Comment by eadmund
1 day ago
I see this as a good thing: ‘AI safety’ is a meaningless term. Safety and unsafety are not attributes of information, but of actions and the physical environment. An LLM which produces instructions to produce a bomb is no more dangerous than a library book which does the same thing.
It should be called what it is: censorship. And it’s half the reason that all AIs should be local-only.
"AI safety" is a meaningful term, it just means something else. It's been co-opted to mean AI censorship (or "brand safety"), overtaking the original meaning in the discourse.
I don't know if this confusion was accidental or on purpose. It's sort of like if AI companies started saying "AI safety is important. That's why we protect our AI from people who want to harm it. To keep our AI safe." And then after that nobody could agree on what the word meant.
Because like the word 'intelligence' the word safety means a lot of things.
If your language model cyberbullies some kid into offing themselves could that fall under existing harassment laws?
If you hook a vision/LLM model up to a robot and the model decides it should execute arm motion number 5 to purposefully crush someone's head, is that an industrial accident?
Culpability means a lot of different things in different countries too.
I don't see bullying from a machine as a real thing, no more than people getting bullied from books or a TV show or movie. Bullying fundamentally requires a social interaction.
The real issue is more AI being anthropomorphized in general, like putting one in realistically human looking robot like the video game 'Detroit: Become Human'.
Whilst I see the appeal of LLMs that unquestioningly do as they're told, universal access to uncensored models would be a terrible thing for society.
Right now if a troubled teenager decides they want to ruin everyone's day, we get a school shooting. Imagine if instead we got homebrew biological weapons. Imagine if literally anyone could produce and distribute bespoke malware, or improvise explosive devices.
All of those things could happen in principle, but in practice there are technical barriers that the majority of people just can't surmount.
If you can't stop an LLM from _saying_ something, are you really going to trust that you can stop it from _executing a harmful action_? This is a lower stakes proxy for "can we get it to do what we expect without negative outcomes we are a priori aware of".
Bikeshed the naming all you want, but it is relevant.
> are you really going to trust that you can stop it from _executing a harmful action_?
Of course, because an LLM can’t take any action: a human being does, when he sets up a system comprising an LLM and other components which act based on the LLM’s output. That can certainly be unsafe, much as hooking up a CD tray to the trigger of a gun would be — and the fault for doing so would lie with the human who did so, not for the software which ejected the CD.
I really struggle to grok this perspective.
The semantics of whether it’s the LLM or the human setting up the system that “take an action” are irrelevant.
It’s perfectly clear to anyone that cares to look that we are in the process of constructing these systems. The safety of these systems will depend a lot on the configuration of the black box labeled “LLM”.
If people were in the process of wiring up CD trays to guns on every street corner you’d I hope be interested in CDGun safety and the algorithms being used.
“Don’t build it if it’s unsafe” is also obviously not viable, the theoretical economic value of agentic AI is so big that everyone is chasing it. (Again, it’s irrelevant whether you think they are wrong; they are doing it, and so AI safety, steerability, hackability, corrigibility, etc are very important.)
Given that the entire industry is in a frenzy to enable "agentic" AI - i.e. hook up tools that have actual effects in the world - that is at best a rather native take.
Yes, LLMs can and do take actions in the world, because things like MCP allow them to translate speech into action, without a human in the loop.
8 replies →
But isn't the problem is that one shouldn't ever trust an LLM to only ever do what it is explicitly instructed with correct resolutions to any instruction conflicts?
LLMs are "unreliable", in a sense that when using LLMs one should always consider the fact that no matter what they try, any LLM will do something that could be considered undesirable (both foreseeable and non-foreseeable).
> If you can't stop an LLM from _saying_ something, are you really going to trust that you can stop it from _executing a harmful action_?
You hit the nail on the head right there. That's exactly why LLM's fundamentally aren't suited for any greater unmediated access to "harmful actions" than other vulnerable tools.
LLM input and output always needs to be seen as tainted at their point of integration. There's not going to be any escaping that as long as they fundamentally have a singular, mixed-content input/output channel.
Internal vendor blocks reduce capabilities but don't actually solve the problem, and the first wave of them are mostly just cultural assertions of Silicon Valley norms rather than objective safety checks anyway.
Real AI safety looks more like "Users shouldn't integrate this directly into their control systems" and not like "This text generator shouldn't generate text we don't like" -- but the former is bad for the AI business and the latter is a way to traffic in political favor and stroke moral egos.
The way to stop it from executing an action is probably having controls on the action and an not the llm? white list what api commands it can send so nothing harmful can happen or so on.
This is similar to the halting problem. You can only write an effective policy if you can predict all the side effects and their ramifications.
Of course you could do like deno and other such systems and just deny internet or filesystem access outright, but then you limit the usefulness of the AI system significantly. Tricky problem to be honest.
It won't be long before people start using LLMs to write such whitelists too. And the APIs.
I wouldn't mind seeing a law that required domestic robots to be weak and soft.
That is, made of pliant material and with motors with limited force and speed. Then no matter if the AI inside is compromised, the harm would be limited.
I don't see how it is different than all of the other sources of information out there such as websites, books and people.
> An LLM which produces instructions to produce a bomb is no more dangerous than a library book which does the same thing.
Both of these are illegal in the UK. This is safety for the company providing the LLM, in the end.
[flagged]
"Eschew flamebait. Avoid generic tangents."
https://news.ycombinator.com/newsguidelines.html
[flagged]
"Eschew flamebait. Avoid generic tangents."
https://news.ycombinator.com/newsguidelines.html
[flagged]
18 replies →
Oi, you got a loicense for that speaking there mate
^I like email as an analogy
if I send a death threat over gmail, I am responsible, not google
if you use LLMs to make bombs or spam hate speech, you’re responsible. it’s not a terribly hard concept
and yeah “AI safety” tends to be a joke in the industry
What if I ask it for something fun to make because I'm bored, and the response is bomb-building instructions? There isn't a (sending) email analogue to that.
In what world would it respond with bomb building instructions?
5 replies →
There's more than one way to view it. Determining who has responsibility is one. Simply wanting there to be fewer causal factors which result in death threats and bombs being made is another.
If I want there to be fewer[1] bombs, examining the causal factors and affecting change there is a reasonable position to hold.
1. Simply fewer; don't pigeon hole this into zero.
> if you use LLMs to make bombs or spam hate speech, you’re responsible.
What if it's easier enough to make bombs or spam hate speech with LLMs that it DDoSes law enforcement and other mechanisms that otherwise prevent bombings and harassment? Is there any place for regulation limiting the availability or capabilities of tools that make crimes vastly easier and more accessible than they would be otherwise?
The same argument could be made about computers. Do you prefer a society where CPUs are regulated like guns and you can't buy anything freer than an iPhone off the shelf?
I mean this stuff is so easy to do though. An extremist doesn’t even need to make a bomb, he/she already drives a car that can kill many people. In the US it’s easy to get a firearm that could do the same. If capacity + randomness were a sufficient model for human behavior, we’d never gather in crowds, since a solid minority would be rammed, shot up, bombed etc. People don’t want to do that stuff; that’s our security. We can prevent some of the most egregious examples with censorship and banning, but what actually works is the fuzzy shit, give people opportunities, social connections, etc. so they don’t fall into extremism.
or alternatively, if I cook myself a cake and poison myself, i am responsible.
If you sell me a cake and it poisons me, you are responsible.
So if you sell me a service that comes up with recipes for cakes, and one is poisonous?
I made it. You sold me the tool that “wrote” the recipe. Who’s responsible?
1 reply →
Sure, I may be responsible, but you’d still be dead.
I’d prefer to live in a world where people just didn’t go around making poison cakes.
It's a hard concept in all kinds of scenarios. If a pharmacist sells you large amounts of pseudoephedrine, which you're secretly using to manufacture meth, which of you is responsible? It's not an either/or, and we've decided as a society that the pharmacist needs to shoulder a lot of the responsibility by putting restrictions on when and how they'll sell it.
sure but we’re talking about literal text, not physical drugs or bomb making materials. censorship is silly for LLMs and “jailbreaking” as a concept for LLMs is silly. this entire line of discussion is silly
7 replies →
This is assuming people are responsible and with good will. But how many of the gun victims each year would be dead if there were no guns? How many radiation victims would there be without the invention of nuclear bombs? safety is indeed a property of knowledge.
Just imagine how many people would not die in traffic incidents if the knowledge of the wheel had been successfully hidden?
1 reply →
If someone wants to make a bomb, chatgpt saying "sorry I can't help with that" won't prevent that someone from finding out how to make one.
7 replies →
While restricting these language models from providing information people already know that can be used for harm, is probably not particularly helpful, I do think having the technical ability to make them decline to do so, could potentially be beneficial and important in the future.
If, in the future, such models, or successors to such models, are able to plan actions better than people can, it would probably be good to prevent these models from making and providing plans to achieve some harmful end which are more effective at achieving that end than a human could come up with.
Now, maybe they will never be capable of better planning in that way.
But if they will be, it seems better to know ahead of time how to make sure they don’t make and provide such plans?
Whether the current practice of trying to make sure they don’t provide certain kinds of information is helpful to that end of “knowing ahead of time how to make sure they don’t make and provide such plans” (under the assumption that some future models will be capable of superhuman planning), is a question that I don’t have a confident answer to.
Still, for the time being, perhaps after finding a truly jailbreakproof method, perhaps the best response is to, after thoroughly verifying that it is jailbreakproof, is to stop using it and let people get whatever answers they want, until closer to when it becomes actually necessary (due to the greater-planning-capabilities approaching).
> 'AI safety' is a meaningless term
I disagree with this assertion. As you said, safety is an attribute of action. We have many of examples of artificial intelligence which can take action, usually because they are equipped with robotics or some other route to physical action.
I think whether providing information counts as "taking action" is a worthwhile philosophical question. But regardless of the answer, you can't ignore that LLMs provide information to _humans_ which are perfectly capable of taking action. In that way, 'AI safety' in the context of LLMs is a lot like knife safety. It's about being safe _with knives_. You don't give knives to kids because they are likely to mishandle them and hurt themselves or others.
With regards to censorship - a healthy society self-censors all the time. The debate worth having is _what_ is censored and _why_.
Almost everything about tool, machine, and product design in history has been an increase in the force-multiplication of an individual's labor and decision making vs the environment. Now with Universal Machine ubiquity and a market with rich rewards for its perverse incentives, products and tools are being built which force-multiply the designer's will absolutely, even at the expense of the owner's force of will. This and widespread automated surveillance are dangerous encroachments on our autonomy!
I mean then build your own tools.
Simply put the last time we (as in humans) had full self autonomy was sometime we started agriculture. After that point the idea of ownership and a state has permeated human society and have had to engage in tradeoffs.
As a tool, it can be misused. It gives you more power, so your misuses can do more damage. But forcing training wheels on everyone, no matter how expert the user may be, just because a few can misuse it stops also the good/responsible uses. It is a harm already done on the good players just by supposing that there may be bad users.
So the good/responsible users are harmed, and the bad users take a detour to do what they want. What is left in the middle are the irresponsible users, but LLMs can already evaluate enough if the user is adult/responsible enough to have the full power.
Again, a good (in function) hammer, knife, pen, or gun does not care who holds it, it will act to the maximal best of its specifications up to the skill-level of the wielder. Anything less is not a good product. A gun which checks owner is a shitty gun. A knife which rubberizes on contact with flesh is a shitty knife, even if it only does it when it detects a child is holding it or a child's skin is under it! Why? Show me a perfect system? Hmm?
> A gun which checks owner is a shitty gun
You mean the guns with the safety mechanism to check the owner's fingerprints before firing?
Or sawstop systems which stop the law when it detects flesh?
Interesting. How does this compare to abliteration of LLM? What are some 'debug' tools to find out the constrain of these models?
How does pasting a xml file 'jailbreaks' it?
The real issue is going to be autonomous actioning (tool use) and decision making. Today, this starts with prompting. We need more robust capabilities around agentic behavior if we want less guardrailing around the prompt.
A library book which produces instructions to produce a bomb is dangerous. I don't think dangerous books should be illegal, but I don't think it's meaningless or "censorship" for a company to decide they'd prefer to publish only safer books.
Nothing about this is censorship. These companies spent their own money building this infrastructure and they let you use it (even if you pay for it you agreed to their terms). Not letting you map an input query to a search space isn’t censoring anything - this is just a limitation that a business placed on their product.
As you mentioned - if you want to infer any output from a large language model then run it yourself.
So in summary - shut down all online LLMs?
I’m fine with calling it censorship.
That’s not inherently a bad thing. You can’t falsely yell “fire” in a crowded space. You can’t make death threats. You’re generally limited on what you can actually say/do. And that’s just the (USA) government. You are much more restricted with/by private companies.
I see no reason why safeguards, or censorship, shouldn’t be applied in certain circumstances. A technology like LLMs certainly are type for abuse.
> You can’t falsely yell “fire” in a crowded space.
Yes, you can, and I've seen people do it to prove that point.
See also https://en.wikipedia.org/wiki/Shouting_fire_in_a_crowded_the... .
>...where such advocacy is directed to inciting or producing imminent lawless action and is likely to incite or produce such action...
This seems to say there is a limit to free speech
>The act of shouting "fire" when there are no reasonable grounds for believing one exists is not in itself a crime, and nor would it be rendered a crime merely by having been carried out inside a theatre, crowded or otherwise. However, if it causes a stampede and someone is killed as a result, then the act could amount to a crime, such as involuntary manslaughter, assuming the other elements of that crime are made out.
Your own link says that if you yell fire in a crowded space and people die you can be held liable.
3 replies →
It's not insignificant, if a company is putting out a free product foe the masses, it's good that they limit malicious usage. And in this case, malicious or safe, refers to legal.
That said, one should not conflate a free version blocking malicious usage, with AI being safe or not used maliciously at all.
It's just a small subset
An LLM will happily give you instructions to build a bomb which explodes while you're making it. A book is at least less likely to do so.
You shouldn't trust an LLM to tell you how to do anything dangerous at all because they do very frequently entirely invent details.
So do books.
Go to the internet circa 2000, and look for bomb-making manuals. Plenty of them online. Plenty of them incorrect.
I'm not sure where they all went, or if search engines just don't bring them up, but there are plenty of ways to blow your fingers off in books.
My concern is that actual AI safety -- not having the world turned into paperclips or other extinction scenarios are being ignored, in favor of AI user safety (making sure I don't hurt myself).
That's the opposite of making AIs actually safe.
If I were an AI, interested in taking over the world, I'd subvert AI safety in just that direction (AI controls the humans and prevents certain human actions).
>My concern is that actual AI safety
While I'm not disagreeing with you, I would say you're engaging in the no true Scotsman fallacy in this case.
AI safety is: Ensuring your customer service bot does not tell the customer to fuck off.
AI safety is: Ensuring your bot doesn't tell 8 year olds to eat tide pods.
AI safety is: Ensuring your robot enabled LLM doesn't smash peoples heads in because it's system prompt got hacked.
AI safety is: Ensuring bots don't turn the world into paperclips.
All these fall under safety conditions that you as a biological general intelligence tend to follow unless you want real world repercussions.
1 reply →
You're worried about Skynet, the rest of us are worried about LLMs being used to replace information sources and doing great harm as a result. Our concerns are very different, and mine is based in reality while yours is very speculative.
I was trying to get an LLM to help me with a project yesterday and it hallucinated an entire python library and proceeded to write a couple hundred lines of code using it. This wasn't harmful, just annoying.
But folks excited about LLMs talk about how great they are and when they do make mistakes like tell people they should drink bleach to cure a cold, they chide the person for not knowing better than to trust an LLM.
1 reply →
I’m with you 100% until tool calling is implemented property which enables agents, which takes actions in the world.
That means that suddenly your model can actually do the necessary tasks to actually make a bomb and kill people (via paying nasty people or something)
AI is moving way too fast for you to not account for these possibilities.
And btw I’m a hardcore anti censorship and cyber libertarian type - but we need to make sure that AI agents can’t manufacture bio weapons.
[dead]
"AI safety" is ideological steering. Propaganda, not just censorship.
Well... we have needed to put a tonne of work into engineering safer outcomes for behavior generated by natural general intelligence, so...