Artificial general intelligence is here, and it’s useless

Link post

*Disclaimer: I originally wrote this for my own blog and as an editorial for skynet today, a few weeks ago I started reading LW and I thought people here might enjoy it, I think it’s in the spirit of the conversation here. So I made a few edits and the text bellow is the result.*

One of the most misunderstood ideas that’s polluting the minds of popular “intellectuals”, many of them seemingly accustomed with statistics and machine learning, is the potential or threat that developing an artificial general intelligence (AGI) would present to our civilization.

This myth stems from two misunderstanding of reality.

One such misunderstanding is related to the refinement and extensibility of current ML algorithms and FPA hardware, however discussing that always leads of people arguing “what ifs” (e.g What if cheap quantum computers with very efficient I/​O become a thing ?), thus, I won’t pursue that line of thought here.

A second, more easy to debunk misunderstanding, is related to the practicality of an AGI. Assuming that our wildest dreams of hardware were to come true… would we be able to create and AGI and would this AGI actually have any effect upon the world other than being a fun curiosity ?

1. We already have AGIs

AGI is here, as of the time of writing there are an estimate of 7,714,576,923 AGI algorithms residing upon our planet. You can use the vast majority of them for less than 2$/​hour. They can accomplish the vast majority intellectual task that can be well-defined by humans, not to mention they can invent new tasks themselves.

They are capable of creating new modified version of themselves, updating their own algorithms, sharing their algorithms with other AGIs and learning new complex skills. To add to that, they are energy efficient, you can keep one running for optimally for 5 to 50$/​day depending on location, much less than your average server farm used to train a complex ML model.

This is a rather obvious observation, but one that needs to be noted nonetheless. Nobody has ever complained about the severe lack of humans in the world. If I were to ask anyone what they would envision as making a huge positive impact upon the future, few would answer vastly increasing birth rates.

So, if we are agreed on the fact that 70 billion people wouldn’t be much better than 7 billion, that is to say, adding brains doesn’t scale linearly… why are we under the assumption that artificial brains would help ?

If we disagree on that assumption, than what’s stopping the value of human mental labor from sky-rocketing if it’s in such demand ? What’s stopping hundreds of millions of people with perfectly average working brains from finding employment ?

Well, you could argue, it’s about the quality of the intelligence, not all humans are intellectually equal. Having a few million artificial Joe nobody wouldn’t help the world much, but having even a few dozen Von Neumanns would make a huge difference. Or, even better, it’s about having an AI that’s more intelligent than any human that we’ve yet encountered.

This leads me to my second point.

2. Defining intelligence

How does one define an intelligent human or intelligence in general ?

An IQ test is the go-to measure of human intelligence used by sociologists and psychologists.

However, algorithms can readily out-score humans in IQ tests, for a very long time, and they are able to do so reliably with more and more added constraints.

The problem here is that IQ tests are designed for humans… not machines.

But, let’s assume we come up with a “machine intelligence quotient” (MIQ) test that our potential AGI friends cannot use their perks to “cheat” on.

But how do we design it to avoid the pitfalls of a human IQ test when taken to the extremes ? Our purpose, after all, is not to “grade” algorithms with this test in a stagnant void, but to improve them in such a way that scoring higher on the tests means they are more “intelligent” in general.

In other words, this MIQ test needs to be very efficient at spotting intelligence outliers.

If we come back to an IQ test, we’ll notice it’s rather inaccurate at the thin ends of the distribution.

Have you ever heard of Marilyn vos Savant or Chris Langan or Terence Tao or William James Sidis or Kim Ung-yong ? These are, as far as IQ tests are concerned, the most intelligence members that our species currently contains.

While I won’t question their presumed intelligence, I can safely say their achievements are rather unimpressive. Go down the list of high IQ individuals and what you’ll mostly find is rather mediocre people with a few highly interesting quirks (can solve equations quickly, can speak a lot of languages, can memorize a lot of things… etc).

Once you consider the top 100,000 or so, there is most certainly a great overlap with people that have created impressive works of engineering, designed foundational experiments, invented theories that explain various natural phenomenon, became great leaders… etc (let’s call these people “smart”, for the sake of brevity).

But IQ is not a predictor of that, it’s more of a filter. You can almost guarantee that a highly smart person (as viewed by society) will have a high IQ, but having the highest IQ doesn’t even guarantee you will be in the top 0.x% of “smart” people as defined by your achievements.

So even if we somehow manage to create this MIQ to be as good of an indicator of whether or not a machine is intelligent as IQ is for humans, it will still be a bad criterion of benchmark against.

Indeed, the only reason IQ is a somewhat useful marker for humans is because natural selection did not use IQ, start having some mad-scientists IQ-based human optimization farms and soon enough you might end up with individuals that can score 300 on an IQ test but can’t hold civil conversations or operate in society or tie their shoelaces.

This seems like an obvious point, a good descriptor of {X} becomes bad once it starts being used as a guidelines to maximize {X}. This should be especially salient to AI researchers and anyone working in automatic differentiation based modeling, optimizing a model for a known criterion (performance on training data) does not guarantee success on future similar data (testing/​validation set), indeed, optimizing it too much can lead to a worst model than stopping at some middle ground.

We might be able to say “An intelligent algorithm should have an MIQ of at least 100”, but we’ll hardly be able to say “Having an MIQ of 500 means that an algorithm has super-human intelligence”.

The problem is that we aren’t intelligent enough to define intelligent. If the definition of intelligence does exist, there is no clear path to finding out what it is.

Even worse, I would say it’s highly unlikely that the definition of intelligence exists. What I might consider intelligent is not what you would consider intelligent… our definitions may overlap to some extent, we’ll likely be able to come up with a common definition of what constitutes “average” intelligence (e.g. the IQ test), but they would diverge towards the tail.

Most people will agree as to whether or not someone is somewhat intelligent, but they will disagree to no end on a “most intelligent people in the world” list.

Well, you may say, that could indeed be true, but we can judge people by their achievements. We can disagree all day on whether or not some high IQ individual like Chris Langan is a quack numerologist or a misunderstood god. But we can all agree that someone like Richard Feynman or Tom Mueller or Alan Turing is rather bright based on their achievements alone.

Which brings me to my third point.

3. Testing intelligence

The problem of our hypothetical superhuman AGI, since we can’t come up with a simple test to determine it’s intelligence, is that it would have to prove itself as capable.

This can be done in three ways:

  1. Use previous data and see if the “AI” is able to perform on said data as well or better than humans.

  2. Use very good simulations of the world and see if the “AI” is able to achieve superhuman results competing inside said simulations.

  3. Give the “AI” the resource to manifest itself in the real world and act in much the same way the human would (with all the benefits of having a computer for a brain).

Approach number (1) is how we currently train ML algorithms, and it has the limitation of only allowing us to train on very limited tasks where we know all the possible “paths” once the task is complete.

For example, we can train a cancer detecting “AI” on a set of medical imaging data, because it’s rather easy to then take all of our subjects and test whether or not they “really” have cancer using more expensive methods or waiting.

It’s rather hard to train a cancer curing “AI”, since that problem contains “what ifs” that can’t be explored. There are limitless treatment options and given that a treatment fails (the person dies of cancer)… we can’t really go back and try again to see what the “correct” solution was.

I wrote a bit more about this problem here, if you’re interested in understanding it a bit more. But, I assume that most readers with some interest in statistics and/​or machine learning have stumbled upon this issue countless times already.

This can be solved by (2), which is creating simulations in which we can test an infinite amount of hypothesis. The only problem being that creating simulations is rather computationally and theoretically expensive.

To go back to our previous example of cancer curing “AI”, currently the “bleeding edge” of biomolecular dynamics simulation, is being able to simulate a single medium-sized gene, in a non reactive substance, for a few nanoseconds, by making certain intelligent assumptions that speed up our simulation by making it a bit less realistic, on a supercomputer worth a few dozens of millions of dollars… Yeah, the whole simulation idea might not work out that well after all.

Coincidentally, most phenomenon that can be easily simulated are also the kind of phenomenon that are either very simplistic or fall into category (1), go figure.

So we are left with approach (3), giving our hypothetical AGI the physical resources to put its humanity-changing ideas into practice. But… who’s going to give away those resources ? Based on what proof ? Especially when the proposition here is that we might have to try out millions of these AGIs before one of them is actually extremely smart, much like we do with current ML models.

The problem, of course, is that resources are finite and controlled by people (not in the sense that they are stagnant, but we can’t create an infinite amount of resources on demand, it takes time). Nothing of extrinsic value is free.

Which leads me to my fourth point about AGI being useless.

4. The problem of gradual tool building

The process by which humanity advances is, once you boil it down, one of building new tools using previous tools.

People in the bronze age weren’t unable to smelt steel because they didn’t have the intelligence to “figure it out”, but because they didn’t have the tools to reach the desired temperature, evaluate the correct iron & carbon mixture, mine the actual iron out of the ground and establish the trade networks required to make the whole process viable.

As tools of bronze helped us build better tools of bronze, that helped us discover more easy to mine iron deposits, build ships and caravans to trade the materials needed to process said iron and build better smelters, we were suddenly able to smelt steel. Which leads us to being able to build better and better tooling out of steel… etc, etc.

Human civilization doesn’t advance by breeding smarter and smarter humans, it advanced by building better and better tools.

The gradual knowledge that we acquire is mostly due to our tools. We don’t own our knowledge of particle physics or chemistry to a few intelligent blokes that “figured it out”. We owe it to the years of cumulative tool building that lead us to being able to build the tools to perform the experiments that gave us insights into the very world we inhabit.

Take away Max Planck and you might set quantum mechanics back by a few years, but in a rather short time someone would have probably figured out the same things he did. This is rather obvious when you look at multiple discoveries throughout history, that is to say, people discovering the same thing, at about the same time, without being aware of each other’s works.

Have Max Planck be born among a tribe of hunter gatherers in the neolithic period, and he might become a particularly clever hunter or shaman… but it’s essentially impossible for him to have the tooling which allowed him to make the same discoveries about nature as 20th century Plank.

However, to some extent, the process of tool building is inhibited by time and space. If we decided to build a more efficient battery, or a more accurate electronic microscope, or a more accurate radio telescope, we wouldn’t be limited by our intelligence, but by the thousands of hours required for our factories to build the better tools required to build better factories to build even better tools required to build even better factories in order to build even more amazing tools… etc.

Thousands of amazing discoveries, machines and theories lie within our grasp and one of the biggest bottlenecks is not intelligence but resources.

Not matter how smart your interstellar spaceship design is, you will still need rare metals, radioactive materials, hard to craft carbon fiber and the machinery to put it all together. Which is rather difficult, since we’ve collectively decided those resources are much better spent on portable masturbation aids, funny looking things to stick on our bodies and giant bombs just in case we need to murder all of humanity.

So, would a hypothetical superintelligent AGI help this process of tools building ? Most certainly. But it will probably end up with the same bottlenecks people that want to create amazing things face today, other people not wanting to give up their toys and the physical reality requiring the passage of time to shape.

Don’t get me wrong, I’m not necessarily blaming people for choosing to focus on 3rd printing complexly shaped water-resistant phallus-like structures instead of using those resources to research senolytic nanobots. As I’ve mentioned before, defining intelligence is hard, and so is defining progress. What might seem “awesome” for the AGI reading this article could be rather boring or stupid for a few other billions of AGIs.

In conclusion

  • Artificial general intelligence is something we have plenty of here on Earth, most of it goes to waste, so I’m not sure designing AGI based on a human model would help us much.

  • Superhuman artificial general intelligence is not something that we can define, since nobody has come up with a comprehensive definition of intelligence that is self-sufficient, rather than requiring real world trial and error.

  • Superhuman artificial general intelligence is not something we can test, since we can’t gather statistically valid training datasets for complex problem and we can’t afford to test via trial and error in the real world.

  • Even if superhuman artificial intelligence was somehow created, there’s no way of knowing that they’d be of much use to us. It may be that intelligence is not the biggest bottleneck to our current problems, but rather time and resources.

If you think you’re specific business case might require an AGI, feel free to hire one of the 3 billion people living in South East Asia that will gladly help you with tasks for a few dollars an hour. It’s likely to be much cheaper, Amazon is already doing it with Alexa, since it turns out to be somewhat cheaper than doing it via machine learning.

Is that to say I am “against” the machine learning revolution ?

No, fuck no, I’d be an idiot and a hypocrite if I thought machine learning wouldn’t lead to tremendous human progress in the following decades.

I specifically wanted to work in the area of generic machine learning. I think it’s the place to be in the next 10 or 20 years in terms of exciting developments.

But we have to stop romanticizing or fear-mongering about the pointless concept of a human-like intelligence being produced by software. Instead, we should think of machine learning (or “AI”, if you must really call it that) as a tool in our great arsenal of thinking tools.

Machine learning is awesome if you apply it to the set of problems it’s good at solving and if we try to extent that set of problems by being better and collecting and building algorithms we might be able to accomplish some amazing feats. But the idea that an algorithm that can mimic a human would be of particular use to us, is similarly silly to the idea that a hammer which also serves as a teaspoon would revolutionize the world of construction. Tools are designed to be good at their job, not much else.

And who knows, maybe someday we’ll combine some of these awesome algorithms we’ve developed, add a few extra bits, shield them inside a realistic-looking robot body and realize that the “thing” we’ve create might well be a human… then it can join us and the other billions of humans in being mostly useless at doing anything of real value.