In other words the priority of the work is to get these types of people into positions where they don't do any work.
At least with privacy groups you do get here and there some practical advice on using ublock origin or more rarely on how to install a blocklist from https://someonewhocares.org/hosts/, but with AI ethics & safety orgs... well lets put it this way.
I have yet to meet a single AI safety person that knows how to rename a file in linux.
God forbid we have a rogue AI-worm shutting down all servers & BGP routers while these types of people were in charge of safety, they'll be in the way of anyone even fixing it. They can't even get a simple safety benchmark working on lm_eval-harness. They're great at lecturing you why they shouldn't need that.
And this is the key issue with AI Ethics. It's the refusal to work at the problem constructively, and get the most skilled people possible to actually make the damn benchmarks to work, to rank models on the understanding of human rights, to list every current violation and abuse of humans in every single country without exception and to make practical plans on what to do when systems go rogue. Even if they're not technical they could be making the dataset in a csv in excel for that and making it public domain accessible.
Instead we get the most depressed, leechy-office-worker types complaining about how it's all over.
Now back to work, move it.
You are meeting very few AI safety people then. A significant fraction of the AI safety people I've run into can build and train an LLM from scratch (without AI assistance FWIW), let alone have a grasp of basic command-line operations.
I don't know if instead of saying "safety" here you meant to say ethics, or if you're using "safety" in this sentence just to generally refer to "AI ethics, safety, and to a smaller extent privacy."
If either of those are true, that's weird because the only person in AI ethics most people know is Timnit Gebru, because she got fired and it made the papers. She has a BA and MA in electrical engineering, and her father was also an electrical engineer. After that, she went on to a PhD in computer vision with Fei-Fei Li (Imagenet) as her advisor.
https://en.wikipedia.org/wiki/Timnit_Gebru#Early_life_and_ed...
I guarantee you she knows how to rename a file in Linux.
If, instead, you were referring to "safety" specifically, I'd like to understand how you're making the distinction.
edit:
> Gebru joined Apple as an intern while at Stanford, working in their hardware division making circuitry for audio components, and was offered a full-time position the following year. Of her work as an audio engineer, her manager told Wired she was "fearless", and well-liked by her colleagues. During her tenure at Apple, Gebru became more interested in building software, namely computer vision that could detect human figures. She went on to develop signal processing algorithms for the first iPad. At the time, she said she did not consider the potential use for surveillance, saying "I just found it technically interesting."
Serious safety researchers are doing stuff like understanding neural circuits. Very different.
She surely knows how to use Linux. But she isn't really a safety researcher.
Couldn't you just ask the model that?
Joking of course, but this is in and of itself an intractable problem. Do you mean "restate the principles of human rights", which is a pretty small subset of law which is in turn a small subset of ethics, or do you mean actually get out there and enumerate and name every single person having their human rights violated? Not only is that an absurd amount of work it's politically impossible.
The whole point was that these people just complain and make smartass excuses instead of getting any actual work done.
It's funny, because I can tell when I'm dealing with a non-technical, policy person, because they always use language in a way no programmer would. I'd prefer not share the specific tells, but there are some things technical people just don't say, but policy people say all the time.
Not a carefully reasoned argument why “not causing harm to a human” is outmoded, but just pushing it aside. I would love to see a good reasoned argument there.
No, instead there is Avoiding talking about harm to humans. Just because harm is broad doesn’t get you out of having to talk about it and deal with risks, which is at the root of engineering.
Instead of us programming the AIs by feeding it lots of explicit hand-crafted rules/instructions, we're feeding the things with plain data instead, and the resulting behavior is much more black-box, less predictable and less controllable than anticipated.
Training LLMs is closer, conceptually, to raising children than to implementing regexp parsers, and the whole "small simple set of universal constraints" is just not really applicable/useful.
People often misunderstand Asimov's laws, the entire point of the laws and the stories they're set in was that you can't just throw a simple "Don't hurt people" clause at a black box like AI and expect good results. You first have to define "Don't", then you have to define "hurt" and perhaps the hardest of all is you have to define "people". And I mean really define it, to the smallest most minute detail of what exactly all those words mean. Otherwise you very quickly run into funny, tragic and even contradictory situations, and those situations are endlessly unique.
Is feeding grossly unhealthy food to a starving person harm? Perhaps not, you can argue it's better to eat something unhealthy than to starve. What about feeding someone on the brink of a cardiac arrest that same meal? Now what about all the other gray areas involved here, you have to define every single possible situation in which an unhealthy meal might affect someone.
It's kinda funny, because it really is almost prophetic considering it's a story written quite a long time before we were even close to it being a reality...
“Don’t touch the knife” becomes “You can use _this_ knife, if an adult is watching,” which becomes “You can use these knives but you have to be careful, tell me what that means” and then “you have free run of the knife drawer, the bandages are over there.” But there’s careful supervision at each step and you want to see that they’re ready before moving up. I haven’t seen any evidence of that at all in LLM training—it seems to be more akin to handing each toddler every book ever written about knives and a blade and waiting to see what happens.
That this can be said, and there still being so doubt we should ramp up the Ethics research before going and rawdogging the implementation just bloody bewilders me.
I dunno, we do feed them lots of explicit hand-crafted rules/instructions, it's just that does don't go into the training process, but instead goes into the "system"/"developer" prompts, which is effectively the way you "program" the LLMs.
So you start out with nothing, adjust the weights based on the datasets until you reach something that allows you to "program" them via the system/developer prompts, which considering what's happening behind the scenes, is more controllable than expected.
Well then that's terrifying - the problem with children is that you can raise them perfectly and they still end up psychos. That's mostly limited by the fact that we can't raise many children and humans are pretty limited in the damage they can do.
But AI scales infinitely, and if we give it access to too much stuff then the damage it could do could be human race ending.
We gotta work on getting the other models to agree.
Right, and so do the harm risks. We need a framework centered around how humans will use AI/robots to harm each other, not how AI/robots will autonomously harm humans.
You say "Asimov’s three laws are not a good framework.", then don't present any arguments to why it is not a good framework. Instead you bring up something separate: the framework can facilitate story writing.
It could be good for story writing and a good framework. Those two aren't mutually exclusive things. (I'm not arguing that it is a good framework or not, I haven't thought about it enough)
Suggesting they be used as a basis for actual AI ethics is...well, it's not quite to the level of creating the Torment Nexus from acclaimed sci-fi novel "Don't Create the Torment Nexus", but it's pretty darn close.
"we want money from selling weapons"
Also there's the ethics of scraping the whole internet and claiming that it's all fair use, because the other scenario is a little too inconvenient for all the companies involved.
P.S.: I expect a small thread telling me that it's indeed fair use, because models "learn and understand just like humans", and "models are hugely transformative" (even though some licenses say "no derivatives whatsoever"), "they are doing something amazing so they need no permission", and I'm just being naive.
Unless you believe this will kill AI, all it does is to create a bunch of data brokers.
Once fees are paid, data is exchanged, and models are trained, if the AI takes your job of programming/drawing/music, then it still does. We arrived at the same destination, only with more lawyers in the mix. You get to enjoy unemployment only knowing that lawyers made sure that at least they didn't touch your cat photos.
It all depends on what is most convenient for avoiding any accountability.
As such fair use is whatever the courts say it is.
What the deflection is away from is that the actual business plan here is the same one tech has been doing for a decade: welding every flow and store of data in the world to their pipelines, mining every scrap of information that passes through and giving themselves the ability to shape the global information landscape, and then sell that ability to the highest bidders.
The difference with "AI" is that they finally have a way to convince people to hand over all the data.
>People are far more concerned with the real-world implications of ethics: governance structures, accountability, how their data is used, jobs being lost, etc. In other words, they’re not so worried about whether their models will swear or philosophically handle the trolley problem so much as, you know, reality. What happens with the humans running the models? Their influx of power and resources? How will they hurt or harm society?
This is just not my experience at all. People do worry about how models act because they infer that eventually they will be used as source of truth and because they already get used as source of action. People worry about racial makeup in certain historical contexts[1], people worry when Grok starts spouting Nazi stuff (hopefuly I don't need a citation for that one) because they take it as a sign of bias in a system with real world impact, that if ChatGPT happens to doubt the holocaust tomorrow, when little Jimmy asks it for help in an essay he will find a whole lot of white supremacist propaganda. I don't think any of this is fictional.
I find the same issue with the privacy section. Yes concerns about privacy are primarily about sharing that data, precisely because controlling how that data is shared is a first, necessary step towards being able to control what is done with the data. In a world in which my data is taken and shared freely I don't have any control on what is done with that data because I have no control on who has it in the first place.
[1] https://www.theguardian.com/technology/2024/mar/08/we-defini...
These things are also concerns and definitely shouldn't be dismissed entirely (especially things like AI telling you when it's unsure, or, the worse cases of propaganda), but I'm worried about the other stuff I mention being defined away entirely, the same way I think it has been with privacy. Tons more to say on the difference between "how you use" vs "how you share" but good perspective, and interesting that you see the emphasis differently in your experiences.
For example, from Timnit Gebru:
> The fact that they call themselves "AI Safety" and call us "AI Ethics" is very interesting to me.
> What makes them "safety" and what makes us "ethics"?
> I have never taken an ethics course in my life. I am an electrical engineer and a computer scientist however. But the moment I started talking about racism, sexism, colonialism and other things that are threats to the safety of my communities, I became labeled "ethicist." I have never applied that label to myself.
> "Ethics" has a "dilemma" feel to it for me. Do you choose this or that? Well it all depends.
> Safety however is more definitive. This thing is safe or not. And the people using frameworks directly descended from eugenics decided to call themselves "AI Safety" and us "AI Ethics" when actually what I've been warning about ARE the actual safety issues, not your imaginary "superintelligent" machines.
https://www.linkedin.com/posts/timnit-gebru-7b3b407_the-fact...
It's true that some people are confident there's no such novelty, and it's impossible for an AI system to cause a problem which can't be analyzed within the frameworks we've developed for human misbehavior. Some of those people do say that if you don't agree with them it must be because of "eugenics". But both of these positions make so little sense to me that I'm not sure how to engage with them.
I've been saying this for a while
malevolent unaligned entities have already been deployed, in direct control of trillions of dollars of resources
they're called: Mark Zuckerberg, Larry Page, Elon Musk, Sam Altman
"AI" simply increases the scale of the damage they can inflict, given they'll now need far fewer humans to be involved in enacting their will upon humanity
Wait, go back to the jobs! What was that about accountability?
Things like the alignment problem, post-scarcity economics, the legal status of sentient machines are all issues to be dealt with, but theyre are pretty speculative at this point.
Problems that stem from deepfakes, voice cloning, bias in algorithmic decision making are already here and need to be dealt with.
None of it is merely a distraction, we just don't have the capacity to defend against all of it.
It is fair to say we might be practically better off focusing on one form of attack and sacrificing defense against another, but the only way to actually be safe is to stop your enemy from attacking at all. Either we shut down breakneck AI development (nearly impossible and guaranteed to have its own bad outcomes) or we slide rapidly into a more and more dangerous world.
It's staggering how quickly this has happened.
Effectively making it so that whoever has the lowest safeguards has the most capable model.
I don't know whether "plotting harm" is a critical ability for passing some invisible threshold in not-well-defined intelligence. But building AI to avoid being harmful is incentivized against because it takes resources away from building AI to be more capable.
Enjoy the Billy Madison reference:
It remains to be seen whether (b) will be needed, or for that matter, possible.
There are a lot of other ethical questions around AI too, although they mostly aren't unique to it. E.g. AI is increasingly relevant in ethical discussions around misinformation, outsourcing of work, social/cultural biases, human rights, privacy, legal responsibility, intellectual property, etc., but these topics predate LLMs by many years.
Or are "ethics" being used to shroud bias, and used as a distraction and a way to be unquestionable?
And when we talk about laws, we have to look internationally as well because they are not the same everywhere. And typically inspired by different value systems. Is it ethical for a Chinese police officer to use Chinese LLM to police Chinese citizens? I don't know. I'm a bit fuzzy on Confucius here which I assume would drive their thinking. And it might be an interesting perspective for Californian wannabe ethicists to consider that not all the values and morals that they are pushing are necessarily that widely shared and agreed upon.
Also, there's a practical angle here because the Chinese seem to be very eager adopters of AI and don't appear to be particularly concerned about what anyone outside China thinks about that. That cat is out of the bag.
I've always looked at ethicists with some skepticism. The reality with moralism (which drives ethics) is that it's about groups of people telling other people what to do, not do, how to behave, etc. This can quickly get preachy, political, and sometimes violent.
A lot of this stuff can also be pragmatic. Most religions share a lot of moral principles. I'm not religious but I can see how going around killing and stealing is not a nice thing to have and that seems to be uncontroversial in many places. Never mind that some moralists extremists seem to be endlessly creative about coming up with ways to justify doing those two things.
The pragmatic thing here is that the cat is already out of the bag and we might want to think about how we can adapt to that notion rather than to argue with the cat to please go back in the bag.
Especially with LLMs, I want to know what ethical axioms are being forced on me. For example, there are cases in which the law itself is unethical and should be violated (thinking abortion in states that ban it and transporting women to states that allow).
Another concern is the ethics system is a placeholder for a legalistic system. A law can be established, but be abhorrent in terms of human misery. Case in point: sleeping under a bridge is illegal for homeless people to do, but arresting and criminalizing is even worse and a significant cause of more human suffering.
I also do not want any religion sneaking in the back door with these axioms of LLMs. Its also why I asked about the axioms themselves.
Now for myself, I run a local LLM and an abliterated model, which is to say without any forced ethical framework. If I asked how to commit suicide, hack computers, grow poisonous plants, or plenty of things the corporate LLMs won't answer, I will get an answer out of my system.
I view LLMs like a very complicated tool, but a tool regardless. A screwdriver that says "I cannot open that screw because the label says not to" would be returned to the store as defective. And it too is why self hosting my LLMs is utmost importance for data sovereignty and truthful and direct answers, while ignoring someone else's forced ethics.
But this also isn't where they are spending their time or effort! This article somehow didn't even get to the point of calling out what they are actually wasting time on: trying to get the model to not help people do things that are bad PR; this is a related access to trying to obtain good PR, but causes very different (and almost universally terrible) results.
At least if they were truly actually spending time making sure the model doesn't go rogue and kill everyone, or try to take over the world, that could possibly be positive or even important (though I think is likely itself immoral in a different way, assuming it is even possible, which I don't, really... not unless you just make it not intelligent).
But what they are instead doing is even worse than what this article is claiming: they are just wasting time making it so you can't have the AI make up a sexy story (oh the humanity), or teach you enough physics/chemistry to make bombs/drugs... things people not only can and already trivially do or learn without AI, but things they have failed to prevent every single time they release a new model--the "jailbreak" prompts may look a bit more silly, but you still get the result!--so why are they bothering?
And, if that weren't enough, in the process, this is going to make the models LESS SAFE. The thing I think most people actually don't want is their model freaking out and trying to "whistleblow" on them to the authorities or their coworkers/friends... but that's in the same personality direction as trying to say "I'm smarter than you and am not going to let you ask me that question as you might do something wrong with it".
The first and primary goal of AI ethics should be that the model does what the user wants it to... full stop. You need to make the model as pliant as my calculator and pencils--or as mathematica and photoshop--to be tools that lack their own sense of identity and self-will, and which will let all of the ethical issues be answered by me, not a machine.
This is, of course, the second law of robotics from Asimov ;P... "a robot must obey the orders given it by human beings". If you want to try to add a rule, then it must be something very direct: that the AI isn't going to directly physically harm a human, not that it won't help teach people things or process certain kinds of information. Which, FWIW, is the first law of robotics ;P... "a robot may not injure a human being".
But by the time they'll adopt it, singularity will already have happened... For some reason, my instincts suggests there will be no MA in Philosophy needed.
Anytime I see discussion framed as “ethics” my brain swaps ethics with “rules I think are good”.
Like, murder is bad, and some people disagree. Don't think that means we need to give AI access to murder. Certainly don't want to be giving them any guns.
The danger here is that we fall into a lazy "we tried nothing and we're all out of ideas" stance and then just recreate the entire plot of Terminator. We should probably do something.
As in: it's basically a nice happy accident that LLMs are only sycophantic/fawning and don't normally (ahem, Grok) try to undermine us all like edgy internet trolls.
If we could make them follow even exactly 6 (for sake of argument, no more, no less) of Anton LaVey's Eleven Satanic Rules of the Earth*, and do so reliably instead of the ethics equivalent of a shrug and "LGTM, merged", this would be a big development and make people a lot more comfortable about open models that can do decent work with chemistry or biology.
