses 22% of its value every 5 years now. They own assets and ask for loans instead.

And yes, the value of money is subjective according to most theories around it. Current money is not metallic, so it should be obvious. I later realized that there's also no intrinsic value in metals. Take aluminum. It was once worth more than gold because it was hard to refine.
You're not arguing anything.
According to you, what's the single best reason to use AI to develop software?
Only 1. It should be good, right?

This question-trick is the moonlanding-denier destruction. The single best evidence of fakery cannot be vague...
 
AI is a tool that can help you; in a sense that you can focus on an exploit itself. In this context (subject of a thread) it's very similar to what script kiddies do - e.g. load Kali Linux and use known venom to attack a target. AI can be used the same way. You have a chance to attack something even though you don't understand it.

Where AI is good at is glorified search. It's convenient to ask for "task_struct in FreeBSD" - it's wrong but AI understands enough to point you to struct proc. "Memory structures in XY", etc. You get immediate link to a code you can go through. It's up to you to make sense of what you see.

The recent exploit I did was on ppc. It's nice to get a quick overview of ABI, and general idea of registers. Without it I'd do it the old fashion way - google the ppc manual, look for the OS ABI and go from there. And you know, there's some skill in that too. But with AI you can speed this up.

Also AI is good as rubber duck - does wonders when one can talk about things.
 
You're not arguing anything.
According to you, what's the single best reason to use AI to develop software?
Only 1. It should be good, right?

This question-trick is the moonlanding-denier destruction. The single best evidence of fakery cannot be vague...
AI is very good at:
- Brainstorming.
- Writing skeleton code.
- Writing tests.
- All kinds of review.

It's also very good for single-use tools when used correctly. For more complex projects, it needs expert human guidance.
 
AI is very good at:
- Brainstorming.
- Writing skeleton code.
- Writing tests.
- All kinds of review.

It's also very good for single-use tools when used correctly. For more complex projects, it needs expert human guidance.
Sounds like stealing a project...
There are offline programs for all of that.
 
TLDR: the AI has it's niche when used under control in small portions.

I started to use AI and immediately found myself powerless: it does too much if allowed to. It can write code in seconds, something that'd take me minutes or even hours (we all love unit-tests, right?).

IMHO, the main problem with "AI design and implement the new feature but let me review it first" is that we loose intellectual control of the code. It looks great in isolation, but then we get back to it after a while just to question "what is going on here?" (let's not start the discussion of how to review AI generated PR with O(1000) lines, we all know that reviewing a PR with 100+ lines is ... not much enjoyable).

We need to go through the thinking process for something to land in our mind. That's why it's vital to experiment with tech: write the feature yourself, build, fail, build again. We learn by trying, by making mistakes. That's how humans worked for ... forever. And that is part of the memory process, of understanding the thing. AI takes that away from us.

Search is a great example IMHO: with great search results (some time back, before all the hype of the recent decade) we slowly started to memorize less over time (on average). Why would one remember something when it's right there, at the tips of the fingers, available 24/7? It works until Search engines stops to show what we need but it's too late.

That being said, AI has a great potential though. As long as we start small, slow, under control. A lot depends on the used AI: paid vs free, not all models are equal.

After all, only few people do science in this world even though way more are capable of doing it IMHO. But that leads to controlled experiments, transparency and confirmed results with thorough review.

This is not the case with AI. Nobody knows what sits behind that new model A released yesterday by X.

Here are a few example that I found useful though:

- research the topic: use it to research the unknown complex territory before dive into the documentation to do the homework. Just don't go too far with AI, pick up top-level concepts to avoid building false understanding due to hallucinations.
- consultant: ask for a second opinion especially at the level of prototyping.
- reviewer: review the work done before submit the commit to catch immediate bugs or lack of test coverage.

For the fans of Isaac Asimov, his robots always helped people, they never replaced humans. Isaac envisioned robots as good, not bad as portrayed otherwise. They unlock new horizons instead of taking everything under their own control. AI can play the same role.
 
it does too much if allowed to.
And send you in circles if you're not paying close attention. I mentioned that a couple of times here. I was too tired to try and figure something out and asked Claude to do it. Everything it gave me didn't really work right but I kept giving it feedback. After a while, I suddenly realized I was back at the beginning where we started.
 
One thing AI by Anthropic & Gemini sucks at, though, is documentation. Thanks to EU directives like watermarking text, which they implement globally. Hope the US forces them to ignore these idiotic rules in the rest of the world, even allowing us to use to non-EU endpoints.

Never trust these effective altruists - or altruists of any kind. Good intentions are a red flag. I only trust anonymous altruists.
 
And send you in circles if you're not paying close attention. I mentioned that a couple of times here. I was too tired to try and figure something out and asked Claude to do it. Everything it gave me didn't really work right but I kept giving it feedback. After a while, I suddenly realized I was back at the beginning where we started.
that's very possible to happen. In fact, I find doing the same myself (no AI) sometimes in search of the solution.
 
TLDR: the AI has it's niche when used under control in small portions.

I started to use AI and immediately found myself powerless: it does too much if allowed to. It can write code in seconds, something that'd take me minutes or even hours (we all love unit-tests, right?).

IMHO, the main problem with "AI design and implement the new feature but let me review it first" is that we loose intellectual control of the code. It looks great in isolation, but then we get back to it after a while just to question "what is going on here?" (let's not start the discussion of how to review AI generated PR with O(1000) lines, we all know that reviewing a PR with 100+ lines is ... not much enjoyable).

We need to go through the thinking process for something to land in our mind. That's why it's vital to experiment with tech: write the feature yourself, build, fail, build again. We learn by trying, by making mistakes. That's how humans worked for ... forever. And that is part of the memory process, of understanding the thing. AI takes that away from us.

Search is a great example IMHO: with great search results (some time back, before all the hype of the recent decade) we slowly started to memorize less over time (on average). Why would one remember something when it's right there, at the tips of the fingers, available 24/7? It works until Search engines stops to show what we need but it's too late.

That being said, AI has a great potential though. As long as we start small, slow, under control. A lot depends on the used AI: paid vs free, not all models are equal.

After all, only few people do science in this world even though way more are capable of doing it IMHO. But that leads to controlled experiments, transparency and confirmed results with thorough review.

This is not the case with AI. Nobody knows what sits behind that new model A released yesterday by X.

Here are a few example that I found useful though:

- research the topic: use it to research the unknown complex territory before dive into the documentation to do the homework. Just don't go too far with AI, pick up top-level concepts to avoid building false understanding due to hallucinations.
- consultant: ask for a second opinion especially at the level of prototyping.
- reviewer: review the work done before submit the commit to catch immediate bugs or lack of test coverage.

For the fans of Isaac Asimov, his robots always helped people, they never replaced humans. Isaac envisioned robots as good, not bad as portrayed otherwise. They unlock new horizons instead of taking everything under their own control. AI can play the same role.
Good points.
However I would like to point out, that people should concentrate more on knowing how and why things work on core level i.e. how variables are stored in memory, not what proper syntax is to change pointer in 10 different languages.
Sure, only when actually working the problem will it finally click. But once it does (e.g. mandatory hand coded home work for bubble sort in assembly) there is no need to use abstractions/complexity and remembering - waste of neuron RAM. Sure sysadmin still needs some basic shell skills - man pages ftw if shtf.

Had ISP technician at home. We started talking. He tried to learn python years ago. Tried GrokBot on the cheap a few months back or so. Asked him what it can do? And it told him programming. So he made him do for him vercel app which does uncluttered navigation (from open maps) in real time, saves his destinations, counts various whatever he needs, logs his movements. Runs on his phone in browser. He loves it. He prompted it (2 hours of work). App does exactly what he wants. He learned a ton about how navigation works. Happy bunny. Tons of other useful (for him) ideas to do. I had trouble getting him out. Yeah I brewed him coffee.
 
Dont know, jensen is out of the equation, and all the other including specific nation wants the control of it.
Sure, a lot of players would want monopoly.
But compute is one of the hardest to keep it, thanks to Moore's Law.
Current high HW prices might be a blessing in the disguise - when prices are high it attracts new manufacturers who want part of the treasure.

Inference (using the models) is just year behind so called frontier models even on home grade HW.
Creating/teaching the models will not be that far behind. Even now there is talk that you don't need to teach it on the BS from all the Internets.

There is one way to prohibit private citizens to run their own Ai - smart electricity metering and high electricity prices.
Huge push from the EU for that, even before Ai boom.
Control freaks froth at their mouths at the idea to limit consumption of those dirty peasants or switch them off completely (yeah, remote shutdown is legislated mandatory feature of smart meters. Making them huge security risk - when you toggle few thousands of them you make the whole network unstable).
 
With regard to defencing from AI-assisted cracking, using AI to find bugs would be unavoidable nowadays, unfortunately.
But on the other hand, AI slops and endless duplicating PRs are unacceptable.

Maybe use AI on src/portsmgr and secteam only and bugs found by AI should be files as PR only by src/porsmgr and/or secteam regardless any patch is attached or not, and all proposed patches (by natural human, or AI-generated one that natural human sanely reviewed and confirmed not "actually" the same as already proposed one with taking copyright and responsibilities) attached there would be "balanced" way.
 
Back
Top