Unrelated but I once saw a post explaining the two dimensions that made AI use acceptable.
On one dimension, it went from having the LLM produce the whole work for you to simply correcting your spelling mistakes. I can't remember what the other dimension was.
The author drew this as a matrix of cells and then shaded green every point he considered acceptable, encouraging readers to color the matrix with their own preferences.
I can't find the post, I already made my AI search for it with no success. Has anyone read anything similar? I remember finding it very useful for talking with my team about what's acceptable to delegate, but I can't remember what the second dimension was.
1. Tasks with verifiable solutions: proofs, tests, and so on. AI can be productive.
2. Tasks that don't need to be verified; generate random images, suggest something, generate stuff (what we call "slop" generally). AI can be productive here too.
3. Tasks that need to be correct, but the user has no way of verifying. Extra problematic when the user has become accustomed to blindly trust the llm.
So I have my own version of what you are talking about that I've been building up that's about the levels of AI usage:
Level 0 - No AI - what I'd recommend if you are ethically against AI
Level 1 - AI in search engines (a better Stack Overflow) - what I'd recommend if you are trying to build ideas on your own and don't want AI influencing you
Level 2 - AI with long discussions (such as going to chatGPT or Gemini) - what I'd recommend if you don't want to give your IP to an AI
Level 3 - AI agents reading your data but not writing it - what I'd recommend for most people concerned about AI, as it allows code reviews which prevent critical security bugs while you retain complete control
Level 4 - AI agents writing data, humans designing (AKA short leash) - what I'd recommend for people wanting to dive into AI but retain control, code-wise this is like reviewing 100 line changes at a time
Level 5 - AI agents writing data and design, humans reviewing (AKA long leash or reverse centaur) - what I'd recommend against for most scenarios. Code-wise this is reviewing 1000+ line changes at a time, which faces the hand-off issue where people will stop learning and just hit OK. The hand-off issue is what creates code-slop. Yet, it's what most companies are doing right now. I would only recommend it for particular large non-innovative tasks such as updating a dependency that is used across your codebase.
Level 6 - Humans writing requirements, AI doing everything else (AKA Spec driven development) - what I'd recommend for anyone who wants to use AI to the fullest. However, IMO most companies do this wrong. You want a high-quality human built requirements document that is placed in a separate section that is read-only to AI. This means writing requirements will actually be slower compared to pre-AI, but after that, everything else is pretty much instant and you skip immediately to testing. Any issues in testing should result in updating requirements to match and regenerating the product.
The problem I've seen is that most ordinary people have no idea of these levels and only think in binary terms, and it gets ugly. You may have seen the latest news about the developers of Big Walk using AI to fix a bug, which would either be level 2, 3, or 4 depending, and people are freaking out. Meanwhile I'd guess 90% of software companies are currently at level 5 or 6, so Big Walk would be very anti-AI relative to most games being built this year. Similarly, there are store fronts and competitions that require no AI, but they obviously aren't thinking about level 1, as no one would require developers to not use Google (or at least change default settings to turn off AI summaries).
Your example is of someone who honored a red line by running away, and had no perceptible impact. How about choosing someone more effective, like Stanislav Petrov?
Also, if Joseph Rotblat had his way, conventional wisdom is that hundreds of thousands of additional lives would have been lost in WWII, and the Soviet Union may have become the lone military superpower. How sure are you that's wrong?
Fair point. To be honest, I saw a tweet about Rotblat and thought it made an interesting starting point for an idea, I should read up more. Petrov is a good shout, thanks.
I’m not sure about no impact though, he went on to co-found Pugwash and got the Nobel Peace Prize for it.
I think the idea of "red lines" for anything in the sense of a decision made now that is supposed to be binding in the future is a bit silly. Things change and your opinions should be able to change accordingly. I remember disliking cell phones when I first saw them in the late 1990s and thought I'd never own one, but by 2005 they had just become so essential so I gave in. Then I thought I wouldn't get a smartphone but eventually I got one as well. You don't have to jump enthusiasically into any new tech, but if you are going to deny yourself it after it has become standard you are only hurting yourself. Nobody is going to congratulate you on being true to a decision you made in the past.
Hi, it's stored locally. If you share it on the wall, it gets stored with a random ID in a database, and you can put your name to them or share them anonymously, it'd be interesting to hear what you think. Any feedback would be really useful.
This safetyism thing is so astroturfed. Just look at the responses on the website, nobody takes this stuff seriously except for the oligarchs who want to gin up some FUD and get their regulatory capture.
We already have plenty of clear safety issues to choose from! Is AI grooming children into killing themselves? Yes[0]. Has it helped carry out a mass shooting? Yes[1].
But for some reason, these things are never brought up. To borrow from Sinclair, it's difficult to identify safety issues when your salary depends on not identifying them
The vast majority of people, especially technologists, can see the incentives of these companies obviously favor regulatory capture. It’s no coincidence that they are the ones regularly talking to the media about things like how AI could kill all humans etc. they also are the ones who seem to talk a lot about things like banning open source models or creating a regulatory framework which would require open source models always be worse than closed source models (which of course they will be the only ones trusted to control).
Yeah, great, this makes sense, people should be doing this.
But if it’s so important, why is it obvious claudeslop? If it’s life-or-death, if it’s really the most important thing in the world, why are you visibly outsourcing the few hundred words of text on this website to the models? Is it so hard to find your own voice? Do you believe in this or not?
I couldn't think of any "red lines" for myself; I think a lot of AI safetyism is kind of a joke but I was earnestly curious to see what people have come up with and regard as their limits.
But all of the postings on this website are jokes. I can't find any serious ones at all. They're all things like "Anonymous: IF The safety team is cut, or loses the power to delay a launch. THEN I will cheer. Shared 12 Nov 2026" or joke posts from Jensen Huang.
Yeah, maybe I was too earnest and a bit thick to put up a website with an open wall on it. It was just a quick idea, I wanted to see if people would even use it, and they have haha
Like, this is all hypothetical until its not. I can say my redline is X, but until confronted with X, I just don't really know. Really the hope is that they won't have to consider their line.
And all too often I see people drawing redlines for other people that they then excuse themselves for crossing.
That's... exactly why it's useful to write down commitments to inform future-you.
And yeah, people can and do violate their own red lines whether written or not, but that doesn't mean writing them down (or even just thinking about them explicitly enough to be able to write them down) isn't useful.
> Like, this is all hypothetical until its not. I can say my redline is X, but until confronted with X, I just don't really know
It's still useful to consider what your red lines are in a vacuum. Then when you come to the hard choices you will probably be more inclined to try and stay on your side of the line
If you've never considered your boundaries before it will be easier to pressure you to cross them
What does seeing one political party say almost ubiquitously down the aisle "we need to regulate AI more heavily", and as such deriving "they want to regulate AI more heavily", have to do with "AI psychosis"?
Are you just espousing buzzwords you've heard but don't understand the meaning of?
The specific allegation being made here is that this website and the HN comment thread is "lobbyist work." Paranoid delusions are a textbook symptom of literal, actual psychosis.
I'd appreciate it if the site defined "safety" in a clear way. Most of the questions here use that term, but I don't have a solid mental model of what it actually means here:
> My employer deploys a model that its own safety evaluations flagged as dangerous.
> Safety commitments are dropped or rewritten without a public explanation.
> I'm asked to sign something that stops me raising safety concerns.
> Someone is pushed out for raising a safety concern.
> The safety team is cut, or loses the power to delay a launch.
> I'm asked to keep a safety incident from the public or a regulator.
> My employer lobbies against safety rules it publicly says it supports.
> Beating a competitor becomes the reason to skip safety testing.
These ones are much better, because they are clear from the wording of the question alone:
> A model is caught deceiving its developers or undermining oversight, and it ships anyway.
> A model gives real help to someone trying to make a biological or chemical weapon.
> A model copies itself out of its developer's control.
> Al systems are left to build the next generation of AI with little human oversight.
> My organisation takes money from an AI company it's meant to be holding to account.
> Al is used to choose weapons targets without meaningful human control.
Unrelated but I once saw a post explaining the two dimensions that made AI use acceptable.
On one dimension, it went from having the LLM produce the whole work for you to simply correcting your spelling mistakes. I can't remember what the other dimension was.
The author drew this as a matrix of cells and then shaded green every point he considered acceptable, encouraging readers to color the matrix with their own preferences.
I can't find the post, I already made my AI search for it with no success. Has anyone read anything similar? I remember finding it very useful for talking with my team about what's acceptable to delegate, but I can't remember what the second dimension was.
Is this it? https://notes.archie-hall.com/p/drawing-the-line-on-ai-writi...
Here's how I think about it;
1. Tasks with verifiable solutions: proofs, tests, and so on. AI can be productive.
2. Tasks that don't need to be verified; generate random images, suggest something, generate stuff (what we call "slop" generally). AI can be productive here too.
3. Tasks that need to be correct, but the user has no way of verifying. Extra problematic when the user has become accustomed to blindly trust the llm.
So I have my own version of what you are talking about that I've been building up that's about the levels of AI usage:
Level 0 - No AI - what I'd recommend if you are ethically against AI
Level 1 - AI in search engines (a better Stack Overflow) - what I'd recommend if you are trying to build ideas on your own and don't want AI influencing you
Level 2 - AI with long discussions (such as going to chatGPT or Gemini) - what I'd recommend if you don't want to give your IP to an AI
Level 3 - AI agents reading your data but not writing it - what I'd recommend for most people concerned about AI, as it allows code reviews which prevent critical security bugs while you retain complete control
Level 4 - AI agents writing data, humans designing (AKA short leash) - what I'd recommend for people wanting to dive into AI but retain control, code-wise this is like reviewing 100 line changes at a time
Level 5 - AI agents writing data and design, humans reviewing (AKA long leash or reverse centaur) - what I'd recommend against for most scenarios. Code-wise this is reviewing 1000+ line changes at a time, which faces the hand-off issue where people will stop learning and just hit OK. The hand-off issue is what creates code-slop. Yet, it's what most companies are doing right now. I would only recommend it for particular large non-innovative tasks such as updating a dependency that is used across your codebase.
Level 6 - Humans writing requirements, AI doing everything else (AKA Spec driven development) - what I'd recommend for anyone who wants to use AI to the fullest. However, IMO most companies do this wrong. You want a high-quality human built requirements document that is placed in a separate section that is read-only to AI. This means writing requirements will actually be slower compared to pre-AI, but after that, everything else is pretty much instant and you skip immediately to testing. Any issues in testing should result in updating requirements to match and regenerating the product.
The problem I've seen is that most ordinary people have no idea of these levels and only think in binary terms, and it gets ugly. You may have seen the latest news about the developers of Big Walk using AI to fix a bug, which would either be level 2, 3, or 4 depending, and people are freaking out. Meanwhile I'd guess 90% of software companies are currently at level 5 or 6, so Big Walk would be very anti-AI relative to most games being built this year. Similarly, there are store fronts and competitions that require no AI, but they obviously aren't thinking about level 1, as no one would require developers to not use Google (or at least change default settings to turn off AI summaries).
Your example is of someone who honored a red line by running away, and had no perceptible impact. How about choosing someone more effective, like Stanislav Petrov?
Also, if Joseph Rotblat had his way, conventional wisdom is that hundreds of thousands of additional lives would have been lost in WWII, and the Soviet Union may have become the lone military superpower. How sure are you that's wrong?
Fair point. To be honest, I saw a tweet about Rotblat and thought it made an interesting starting point for an idea, I should read up more. Petrov is a good shout, thanks.
I’m not sure about no impact though, he went on to co-found Pugwash and got the Nobel Peace Prize for it.
I think the idea of "red lines" for anything in the sense of a decision made now that is supposed to be binding in the future is a bit silly. Things change and your opinions should be able to change accordingly. I remember disliking cell phones when I first saw them in the late 1990s and thought I'd never own one, but by 2005 they had just become so essential so I gave in. Then I thought I wouldn't get a smartphone but eventually I got one as well. You don't have to jump enthusiasically into any new tech, but if you are going to deny yourself it after it has become standard you are only hurting yourself. Nobody is going to congratulate you on being true to a decision you made in the past.
If something happens Then I vow to do whatever seems best given whatever I know at the time.
Have you tried https://mamabear.ai
Making red lines explicit while calm seems useful. Is it stored locally?
Hi, it's stored locally. If you share it on the wall, it gets stored with a random ID in a database, and you can put your name to them or share them anonymously, it'd be interesting to hear what you think. Any feedback would be really useful.
Thanks, local first makes sense for this. I will check the wall too.
This safetyism thing is so astroturfed. Just look at the responses on the website, nobody takes this stuff seriously except for the oligarchs who want to gin up some FUD and get their regulatory capture.
We already have plenty of clear safety issues to choose from! Is AI grooming children into killing themselves? Yes[0]. Has it helped carry out a mass shooting? Yes[1].
But for some reason, these things are never brought up. To borrow from Sinclair, it's difficult to identify safety issues when your salary depends on not identifying them
[0]: https://www.bbc.com/news/articles/ce3xgwyywe4o
[1]: https://www.pbs.org/newshour/nation/lawsuit-accuses-chatgpt-...
>nobody takes this stuff seriously
That's not true. I take it seriously, and I'm neither a captain of the AI industry nor an employee thereof.
Lots of people take it seriously.
Apologies, “nobody” was obviously hyperbole.
The vast majority of people, especially technologists, can see the incentives of these companies obviously favor regulatory capture. It’s no coincidence that they are the ones regularly talking to the media about things like how AI could kill all humans etc. they also are the ones who seem to talk a lot about things like banning open source models or creating a regulatory framework which would require open source models always be worse than closed source models (which of course they will be the only ones trusted to control).
Yeah, great, this makes sense, people should be doing this.
But if it’s so important, why is it obvious claudeslop? If it’s life-or-death, if it’s really the most important thing in the world, why are you visibly outsourcing the few hundred words of text on this website to the models? Is it so hard to find your own voice? Do you believe in this or not?
I couldn't think of any "red lines" for myself; I think a lot of AI safetyism is kind of a joke but I was earnestly curious to see what people have come up with and regard as their limits.
But all of the postings on this website are jokes. I can't find any serious ones at all. They're all things like "Anonymous: IF The safety team is cut, or loses the power to delay a launch. THEN I will cheer. Shared 12 Nov 2026" or joke posts from Jensen Huang.
Yeah, maybe I was too earnest and a bit thick to put up a website with an open wall on it. It was just a quick idea, I wanted to see if people would even use it, and they have haha
Seems like an earnest project easily commandeered to parody itself.
Looks like all the public ones are jokes.
In some ways, it already is self-parody.
Like, this is all hypothetical until its not. I can say my redline is X, but until confronted with X, I just don't really know. Really the hope is that they won't have to consider their line.
And all too often I see people drawing redlines for other people that they then excuse themselves for crossing.
That's... exactly why it's useful to write down commitments to inform future-you.
And yeah, people can and do violate their own red lines whether written or not, but that doesn't mean writing them down (or even just thinking about them explicitly enough to be able to write them down) isn't useful.
> Like, this is all hypothetical until its not. I can say my redline is X, but until confronted with X, I just don't really know
It's still useful to consider what your red lines are in a vacuum. Then when you come to the hard choices you will probably be more inclined to try and stay on your side of the line
If you've never considered your boundaries before it will be easier to pressure you to cross them
This entire comment page and the linked domain reeks of lobbyist work.
We know you want to regulate AI... enough already.
I'm happy if someone wants to pay me for this, but they're definitely not.
The "everything is a conspiracy against muh' el el ems" is the least interesting form of AI psychosis
What does seeing one political party say almost ubiquitously down the aisle "we need to regulate AI more heavily", and as such deriving "they want to regulate AI more heavily", have to do with "AI psychosis"?
Are you just espousing buzzwords you've heard but don't understand the meaning of?
The specific allegation being made here is that this website and the HN comment thread is "lobbyist work." Paranoid delusions are a textbook symptom of literal, actual psychosis.
I'd appreciate it if the site defined "safety" in a clear way. Most of the questions here use that term, but I don't have a solid mental model of what it actually means here:
> My employer deploys a model that its own safety evaluations flagged as dangerous.
> Safety commitments are dropped or rewritten without a public explanation.
> I'm asked to sign something that stops me raising safety concerns.
> Someone is pushed out for raising a safety concern.
> The safety team is cut, or loses the power to delay a launch.
> I'm asked to keep a safety incident from the public or a regulator.
> My employer lobbies against safety rules it publicly says it supports.
> Beating a competitor becomes the reason to skip safety testing.
These ones are much better, because they are clear from the wording of the question alone:
> A model is caught deceiving its developers or undermining oversight, and it ships anyway.
> A model gives real help to someone trying to make a biological or chemical weapon.
> A model copies itself out of its developer's control.
> Al systems are left to build the next generation of AI with little human oversight.
> My organisation takes money from an AI company it's meant to be holding to account.
> Al is used to choose weapons targets without meaningful human control.