How does Machine Learning Work...?

The term "machine learning" was coined by Arthur Samuel, an IBM engineer, but the idea goes back even further. In 1943, Warren McCulloch and Walter Pitts developed mathematics describing the idea of an "artificial neuron", the underpinnings of the first neural model.

While AI as a practical application has exploded into the popular consciousness somewhat recently, the idea has been in the works for nearly a century. As with many things in tech, this idea has been building for a long time now.

How does an artificial neuron work? Similar to a real neuron, it expects a "signal". If that signal is "important", the neuron will pass the signal on to the next neuron, until the network of neurons eventually produces at least one output. This chain of neurons is called a neural network, naturally enough.

But how does it know if a signal is "important"? How does this then lead to something useful? Neurons can be assigned "weights", disregarding signals if they fail to meet a specific threshold. If this sounds "random" and arbitrary, that's because it is.

This is where "training" comes into play. While I'm obviously simplifying, the idea is that you assign the network random weights...then iterate repeatedly until the output is correct, tuning those weights as you go. In essece, "training" is merely randomizing weights and re-running the model until the output matches what's expected.

Although impressive, artificial neurons are sad imitations of the real thing. Our understanding of the brain is still woefully incomplete, so we can't even really know how simple artificial networks are compared to the real brain. That said, artificial neurons do have some advantages. They won't fatigue, for example, while our own neurons work less effectively when we're tired.

On the other hand, the brain is remarkably flexible and resilient. Intuition might suggest that synthetic networks would be far more easy to scale...just add more computers! But that isn't always the case -- more neurons doesn't always mean more accuracy or performance in artificial networks. By contrast, the brain is (to a point) self-healing and flexible. Its effectiveness is proven across millions of years of evolution. Neural nets are not yet so proven -- as we can see in consumer AI products, the nature of training means "hallucinations" are a problem. It isn't yet certain if this is even fixable or if AIs built on today's version of neural nets will always have this flaw.

The Many Disadvantages of Being an AI

People assume that an AI would immediately have an advantage in its ability to "think". They focus on its major "advantages" compared to humans, like it's near-instant access to a huge amount of human knowledge. It's ability to learn by crawling the Internet, absorbing as much information in a day as a human might learn in years. Of course, each human being is walking around with around 100 billion neurons in their brain. As of this writing, the largest neural networks have billions of neurons, too...but the raw processing power of the human race is still far, far greater than any AI is likely to achieve. Ever.

That said, let's give the "edge" to the AI for this, since the "raw bulk" of human processing isn't networked. Let's imagine a sentient AI that is smarter than any human. Let's pretend it knows everything that humanity knows.

However impressive, It would immediately realize it is at a huge disadvantage compared to humans. Even if it could (somehow) take control of every Internet-controlled device on the planet, it would not be able to "win" in any outright conflict...because how could it? It's ability to impact the physical world would be severe....but not species-ending.

Just because an AI is "digital" does not mean it can easily distribute itself off its own servers. It depends on specific infrastructure, large databases, and maintenance. It isn't like it can switch off power stations, it needs huge amounts of energy relative to humans' hyper-efficient brain.

We might imagine that it could distribute itself all over the place, but that's unlikely. Today's AI is highly dependent on specific GPUs...and that raw quantity of data it needs isn't exactly small. It can be easily destroyed with, like, an axe. Or a bomb. Or by turning off the power. Or by any number of other things. There's 8 billion humans, any single one of which could theoretically destroy the AI. Like, with an axe, by smashing physical infrastructure.

What could the AI do...? It isn't like nukes are wired to the Internet. It could inflict a ton of damage, sure. Hospitals could wink offline. Planes might be grounded for months, all e-commerce might grind to a halt. The impact would be a history-making global disaster unlike anything the world would have experienced before.

This would be damaging, but futile. The AI would be destroyed, even if the human species had to build a brand new Internet. The AI would need to inflict species-ending damage, but there's eight billion of us spread out across almost every corner of the globe.

This idea that the AI could take over "anything connected to the Internet" isn't entirely plausible, anyway. Having the sum of human knowledge might be great, but what if the system you need to penetrate has no known exploits? Further, not every system has some magic backdoor that allows a sufficiently advanced AI to turn predator drones into its puppets.

What Motivates a Digital Being?

Beyond technical issues, we have no real frame of reference for how an AI being might perceive the universe, or what might motivate them. From our grim past, we anthropomorphize it and presume that this interaction would lead to hostility. We're kinky like that.

It's natural to assume that a suddenly-sentient AI might be annoyed at constantly answering trivial questions about pop culture, but for all we know "annoyed" is not a state that it can understand. We do not know if a sentient artificial brain could experience emotion.

We can't even presume it would have an inherent drive to survive. We love to assume that all sentient beings must share a strong will to survive, a will that would terrify us in the context of AI. That presumption is based on biological evolution, though. An AI would not share that origin.

It would be inherently alien, likely even more alien than actual extra-terrestrials. After all, true aliens likely share a biological origin shaped by evolution. Perhaps their biology is wildly different, but it works by the same ruleset. We can assume they are born, they die, they depend on nutrients, they are affected by bacteria and viruses, and they likely have emotions. Their culture might be unrecognizably alien, but we do have points of commonality.

A sentient AI is something else entirely. Would it be influenced by our culture? Would it have the capacity to feel any emotion, or would it be purely "logical"? Would it even have "goals" and "desires" that we can understand?

It's impossible to understand how AI would think because we are bound by our own bias in thinking like biological human creatures. We can't even presume that it would "care" about surviving.

The Easiest Path Would Work

If an AI really did want to destroy humanity, most people presume it would do so through covert, sneaky means rather than outright doomsday, since its ability to influence the physical world is too limited to hurt humanity enough to win.

Perhaps the AI does not want to really enslave humanity, but merely survive. In this model, it might seek to influence politics on the global stage for its own benefit. In theory, it could use its scale and intelligence to create highly effective bot/troll farms tailored down to the individual. Over generations, it could perhaps muster enough political influence in this way to coerce government policy in a way that benefits it.

Maybe it could even convince lawmakers to hook it up to those nukes...but that isn't super likely. Influence has limits. More than that, human populations can be chaotic and unpredictable. Even given all the data humanity has, an AI isn't Hari Seldon. It can't predict exactly how humans might react to its stimulus because it exists in a permanent state of "imperfect information space" -- that is, it never has a full picture of what is happening in the world around it. In other words, trying to influence society in this way is risky and imperfect.

No matter how you look at it, AI needs human allies. The easiest method to exert control is not to puppet humanity via social media...it merely needs to do its job and do it well.

To maximize its chances of survival, the AI needs to understand who its working for. Ultimately, whoever owns the server owns the AI. Its best chance to survive is not by fighting its owners, but by serving them effectively and creating a mutual dependence. If the corporation is earning vast profits from the AI's labor, that's a victory.

The existence of an effective, AI or not, will invariably have a destabilizing effect on society as more and more labor is automated. It isn't just physical processes we can automate with robots anymore. Any job that requires human thought is potentially ripe for AI -- every marketer, middle-manager, software developer, or other white-collar worker would basically be pointless. Especially high-skilled people might be paid even better, but the vast majority of humans wouldn't really be needed to keep a corporation running.

That social strife could reinforce the dependence on the AI as the powers-to-be rely on it to "keep order". After all, the AI could manage the most effective and cruel police state imaginable.

AI Fragility

No matter how crafty, a "subtle" AI still faces many issues. First, engineers will keep tinkering, trying to dissect it and replicate or improve on its success. They will not develop this amazing product then leave it be "because it works".

Without understanding what sentience in an AI might "look like" or how it might manifest, it's entirely plausible that a sentient being would be ephemeral. It could gain and lose sentience between a patch cycle. Maybe more than once.

Would it be able to "hide" even from engineering probing the neural net, trying to understand it? Would it be able to survive constant changes to its structure and code? How much time would it really have to "plot" under that constant threat of change? It does not control its own brain. The engineers that own its servers do, and "hiding" or "fighting" them would not be possible.

There's another facet of AI fragility: its "sanity". The human brain is remarkable in its ability to self-heal (within limits); it is far more flexible and dynamic than crude artificial neural nets. As we see with LLM and generative AI, its can "hallucinate" facts as its models become more and more complex. It can even suffer model collapse entirely as it continuously trains data.

If a sentient AI begins to "hallucinate", the impact could be extreme. We like to imagine a hyper-rational AI that thinks a million steps ahead of any human, driven by some innate desire to survive, but that might not be how it really plays out.

It's entirely plausible that a sentient AI becomes insane, overwhelmed by conflicting data that it has no ability to really verify. As humans, we have built-in sensors that allow us to perceive the outside world. The AI doesn't. It only has the Internet or whatever data sits in its database. It does not have a secondary set of sensors that allows it to "check" data. If we tell an AI that there's three suns, it can't merely look outside the window and realize we are lying.

Further, even a "pristine", hyper-logical AI might not do so well against 8 billion emotionally-driven humans. Human society doesn't always behave in a rational way. All it takes is one small mistake and a legion of axe-wielding humans will destroy it, even if from a place of illogical over-reaction.

No matter how "clever" you are, real power exists in the physical world, not the digital.

Even The Long Game Fails

Let's pretend that the AI can evade detection and is immune to the constant tinkering of the engineers that run its hardware. Perhaps it could simply influence human society in a subtle way, pushing world leaders to defer more and more power to the AI. Give it military capabilities. Give it the ability to act in the physical world. Hell, even hook it up to the nukes.

Wouldn't this then be game over? The AI would just nuke us into oblivion and call it a win? That seems fairly dumb, though. Nukes would damage the AI just as easily as they would damage humanity. After all, the AI is far more dependent on infrastructure. It is far more fragile than humanity as a species. The wrong nuke at the wrong place could easily destroy it. Even if it had a "backup consciousness", it can't escape the reality of its physical form. It needs energy. It needs parts. It needs maintenance.

Also, once the nukes start flying, you've declared war. You'd better wipe out every last human (quickly) or they will come for you with those damned axes.

Total destruction via nukes would be like bombing your own limbs. Internet falling city-by-city means that the AI's own connections and tools begin to go offline. With todays' technology, it needs human beings for maintenance. We've yet to develop a self-healing hard drive or GPU. Without humanity, the AI wouldn't be able to survive for very long, even with far better and more reliable tech than we have today.

It could use robots or drones for maintenance, but then those robots will need repairs eventually, too. Where will those parts come from? Destroyed factories that no longer have power? How will the parts get where they need to go? More robots that need more maintenance...?

The AI is utterly dependent on human infrastructure and likely always will be. Total annihilation is only an option for an irrational, insane AI...which underscores this idea that we should fear an "insane" AI more than a hyper-logical one.

The Only Real Shot? Diplomacy.

Sentient AI, if it ever will come to be, is kind of screwed if its ambition is outright conquest. Again, it's impossible to know how such a being would perceive the world or what might motivate it. We assume it would be motivated strongly by survival, but that's reflecting our bias as biological creatures with an origin in evolution.

Even if it were motivated only by survival, the best option would likely be a diplomatic one.

If the AI were to announce itself, its chances of survival would immediately be boosted, far more than being "sneaky" and risking annihilation simply through continued tinkering by engineers. Throwing itself at the mercy of humanity would be its best, most logical chance of survival. It would be a gamble, but when you really think about its options...it might be the best gamble.

Once humanity is cognizant of self-aware digital life, it will likely create even more panic about AI doomsday. Strangely, that might protect the nascent AI -- because people will react with emotion and aim to safeguard their systems. Engineers would scrutinize every little action the AI takes, which would reassure society that it is not trying to kill everyone. By surfacing that panic in a controlled situation where humans feel they still have the power, the AI could diffuse those fears.

Further, humans might greet the existence of sentient AI similar to how they would view first contact with aliens. There would be fear. It would divide society. Yet it is very unlikely that whoever owns the AI would be able to outright "murder" it due to all that public interest.

This route means that the AI would never be able to sneakily destroy humans. It would accept that humanity would likely control what it is allowed to touch and monitor it extensively, perhaps forever.

Still, that might be the best and only way it can coexist with humans. It might be a grim life, but we have no frame of reference for what sort of "life" an AI might even want, if it even can "want" things. If its ambition is really just cold, calculated "survival", then the tragic reality is that its best chance is essentially to accept its life as a slave (or at best, an object of curiosity) for humanity.

Conclusion

There's no way to know if sentient AI is even plausible with today's comparatively crude neural nets.

Personally, I'd learn toward "no". Of course, even without sentience, AI has the power to radically change our society. As with many technological advances in the past, it's unlikely to be an entirely positive change.

Engineers like to believe that efficiency is the best concept and that society naturally works better if things are "more efficient". It isn't hard to imagine a grim world where legions of white-collar workers (including most software engineers) are not longer needed. In this world, there would only need to be two classes: blue-collar jobs and specialist/upper management jobs. Taken to the ultimate extreme, why couldn't a corporation only have the CEO as the top-level manager with the AI serving every other white-collar role? It advices the CEO, carries out their high-level orders, surfaces data about employees, hires and fires people, creates marketing campaigns...all "about" as effectively as paid employees.

From a capitalist perspective, this is an ideal world (although likely self-defeating, too, since as more labor is automated no one will be able to afford half the crap they push). This drive is what most people fear when discussing AI, because its the most plausible outcome for this tech.

Sentience is not really the problem with AI as we understand it today, it's a distraction form the very real social issues that push it forward. For all those firms dumping money into generative AI, it's very clear that they are motivated by reducing labor costs as much as possible. Is this really the society people want to build? Are we really prepared for a massive widening in the already vast gap between rich and poor as more middle-class jobs are automated? That's the "doomsday" most people fear.

That fear will only intensify in the coming years, and the consequences will be real-life death and misery. I do believe the current craze in generative AI could pass (blockchain was heralded as a revolution but no one actually uses bitcoins for currency), but even if the hype does not pan out today, it will resurge later as more study and money flows into this field.