The Rise of AI

[edyzmedieval]

I'm sure most of you by now recognise the potential impact that Artificial Intelligence will have in our lives in the near to far future. Robots, driverless cars, automated package delivery... everything can be modified. Even software technologies called RPA - Robotic Process Automation - create new opportunities but focus on removing redundancies in business processes. All in all, sounds great, but there's a serious downside and if you know about SkyNet / Black Mirror you can imagine where this is going to go in the future.

I would like to kickstart a conversation on the benefits and perils of AI and obviously the implication of geopolitics on this. What's your own personal take on AI?

Do you think it will go Elon Musk's way - saying we're creating a monster? Or do you think AI is not going to be like that?

Here's a quick report - https://www.digitalhealth.net/2019/0...-intelligence/

Artificial intelligence is “rapidly developing” and a line “needs to be drawn” to establish accountability between clinicians and technology, a report has warned.
Clinicians may find themselves incorrectly trusting decisions made by AI more than they trust their own, The Academy of Medical Royal Colleges report found.
The organisation called for greater recognition of accountability for harm caused by faulty content or by incorrect operation.
“Technology companies are currently focusing on AI that will support clinicians, rather than replace clinical judgement – implying that accountability for mistakes remains with the clinician,” the report said.

Adding to that, here's this - https://www.forbes.com/sites/cogniti...silicon-brain/

The rise of the silicon brain that can give rise to thought, emotion and behavior in a machine seems to be on the way. This is mainly due to rapid advances in software and hardware that are paving the way for next generation computational systems with cognitive abilities modeled after the human brain.

Any thoughts?

[Beskar]

My favourite video on the subject.

[Husar]

My favourite video on the subject:

[drone]

"Thou shalt not make a machine in the likeness of a human mind."

[Beskar]

My favourite video on the subject

Yeah, I have seen that before, but it does have a great conclusion. "There is a point this machine becomes extremely dangerous and that is when you turn it on."

[edyzmedieval]

Reminds me of the Facebook algorithm / AI machine that decided to learn a new language by itself to speed up internal processes. While that isn't as scary as it sounds right now, it's a first step.

[Brandy Blue]

Reminds me of the Facebook algorithm / AI machine that decided to learn a new language by itself to speed up internal processes. While that isn't as scary as it sounds right now, it's a first step.

If I understand correctly, it was nothing like that. It was more like two parrots trying to speak English with each other. Even at the start, they didn't understand themselves or each other, and they degenerated further into gibberish the further they got from the models of correct English they were originally imitating. Facebook didn't shut them down because they were dangerous, but because they'd reached a point where they were no longer useful for reseach.

https://towardsdatascience.com/the-t...e-37c5d680e5a7

[a completely inoffensive name]

Deus Ex level AI is still decades away.

hard to build consciousness when we cant explain it within ourselves. The public (and I mean the general public) perceptions of what makes an AI "intelligent" are based on a faulty assumption that ties conversation = intelligence.
We will find ourselves with AI that can have a free flowing conversation with us to the degree we cant tell whether a human is giving us responses or not. It will respond to new scenarios based on the accumulated knowledge of millions of scenarios fed to it. But truly novel thinking which is really what we are talking about when we talk about AI, is still a mystery.

[Montmorency]

Deus Ex level AI is still decades away.

hard to build consciousness when we cant explain it within ourselves. The public (and I mean the general public) perceptions of what makes an AI "intelligent" are based on a faulty assumption that ties conversation = intelligence.
We will find ourselves with AI that can have a free flowing conversation with us to the degree we cant tell whether a human is giving us responses or not. It will respond to new scenarios based on the accumulated knowledge of millions of scenarios fed to it. But truly novel thinking which is really what we are talking about when we talk about AI, is still a mystery.

Second and fourth sentences contradict. Language has rules and words, but it does not arise out of a mere accumulation of rules and words. Google Translate will give you an inhuman result no matter how many words or rules or even single states you feed it. It's not a matter of processing power.

Why is it plausible to maintain that assembling a thousand monkeys (or less charitably, nematodes) will by their accumulated mass meet the measure of Man? I'm profoundly skeptical that today's classes of machine AI, in any degree of sophistication and co-articulation, could ever match the "intelligence" of a whole specimen of even the simplest mammal. I'd be inclined to change my mind if you could produce an AI trained to land space rockets that could learn how to determine edibility of substances, maximize score in videogames, and file a tax return, without being designed to do so. Do we have any more reason to believe this is possible than to believe that a mouse can "learn" its way into being a dolphin?

In goofier terms, I still don't understand why this movie is considered by so many a plausible expectation for machine AI:

Spoiler Alert, click show to read: 

It should be easier to intervene in a pre-existing complex biological system and produce unintended results than to manufacture one from scratch or to imagine one will mysteriously "emerge" on its own from something different.


That's not to say that even this weak AI can't be dangerous! If you have AI designed to land commercial planes without human supervisors or designate targets for autonomous weapons, and it arrives at a weird process for doing so without engineers catching on, you could clearly have thousands of fatalities on your hands. But this would only be a consequence of us as designers foolishly giving dumb AI more to bite than it can chew...

[Husar]

Why is it plausible to maintain that assembling a thousand monkeys (or less charitably, nematodes) will by their accumulated mass meet the measure of Man? I'm profoundly skeptical that today's classes of machine AI, in any degree of sophistication and co-articulation, could ever match the "intelligence" of a whole specimen of even the simplest mammal. I'd be inclined to change my mind if you could produce an AI trained to land space rockets that could learn how to determine edibility of substances, maximize score in videogames, and file a tax return, without being designed to do so.

This from the guy whose blind brain theory made a lot of men angry about how it made them look more like arrogant machines than "men".

We already know that the brain has specialized areas. Some areas take care of vision, others of hearing, then there's an area for speech, some for emotions and so on. At some point we may be able to make artificial neural networks for these things, to some extent we already have them. Of course as you say, having all the parts is not enough, neurons can change their connections and work together, but it's not entirely unthinkable that we may understand these processes one day or that we may simply find other solutions that work to make the parts work together as required.
If a human brain is "nothing" but an electrochemical machine, then it should be possible to emulate it or create something similar. That's also where the danger may come in, if you emulate all the "functional" parts, but not the emotions and the social parts for example. Then you may end up with something intelligent that doesn't really care about what it does, as long as it achieves whatever goal it has.

How long it will take until we get there, I cannot say. I too find the current implementations of "AI" not really worth of that name, they're very specialized neural networks at best, from what I have seen. That's not to say they can't be useful though.

[Montmorency]

We already know that the brain has specialized areas. Some areas take care of vision, others of hearing, then there's an area for speech, some for emotions and so on.

First, rather than saying the brain has specialized areas it is better to say that certain areas are more necessary or contributive toward "functions" than others. Brains show some modularity, but they also show distribution.

At some point we may be able to make artificial neural networks for these things, to some extent we already have them. Of course as you say, having all the parts is not enough, neurons can change their connections and work together, but it's not entirely unthinkable that we may understand these processes one day or that we may simply find other solutions that work to make the parts work together as required.
If a human brain is "nothing" but an electrochemical machine, then it should be possible to emulate it or create something similar. That's also where the danger may come in, if you emulate all the "functional" parts, but not the emotions and the social parts for example. Then you may end up with something intelligent that doesn't really care about what it does, as long as it achieves whatever goal it has.

How long it will take until we get there, I cannot say. I too find the current implementations of "AI" not really worth of that name, they're very specialized neural networks at best, from what I have seen. That's not to say they can't be useful though.

The whole point is you're making a category error. Can we, from raw genetic code, incarnate a human-like creature with human-like (or greater) intelligence? I suspect we can't. Insofar as we can't, or can't do fabricate any organism above the level of the simplest unicellular proto-life, why would we imagine doing so is available with hardware and software that have many different limitations and principles of operation than biology, where we don't even have examples of machine "intelligence" like we do with biological?

Brains are not machines, though they share the same underlying physical nature. They have in-built "selected" learning capacities; even if this were theoretically available in computational context, how would machines receive the capacity to evolve them? (Finding more efficient ways to perform the same task is not evolution.) They have no natural environment, only what their creators give them. It is physically impossible for creators to match the creative potential of all Creation. It would be like trying to build an entire planet by hand. The only plausible approach for building a new planet is to somehow gather bazillions of tons of raw matter and knock them all together, hoping for gravity to do its work. In other words, enable complexity to arise as an emergent property from more parts than can be counted or programmed.

How would this be applied in the AI context? In any case we're talking about processes that operated over billions of years on scales beyond comprehension. Basically you need to develop a computer or network a million times more powerful than anything in existence, give it mobility, sensors, and the ability to learn without being taught how to learn or what to learn, and hope it does something rather than nothing...

This from the guy whose blind brain theory [Note that it isn't my theory] made a lot of men angry about how it made them look more like arrogant machines than "men".

I've tried to avoid it, and the discipline of neuroscience for 3 years so far. Muddling in human politics is much more comfortable.

This line of thinking actually supports the idea that the greater prospect (or threat) in the conceivable long-term is biological intelligence. Before we even approach the question of new life, who knows what targeted manipulations we may devise for our own brains, without ever knowing precisely what we are accomplishing or how? Consider how much brute head trauma can do to a person's behavior and psychology - then consider how, for example, optogenetic control permits indirect causation over simple behavioral patterns in simple organisms (so far). Conceivably, it is within the same spectrum of ability to move to directly 'piloting' organisms, and then to modulating their total physiology. Ethical mores against human experimentation or modification won't last forever...

http://www.jneurosci.org/content/38/4/784

This makes the philosophical debate about free will and determinism in a state of underdetermination by current neuroscientific findings. Consequently, and referring to the quotation that started this essay, we might conclude by saying that: Neuroscience may in no way interfere with our first-person experience of the will, it can in the end only describe it … it leaves everything as it is.

That may change. Imagine an applied neural interface that could target your feeling of volitional control over your thoughts and actions. Changing nothing else* but erasing your sense of self and agency. Imagine walking around and doing all the things Husar does every day - but no phenomenology of being Husar, no "qualia". This would effectively demonstrate that humans are ultimately approximate to philosophical zombies. To say nothing of other more practical sci-fi tropes... I maintain all this is more plausible than ISS Alexa refusing to open the pod bay doors (in a way that can't be fixed with some routine maintenance). We have all the matter we need to work with, rather than trying to recreate it.

Spoiler Alert, click show to read: 

*I don't think it could plausibly get that precise in application; there will surely be fractal layers of unintended and unexpected consequences to tinkering with our intracranial galaxies. Maybe losing the sense of intentionality would be a byproduct of trying to reduce symptoms of depression or something. Who knows what is dissociable and under what conditions. How many secondary effects there already are in pharmacology. There would be more precision and greater effect, but also more uncertainty, in going straight to the source. What I'm talking could make swallowing a pill comparable to bashing your head against a wall until you're a math whiz.

Now please accept everything I said uncritically and let me return to not thinking about it.

[Husar]

Now please accept everything I said uncritically and let me return to not thinking about it.

I don't. Only two short remarks:

1. I don't think the computation power is the big issue, NVidia's tensor cores already have plenty of that for neural networs, enough to calculate 8 million pixels for upscaling in a matter of milliseconds. The rest would be finding out how the human parts work first. The reason we can't model some functions of our brain is that we don't even know how exactly they work in our brain. If I show you a black box and tell you to model the mechanism inside without opening it, you would surely have a hard time doing it. And yes, I do think our brains are machines.

2. I don't even see why we would really want to model our brains as 1:1 copies. Getting somewhat close or keeping things specialized might actually be a better and more useful idea. If we model something super intelligent, that can calculate, react, sense and what not much better and faster than we can and is just as greedy and competitive as we are, well, did you ever watch Terminator? And no, I don't think our rebel bunch would stand much of a chance in reality, the machines in the movie are unrealistic. The promise of AI is probably more in certain selective and specialized applications. Having an actually conscious AI would be very...strange at the very least.

[lars573]

Deus Ex level AI is still decades away.

hard to build consciousness when we cant explain it within ourselves. The public (and I mean the general public) perceptions of what makes an AI "intelligent" are based on a faulty assumption that ties conversation = intelligence.
We will find ourselves with AI that can have a free flowing conversation with us to the degree we cant tell whether a human is giving us responses or not. It will respond to new scenarios based on the accumulated knowledge of millions of scenarios fed to it. But truly novel thinking which is really what we are talking about when we talk about AI, is still a mystery.

TLDR: We'll be able to build a machine spirit long before an Abominable Intelligence.

[Seamus Fermanagh]

This fellow gave a talk to our faculty on the media/public perception of the dangers of AI. They've done some interesting research on the assessment of the nature of the threat by experts in the field in question versus the general population.