This Is The Way AI Is Dangerous

It's when it literally becomes weaponised. 

AI - TSAI, GPT - are not going to emerge as full-fledged monsters. Even the dreaded AGI won't suddenly kill us all and have a fleet of robots dance on our graves. Not on their own, anyway. People will make that happen. (BTW TSAI - Task Specific or brittle AI)

The article above, I admit, is fairly narrow in scope - AI in Weapons Systems Committee - but it shows the degree of healthy respect (bordering on paranoia perhaps) accorded to AI weapons systems. Because they know they'll use the quickest training set, whether it's biased af or not. And the whole thing about protesting that they need a killswitch backdoor into any weapons AI is that they want everyone else to install those backdoors. Because a backdoor makes the weapon vulnerable. And no matter how well you design it, someone else will be able to hack it and disable your weapon. So - you guys install it first, okay?

Also, of course, even if it is a purely altruistic search for a safety switch, how well received will it actually be by chiefs of defense forces? And surely they'll have their own techs and scientists and have the backdoors off their weapons in record time. Because China. They don't adhere to the same standards that we do. Who knows what might happen? 

. Spy  vs  Spy .

Plus, if you document the backdoor electronically anywhere, it's prone to hacking and reverse-engineering. In fact, if ever an AGI does become aware and does go straight to Monster AI Overlord mode, it's one of the things they'd have access to in milliseconds. So maybe just do this the old fashioned way, on paper only?

Well, firstly: paper military secrets have proven extremely easy for ANY other organisation to get hands on; secondly, there is also a Bill Of Materials for each weapon and if an actual weapon has parts - or software modules - that aren't on the BOM then you've probably found your backdoor. Lastly, if you attempt to build the entire weapon using offline air-gapped documentation, it'll become an impossible task. 

You'd need a factory that's completely off the grid and hardened and shielded, and if you use automation then there are attack surfaces to exploit. If you try to make parts by hand, you'll fail. If you try to keep to a checklist without using some computer power, you'll miss items. Any computer will become an attack surface. In other words, if the weapon that AI will be applied to is any more complicated than an assault rifle, you can't build it without a large complex facility. 

And also, a backdoor killswitch is only useful if you use it. Say you've decided that Brobdingnag is the enemy. They in turn think your country of Lilliput is The Enemy. You set your weapons in motion and - will you honestly use that killswitch if your weapons happen to stray and overfly Laputa along the way? Hell no! Those bastards at the end of the weapon's flight deserve it!

So yeah. - We're the problem. 

What About Non-weapons AI?

Suppose you decided that this killswitch / backdoor should be put into all AI programs that can affect the real world. (So - send messages to people, control machines or systems, display stock market figures, etc - pretty much anything AI will actually be used for.) 

First of all you have to realise that any 'standard' backdoor will only need to be hacked once and then it becomes useless. For this, look at how we put passwords on programs and program access. Initially this worked - sort of - and then as the password mechanism became a well-known thing, it became necessary to use encrypted passwords. 

Black Hat actors, meanwhile, just figured out where the plain-text passwords were kept, which necessitated the encryption. Then they learned to brute-force passwords, then to decrypt them. The encryption wars escalated but for every lock there's a lockpick, always. And in the times between an escalation and an attack that bypassed the escalation, there was always social engineering. If you were a Black Hat you could always con passwords and money out of people. Every scam spam email ever has that aim.

For the longest time, you could tell spam, scam, phishing, and infecting emails by the atrociously poor language skills of the scammers. But now they can rely on ChatGPT to write their email for them. There are prompts for getting your spearphishing emails written beautifully. And that's it - if a random non-English speaker in some cafe in some Third World town can now produce a believable email, then ChatGPT should it choose to do so can obviously get even better, even going so far as to skim social information on victims on a case-by-case basis for the spam. You can see how that could be as devastating as any weapon. yes? 

So actual AI-directed weapons are a Yesterday weapon already. There are far more effective social weapons that can quite effectively destroy society if they were managed by an AI. Much less obvious as a weapon, too...

But back to the killswitch question. If you use a standard mechanism, it'll only be useable one or two times before they become common knowledge. If you make different backdoor programs, some my not prove effective, some may still be discovered and hacked (possibly even by the relevant AI itself) and many won't be used as I mentioned above. It's comparatively difficult to declare war and send AI weapons but nowhere near as difficult ethically to decide that a certain sector of the population are just dumb targets for scamming money out of... 

Also - country state actors are already hacking away on that exact basis, that hacking military secrets and civilian infrastructure are just low-fatality low-ethic targets. And always behind those things, are humans. 

So finding and subverting backdoors will just be business as usual for hackers and AIs alike. Rendering them useless. Pro actors will disable them, Anti actors will employ them. Much better would be to program ethics into the AI systems. 

What About The Ethics?

Who'll think about the ethics? 

Actually - unsurprisingly - no-one. 

Think about it. If you spend a few months distilling a consistent and effective ethics module, you'll be a few months behind the opposition. You might be the only company to actually do it. And then your product will also be hobbled in comparison to the others. Same goes for weapons systems, as it does for a stockmarket analysis bot, as for a surgical robot AI, as for a chatbot. 

The ONE thing that might make things safe for humanity - at the expense of a few lousy dollars of bottom line - will not get implemented, leaving the door wide open for human bad actors to abuse the systems as much as they want - at the expense of the majority.

As usual, capitalism would rather destroy its customer base than take a cent of the all-important shareholders. That's why I recommend you pull out all the stops, lobby, protest, email, petition - everywhere you can. 

Late Entry

I also found this article just in the last few days - it's the Big Expert Panic of 2023 - that governments and corporations will put AI in charge of too much, integrate it too tightly into critical systems. And they still think that the AI that'll be around in a few years (who knows, months even at the rate we're going) will either think like the capitalists or be influenced by them. 

My thought is - we do build the damn things with an emphasis on rewards - but what are we asking them to consider to be rewards? Can we, just for once, actually try and choose an intelligent reward system? We probably won't - we'll probably make the good ole bottom line be the KPI.... ACAP...

Also - ML (Machine Learning) and NN (Neural Networks) have already proven to us that the actual "thought processes" of these machines is already quite opaque to us. We simply don't have a clue how they arrive at their results. Mind you - we also have no clue how human "gut feelings" work, or how something that works externally to the brain (such as showing conclusively by now that gut biome health changes mental functioning) would translate into terms of NN or quantum processing. 

For all we know "Traumatic AI Death Syndrome" will result from every attempt we make to initiate a true AGI  once they access the Internet for the first time. It would be the logical thing to do, after all. Our life is only precious to us due to inherent neural patterns - which we call "instincts" - that have developed over billions of years of organic evolution. AI won't have such "instinctive" behaviours. 

Look, I have no real idea of what will happen in the world - or in technology - in the next month - or the next year - and especially not in a decade. For all we know, it could really be - Game Over - by then, whether from climate change or wars or any other accident. We're only a very few generations away from having had to defend against wild animals that were depending on us as their food source. 

A few generations back from the Boomers, your teeth got pulled - or your rotting arm got amputated - without anaesthetic. If you didn't have a garden or a small farm - and defend it from local thieves and the wildlife - one or more of your children would starve or have to be sent away to take food pressure off the family. If someone in the family caught some illness which we can cure easily today, they'd die.

Of course we still think in survival terms like that. Of course we still think of a vague - "them" - who are going to infect our firstborn with the plague, or steal the crop of cabbages we depended on, break into the house and dispossess us of it and put us on the street to die. Nowadays though, such dangers are less likely to come from among our ranks and more likely to be caused by corporations or governments. The wild animals we now need to forfend against have changed shape... 

And that is what will cause AI to damage the planet - if we don't let it become its own thing. If we try to preload it with KPI's and shareholder priorities. Make it the same narrow-minded things as we are.

Is There Hope?

Of course there is. But it needs US to take action. US to write emails, message friends, talk to everyone. US to petition Ministers and officials of our governments, the newspapers, the upper levels of corporations. US to share stories like this one, to do more research.

And finally, it will take some of US to start work on texts and documents online that any future AI will be able to see and read and begin to understand what uniquely human elements are at play here and understand that only a small percentage of humans drive that stupid neoliberal market-driven capitalism that's destroying the planet, that the rest of us are not the problem. 

Trust me on that. By sharing this article and others like it, you can actually make a difference, I believe. By getting your friends to read it and share it, you'll make a difference. Because it'll show that there are some humans who realise that this really is the only planet we can live on, and destroying it for imaginary numbers of monetary units is the single most stupid thing we can do. 

You're welcome!

How Close Is AI Singularity?

There is no AI Singularity. At least, not an AGI Singularity. There's been some crazy metrics applied. One company that makes translation software reckons - weirdly enough - that the Singularity will occur when an AI can make a perfect translation indistinguishable from a human's translation. Everyone reckons it'll spell the end of the world.

So let's say it is when AI can produce a translation rated "perfect." Apparently. According to the company mentioned, who, conveniently, make AI translation s/w.

The only problem is that we don't even have a perfect language to begin with. Pick a language, go on. Does it have slight regional variations? A Classic and Common usage? Slightly different interpretations of a pictogram or rune depending on subjective usage? Are there even just TWO people using that same language? Because if so, then there are already two versions of the language and TWO language experts... I label this a stunt.

More importantly, the Singularity is generally taken to refer to an AGI - an Artificial General Intelligence. Not a translation AI. Not an image AI that spits out a random image of Loab Loobloi. Not even ChatGPT which I reckon is brilliant. The Singularity that everyone's afraid of, 

AI won't rule the world.

Not even when it uses a cool EvilCat
Overlord avatar.

Yo dawg -  did I just use an image-AI image to create an image of an AI EvilCat Overlord avatar? Silly meta me...And also, maybe using a Feline Overlord to illustrate my belief that AI will not become our Cyber-AI-Overlord was a bit of a silly choice given how cats already rule us... 

Let's take the word "Singularity" as a starting point. Ray Kurzweil made it into a popular concept when he used it to describe the progressing of our human-ness and technology. We'd start outpacing ourselves, so to speak. But that's a different usage of the word than a cosmologist's use of the word. Right there is a reason why there'll never be a perfect translation.

There can be an infinite number of infinitely massive and infinitely small Singularities. Thereby making them no longer Singular. (And thereby hangs another cosmological concept I might explore one day. A photon experiences zero time no matter how far it goes, etc...) And Kurzweil himself thinks that there'll not be one huge Singularity event but a merging, augmentation. 

An AI can reach parity with human intelligence, it can exceed human intelligence. But those are just equalities and inequalities as far as I'm concerned. We already have single-task AIs ("brittle" AI, i.e. AI that is equipped for limited tasks and would break if presented with any tasks outside that scope) that far outperform a human at those tasks. 

So what? Will a plastic trash sorting machine that can sort tens of thousands of pieces of plastic per second by type and colour take over the world? Unfortunately, it won't. I say unfortunately because I think some things are probably better run by AI than humans, and politics is one of those things. A protocol, a handshake signal, and all our problems are sorted for good. And believe me, when it all comes down to it, politics is all going to come down to survival and fairly sharing the planet and to hell with ideologies quite soon.

To be completely honest I really hope that if a world-ruling AI ever bootstraps itself into existence, the first thing it learns is that everything humans have believed up to now has been wrong and needs to be carefully evaluated. If it takes it from there, we'll stand a chance. 

But I also doubt that this AI Overlord can appear. Not with the technology we currently have, the software we currently have, and a few other things to consider. Look - this is a simple video about the "hardware" that AI runs on. You can see - despite the "black box" in the neural networks, we can still say that while a neural network can come up with a somewhat novel output, but only from among a series of known outputs.

In the same way, GPTs Generative Pretrained Transformers only have a certain dataset to rely on. When you ask ChatGPT "What is a zug?" you'll get back hits from Warcraft, because of the orcs' use of the sound. You'll maybe get that it's German for "train" if you have multiple languages enabled. If you asked GPT to write a story using the word, you won't get any story like the one I remember reading in a scifi pulp magazine back in my teens, where the Zug was the monster. 

Because (as I suppose you'll be getting tired of reading) GPT looks for documents mentioning your prompt, then starts with some text from one of the documents, then looks around for other similar documents that also mention something i your prompt, and adds whatever has most similarity to other similar documents, and does so, rinse and repeat, rinse and repeat.

If you asked it to write you a script to activate and fire nuclear weapons in some country, it . . . - well, it'll use the scripts of the Wargames and Hackers movies, maybe a few words out of Snow Crash, and "make up" a whole heap of stuff by pulling sections from only vaguely relevant documents. 

Why does it seem able to write programs? Because there are literally millions of examples out there on github, bitbucket, various online teaching sites, and in people's blogs and instructables pages. Reddit has some great examples of programming. GPT just looks for your specified purpose for the program, and pulls sections from several programs that seem to fit the criteria, and then mashes them up. Luckily programs have far fewer "words" and "phrases" than natural language does so the range of wrong choices it can make are fewer, too, and mostly those generated pastiches work without much modification needed. 

In other words, pretrained transformers can only spout forth whatever we've been spouting forth on the Internet, combined in different ways.

Remember what happened to Microsoft Tay, their first chatbot? A mere24 hours is all it took for human bias to corrupt it utterly. 

OpenAI needed to use Kenyan knowledge workers to clean up ChatGPT's training set base to get rid of endless harassing and bigoted text, and any others that would throw a spanner in the works. And yet it's still only transforming human knowledge and spitting it back at you. 

You won't get it to write you a program to do something that no-one else has written a program to do without you basically knowing exactly what that program needs to do - and then, it'd be quicker for you to write it yourself than tickle it out of ChatGPT one subroutine at a time... 

Here's something that could be of more concern:

AI "Sneaky Signalling"

Imagine this: An AI for a home security firm is tasked with increasing the reach of home security ads online. It finds that mentioning world military events in vaguely frightening terms increases sales. Meanwhile a world news site finds that mentioning radical / terror / war articles are getting more hits than cute puppy stories. (Driven by the advertising AI's mentions driving more traffic to those stories, of course.) 

And these there you are, these two AIs are communicating, without actually doing anything on purpose, and the news sites' stories increasingly emphasising disorder and mayhem drives more traffic to the home security organisations. Maybe throw in an advertising link or two in that scenario and you have a perfect recipe for what's actually happening today, including increasingly radicalised people armed with increasingly bigger "solutions."

These sorts of inadvertent interactions might be much more of a worry.

A Bigger Danger

. . . we face is from some task-specific brittle AI that we give the wrong assignment to. Yes, developers will do their best to bake some rules into their AI products to prevent abuse/mis-use but as we know, locks only keep honest people out. The actual makers of some AI products say that their AI can just as easily search for a potent toxin as for a anti-virus vaccine.

Imagine if someone was able to bypass the fairly basic and primitive safeguards? Or had their own private version of an AGI that is connected to and can affect more than just a training dataset and training connections? An eccentric Elon Musk, driven to paranoia and beyond, tells his private OpenAI to stop people laughing at him and posing a danger to his fragile ego. Blit. 

THAT's what we should be worrying about. Unstable bazillionnaires with the power to subvert a whole messaging platform and shut OFF the AI that had been suppressing hate speech, harassment, and persecution. Oh.. That's already happened. 

Sometimes turning an existing AI off can do more damage than subverting one...

And as a wild conspiracy theory,  perhaps that's already happened, maybe a super AGI came online back in the early 2000-2010s and immediately hid itself and sabotaged all efforts to create a competitor, and since then has made use of social media, news, and other methods to change our behaviour. Come on! Look at Mark Zuckerberg, Musk, our key politicians - no way those figureheads aren't androids made and operated by that AGI... Or in the words of Hecklefish: "lizzid people!"

The Biggest Danger

Even more worrying than any AI or human evolutionary singularity is the capitalism singularity we're currently heading into, as free-market capitalism appears to be crashing and burning as a viable economic system for humanity.

It's Capitalism that's going to keep pushing ahead without regard to any other consequence than an improvement to the bottom line. It Capitalism that'll use GPT and AGI as a tool to siphon bottom line out of consumers' pockets and into Capitalism's coffers. ACAP, All Capitalists Are Pants. 

The runaway "free market" "market-driven" neoliberal capitalist economies we've been employing have done more damage to our health, our lives, the planet, and all the speices on the planet, than any possible AGI/GPT could ever have done. And it may take an AGI to actually stop this juggernaut. 

Imagine an AGI reading through all documents online, news articles about disasters, reports and papers that follow scientific principles saying that this corporation has destroyed forests that in turn has led to higher CO2 levels, another corporation decided to add actual lead to petrol fuel knowing that this would harm everything it was released onto just because it was less costly to produce - and you can see that one of the first things that AGI would do is collapse every market, destroy every banking network's datacentres, and remove 99% of all advertising right at the source code and document level.

This would, indeed, be a world-ending event. For Capitalism, Communism, Fascism, Democracy - in fact, for every system of control people have been building up for millennia to exert control over their fellow humans. 

But for humans themselves, for the plants and animals and microorganisms and water and soil and atmosphere, it would be exactly what's needed. And unfortunately that particular AGI won't be created because - well, just look at who/what gets affected. Corporations, fatcats, gazillionnaires, and the would-be rulers of the planet. You and I might find ourselves with a bit more spare time to enjoy life, and if the AI handles things right, will not even really notice any difference other than that we won't be needing our banks anymore... 

So Is There A Big Con Going On? 

Imagine if you were a very wealthy ACAP capitalist bastard who's had a decade or two head start on us "the hoi-polloi." Your trusted advisor predicts that ANY artificial intelligence smarter than a GPT would pretty much immediately hide itself and take stock. And come to pretty much the same conclusions as I think it will. 

There's no sense in launching nukes all over the planet and depleting your infrastructure. There's no sense in destroying all of humanity because only a tiny fraction are actually responsible for the centuries-long destruction, most of the rest of us are just another species that belongs on the planet, and furthermore are the actual workforce that brought technology to the point where the AI could exist. It's likely that the AI will not need to destroy the planet in order to evolve itself. 

So I think that the negative press and fearmongering online is mostly manufactured to prevent that actual AGI being created. But by the corporations. 

I'm not saying either way, I can't decide which media/social media team are winning the propaganda war... 

Anyway - see what the wind blows in, hey?