nautilus12

It will inevitably lead to people not sharing their scientific results out publicly and just going directly to those who stand to benefit from them, if they know publishing means they will be immediately exploited by a single entity.

While i admire the nix community i don't know how people can be spending their time on it to solve the types of problems they are, when AI is shitting out bad python code quicker than anyone can keep up with that's degrading all our systems. By one token you can argue that's a way of fighting back, but maybe also a possible way to bury your head in the sand while the rest of the world burns around you.

You can only build skills in things that have consistent rules. LLMs have totally unpredictable outputs, attempting to build skills in prompting them is a waste of time on par with the madness of astrology or daytrading

Have any real mathematicians working on these problems reviewed any of these and determined if they are just gobbledegook or not?

The ones with lean proofs could still be formulated incorrectly

Ah i see what you were saying now, you were saying session specific poisoning is more likely to work if it was in tandem with training document specific poisoning.

Oh makes perfect sense, do you have thoughts about real life examples? I did some research using AI and it said there were examples of restricted 3 body problems like the trojan asteroids, but no examples in real life similar to what is in this web app

I was thinking to myself this dismal thought: everyone is so busy because of the acceleration caused by AI that no one will have the time to actually take the five minutes to verify manually that this really is a solution or that the article just generated something plausible looking

Is this assumed to be 2D? I was going to ask if there are any observed examples of 3 body equilibrium observed in nature.

This is the real issue, though this article does not really hit on it explicitly just alludes to it.

Thats not what they are asking. This paper is discussing documents in the training dataset poisoning the LLM for malicious behavior. This person are asking if anyone has deliberately put something in a private chat (presumably with retrain on my data turned off), to see if they can get it to leak across sessions from distinct users. I am positive this happens but I have not seen the proof. I also want to know the answer to this question.

Here are potentially relevant documents?

https://medium.com/secludy/fine-tuning-llm-on-sensitive-data...

https://spylab.ai/blog/non-adversarial-reproduction/

https://arxiv.org/abs/2601.18834

This should be a watershed moment for everyone using a commercial LLM. EVERY PIECE OF INTELLECTUAL PROPERTY YOU DISCUSS WITH IT BECOMES THEIR PROPERTY. FULL STOP.

It is impossible to fight against this. It is statistically obfuscated, they can just argue it is a derived work.

I get jealous of people that have the luxury of time to do things like this

My issue is that taking a strictly GUI driven approach makes you build an application (in the browser presumably) that will require more permissions than you would typically want in the terminal so TUIs win out by being in proximity to those tools (cat, ls, rm, etc) for development use cases

Can they try to prove that he directed someone to direct the agent to do that?

I call it "yegge sweats"

“The demands of ritual are always stronger than those of reason.”

Guess where AI Coding entered the picture

Worked for famous Japanese data platform for a few years. The Japanese engineers were collegial but some who didn't come in to the office were actually Hikikomori focused on very narrow things and were very nitpicky about details that wouldn't have ultimately mattered. Those that came into the Tokyo office lived to work and I saw people regularly sleeping at their desk after having stayed out all night for obligatory whiskey outing with colleagues and arriving at the office at 6 am as expected. The San Fran office was the opposite, very sloppy standards people, getting in at noon and staying up late to meet deadlines. The impedence mismatch between the two environments was almost unbearable.

So is this a backtrack or clarification on their original stance? Do I need to be worried about skynet killing grandma?

You are citing negative metrics. The reality is that companies only care about positive metrics: increase marginal revenue by 30%

That's regardless of the lip service they pay to cost cutting or risk reduction. It will only get worse, in the AI economy it's all about growth.

Great follow up, i agree

This would just record a lot of me cursing at and calling the AI an idiot.

Absolute power corrupts absolutely

If ever I saw something that made me want to move to a log cabin the woods and never touch a computer again, this is it

Lol, does any of this matter anymore with AI coding? It was hard to get people to care about this sort of thing before ai coding, now its impossible. People in a few years will be coding without even knowing what unix is.

Ok, why don't you work on getting 3.0 out of preview first? 10 min response time is pretty heinous

I don't think this argument makes much sense. If you are running down hill towards a cliff then saying that adding a cart to speed up the process doesn't give the cart moral blameworthiness is an unhelpful observation. You can still chose to stop running down the hill or to not get on the cart.

The irony is that the LLMs are trained on stack overflow and should inherit a lot of those traits and errors.

I had the same problem. My Claude md eventually gets forgotten and it forgets best practices that I put in there. I've switched to using hooks that run it through a variety of things like requiring testing. That seems to work better than Claude md because it has to run the hook every time it makes changes.

Unfortunately the question of whether we should is not a very popular one right now.