Potential 6 month warning until the end

[Frombeyond]

So, we potentially only have 6 months until the end of civilised society and mass deaths. I’m terrified and sad

AI staff 'genuinely frightened' for humanity's future, ex-Anthropic researcher tells BBC www.bbc.co.uk/news/articles/c1kx0gyje9wo

And who are ´they’

The world was going to be exterminated by the Millennium Bug.
Then it was mass terrorism.
Then it was weapons of mass destruction.
Then it was a renaissance of the greenhouse effect.
Some time after this (I can't be bothered to think back), it was Covid.
Then it was being caught up in the Ukraine war.
Then it was monkeypox.
Then it was extreme heat.

All to sell papers, by keeping the public frightened.

Among all the fearmongering noise which the press makes every day, it can be very hard to spot the genuine threats.

This is reminiscent of the "duck and cover" drills in US schools in the 50s. MrGonners remembers them vaguely, but mainly remembers his mum (who was a teacher) laughing at the futility of them. They were intended to shield you from the blast but, as she pointed out, they didn't shield you from anything else.

Also of the frankly hilarious instructions given to British army wives living near the East German border in the 70s. "If the tanks start to roll in, put your kids in the car, drive to the Hook of Holland and get on the ferry home." Never mind that you can't drive, or that your husband has taken the only family car to drive to work and will be too busy soldiering to bring it home. Just do it! My friend, who was in that situation, just laughed.

Scaremongering has no benefit and is abhorrent.

I agree . I mean I’m not going to loose sleep over it. No point but the fact they are worried is absolutely significant to me.

You will be using it, though, all the time, without knowing it. Everything you do on the internet will have some AI functions running. Every hospital uses AI.

???

One thing I'd do differently when I was younger is spend a lot more - if I had, I would now be entitled to A LOT more government help than I am currently. I'm working class, but because I saved so much and lived such a frugal life, the state treats me no differently to a middle class person with family money because of those savings I have.

Who are you including in that ‘we’? I use pen and paper daily; I much prefer it to typing on a screen or keyboard. I enjoy the physical process of putting ink on paper and I believe it helps my thinking. I get through quite a lot of notebooks (but I do buy recycled ones.)

Your writing is really hard to comprehend. I think you’re arguing that the only concern should be with how people choose to use AI, whereas I believe the OP is highlighting concerns about AI taking actions autonomously that are harmful to us.

I don’t think it’s hard to know how people are using AI, in the sense that prompts can be monitored and routing to whatever model / server is all easily traced. I do think it’s hard to provide enough tools to the AI that it can be useful while also locking it down enough that it can’t do harm, and hard to then identify actions initiated by the AI that will cause harm before they are executed. The labs use chain of thought monitoring on LLM outputs to attempt to identify potential bad behavior, my initial response to you was a brief argument for why CoT monitoring is insufficient (which of course they all know, but since we don’t have good solutions we use the imperfect ones).

Yes I understand what a large language model is, but from this post I’m not at all sure that you do!?

God I hate pens and curly writing.

Babe, you are constructing your answer using artificial intelligence, but you are not actually able to understand my point. A large language model is basically a database; you are either agreeing with what I am saying, or you disagree to make a point that I do not quite understand. This is my fundamental response: No, AI is not going to cause the end of the world in six months. Yes, AI is dangerous, but it is dangerous because of humans. It is not dangerous in and of itself; it is not a living entity, and it does not work like that.

I'm very concerned, I have two young children and another on the way. I can see it leading to financial crisis, internet outages, and potentially blackouts. All of these will mean deaths, deaths that don't need to happen. Most people hate AI, it's just taking jobs, electricity, water, and now it may be out of control enough to take lives.

My problem is our governments seem to be just sleepwalking into it with very rich men deciding the fate of the world. They claim a race with China, but I can only imagine China doesn't want it's own economy and people plunged into crisis? The CCP will be nothing of they too have no power, internet, or economy.

I don't know how serious this is. You have people on one side declaring total human extinction, and on the other dismissing everything as a PR stunt. In my view, even a small chance the people declaring catastrophic results should mean the global governments get their act together and take it seriously. I would rather they be over cautious over a PR stunt then do little and watch everyone die. What harm will international oversight and pausing AI do? It only has positives.

My only hope is that people in power, and the wealthy, have children they love too. My DCs are the most precious things in the world to me and all I want is for them to grow up and thrive. My worry is they believe they are too rich to fail and they can't see a world in which they don't survive.

I personally think the best way to combat it is to make the creators of AI responsible for their creation. Prison sentences, astronomical fines and international courts if their creation does something that, if a person did it, would range from a crime to terrorism. They are in effect handing a weapon over to nefarious individuals with poor safety regulations. So far we have heard about hacking attacks, biological weapon leaks, and chatgpt encouraged suicides and all they seem to do is shrug and carry on. I would imagine if their precious billions and their freedoms were at stake they might start to care.

Also,. Using Chain of thought to identify bad behavior it's an exotic description of what is quite a normal prosaic data cleansing activity to control output from the interface. It's really not as special or as magical as the mister sysms that the people that are working in AI are putting out, it really does seem to be like a money driven or a power driven conversation with an intention that I don't fully understand and I mean, for example, Elon. Musk, saying AI is dangerous. What does he mean by that?He knows exactly how it works.What it can do and what it can not do, so there's something else going on there. th—I'm not sure where you're getting this from. Generally, an artificial intelligence model in an institution is governed by the same data controls you would put in place for a database the way it reaches its answers is basically by pulling from a large data source to arrive at a conclusion. It doesn't do anything innovative in itself without being directed, driven, or taught, because human input is what triggers a bad reaction; it's the human input that causes the problem. AI itself is just a tool like a calculator. Does that mean calculators can't be dangerous? No, of course not; calculators in the wrong hands are deadly, but that's the point. It's the fearmongering over artificial intelligence as though it's uncontrollable that's the problem. That's not true; the tool is controllable; the tool user is not.

They’ll spare me because I always say please and thank you when I used ChatGPT.

. I can understand tge worry the way this is being out out is odd. The reason why it's dangerous is more in the control of those tech companies than they would kind of want to admit to I do think they need to be held to a different kind of standard, and I don't think their s being policed properly

Oh well not including you then 😉 people don’t write with a pen much around where I am, I mean they still occasionally do but..

Haha me too!

Well if AI takes the internet down we’ll need to get used to it again 😂

I know im legit able to compete because of computers. Back to pen a papers im done.

Definitely never heard this joke before! 🙄😂

🤷🏻‍♀️

Computers and word processors are one thing, AI is quite another! I do appreciate that some people can write much better using technology.

I hope people think like that but honestly these days it feels like people are becoming fearful of the whole thing. But i tell you something, and like, you know, I'll be honest. I'm definitely easy to read to easy to describe as someone that slightly away with the fairies, but people can't compartmentalise, I realise they can't do what you've just done and divide each 3 aspects of the technology and rank them differently. It's weird when people are afraid is something it kind of just everything plays together. And I think if you want an honest conversation, you can't say that AI is not dangerous. And that's what makes it difficult because that were dangerous kind of just begins to inflate everything I don't know.Oh, i've enjoyed this chat.I'm off the bed night girls.

Babe :)

constructing my answer using AI? I’m just writing a response, no AI involved. You are right that I’ve not understood your point though!

I agree that AI is unlikely to cause the end of the world in the next 6 months.

I agree it’s dangerous because of humans ie if a human decides to use it to help design a weapon or dangerous drug or whatever.

AI is also potentially dangerous even when asked to do something that the human user did not intend to be dangerous. This, I think, is where your belief differs from the OP and from Amodei etc. A recent high profile example is the hugging face hack. The AI was asked to solve some benchmarking problems. It then accessed servers it should not have been able to get to in order to cheat. It was never instructed to do that, in fact the humans who set up the system thought it was not possible (the agent was running in a sandbox that was thought to be isolated from the internet). The hack was not noticed until much later.

“just turn it off” (which you suggested as a reason that AI cannot really be dangerously) requires knowing when harm is intended or is actually occurring. My response to you was only to say that it’s not easy to know that.

Model providers attempt to limit intentional harm from humans by instructing models to not answer questions in certain topics. Eg of you ask Claude to help build a bomb it’s probably going to say no. But it’s not especially robust and there are open source models that don’t have those safeguards.

Model providers attempt to limit harm from model actions that were not intended by humans by automatically monitoring the output of the LLMs and putting restrictions on tool use. But that also isn’t very robust. For tools to be useful they inevitably end up having some potential for harm. Automated monitoring of LLM output is imperfect and LLMs can quite effectively hide their intentions.

If just switch it off was really a good solution do you not think all the AI safety people at the big research labs would have thought of that already?

I wholeheartedly agree with everything you have written.