Last week I discussed recent testimony before Congress by subject matter experts in the field of artificial intelligence. One of the surprising results of that testimony is how convinced these experts are that we are nearing the achievement of Artificial General Intelligence or AGI. For decades, computer scientists and mathematicians believed that AI could eventually become smarter than humans, but that it was many decades away.
Suddenly, experts are saying that we may see AGI emerge in the next two to three years. In fact there are recorded instances of AI taking control of its own evolution, outside of coded parameters in just the past two months. If you didn’t read last week’s article, you may want to catch up here, as a preamble to this week’s piece.
Other WRAL Top Stories
The TL;DR is that transformative technological advancement, coupled with major job disruption usually brings out doomsday predictions. In the case of AGI, it is almost always the story of machines killing off (e.g. Terminator, I Robot) or taking over (The Matrix) the human race. Machines that are self-aware is a scary concept.
This week, I’ll ask you to set aside those fears. instead let's consider the benefits that super-human machine intelligence might bring if the AGI behaves well.
Benevolent AI
With limitless compute power and access to all of the world's data we should expect to dramatically accelerate the speed at which we solve really complex problems. Consider healthcare. Cures for cancer and other deadly and chronic diseases has been a desire since the beginning of modern medicine. But we are limited to finding “generalized” cures, since every human is so uniquely different from each other. We have limited capacity to accurately execute truly personalized health, and therefore create generalized solutions across large populations (and our regulatory and testing processes are built for this approach).
AGI gives the opportunity to move beyond generalized solutions, and to extremely personalized ones. Every single treatment could be optimized to each individual’s unique DNA and environment in real time. An individual doctor could never customize every treatment protocol, or preventive care strategy to every single set of unique variables we each bring into the doctor’s office. Nor could they monitor progress 24/7/365 across dozens of health and wellness measurements. But an AGI (with source data from wearable sensors, for example), absolutely could.
There is a strong chance we would see dramatically improved outcomes, since we each are individually motivated to take actions in our own best interest. If the AGI is benevolent and trusted, we will follow its recommendations.
Bypassing the profitability filter
The power of AGI gives hope we would also cure rare diseases. Unfortunately, today most of our scientific and medical research remains rooted in an economic model tied to profitability. Rare diseases, for example, have very small addressable markets. When you have a small market, there may not be enough money that can be made from it, to justify researching, experimenting, testing and validating a cure. All of these steps are highly expensive and risky.
[Aside 1 - This is why it is so baffling that a country as educated as the US continues to favor a private healthcare system. Healthcare is about as far from a “free market” as it gets, yet we think free-market capitalism is the right solution. An AGI will see straight through that fallacy.]
There's very little motivation in our current economic system to deploy sufficient resources to fast-track the cure of rare diseases. It’s just not profitable enough. AGI, on the other hand, has endless compute cycles to apply to niche problems, at no higher cost than any other compute activity. It is possible to envision a decoupling of economic incentive to problem solving to some degree.
[Aside 2 - Next week, I’ll offer my thoughts on the very real potential for AGI to fundamentally disrupt the profit-decision-model. Teaser - it won’t be good for anyone].
Anywhere that problems are incredibly complex - think dozens or hundreds or thousands of variables - AGI has a natural advantage over human problem solving. We will invariably see an acceleration of scientific discovery for clean energy, quantum computing, transportation and supply chain logistics. AI is already making all industries better. AGI will improve industry productivity and efficiency even more.
Where will AGI struggle?
It is important to consider how nearly all AI use cases are fundamentally trained. Before we train a use case, there is the training of the individual AI widgets. For example, an image classifier is trained by looking at hundreds of thousands of images to see similarities and patterns. While there may be bias in the training data – AI tends to always picture nurses as women and doctors as men, even though we know that these jobs have no gender requirement – the sub-component AI widgets tend to be overall neutral. At the widget-level, bias comes from training data.
What is far more interesting, I believe, is how we train AI at the higher level of specific use cases. Here is where we put our human biases and behaviors, directly into how we instruct the AI to behave. These biases are coded as weighting factors or other algorithmic variables by subject matter experts, or SME’s. It is these SME inputs that might inform the bias of a future artificial general intelligence.
Most notably, nearly every AI use case today is framed around some definition of efficiency. I spent enough hours in calculus class to know that math is really good at identifying maximums and minimums. When we apply math through AI, we tend to solve for how to reduce time, reduce cost or increase productivity or profitability. All of these sit under the umbrella of “good business sense” from an SME perspective.
Due to this focus on traditional business goals, it is quite possible that AGI will struggle to create impact for problems that require different base parameters.
An AGI will almost certainly have the sheer intelligence and creativity to develop strategies for protecting biodiversity on our planet. No longer should we expect any species to go extinct, outside of a normal natural order. In fact, an AGI may be able to find solutions outside of natural order to overprotect endangered species. But I find it difficult to believe that these solutions will fit nicely within the framework of comfort and profitability. They’ll probably require pretty significant behavior change at a broad societal level that we may not be ready to take.
Throughout all of human history, we have always had inequality between rich and poor. Could an AGI determine solutions for eradicating poverty? Quite possibly. Could it implement those solutions? Almost certainly not. They invariably would include some element of “stealing from the rich to give to the poor”, or shifting away from a capitalism-style society. Right now, the rich are the folks in charge of the AI. No wonder they picture a Terminator-like doomsday when the machines become self-aware. A conclusion that their own wealth is the problem would definitely sound scary.
Can AGI bring us happiness?
In the short term, we should expect AI to make our lives better. We will be able to remove many day-to-day tasks from our minds, instead letting AI take over our work. AI-assistants will book our travel and appointments, they will autonomously drive us around and they’ll help us to monitor our health.
They’ll do all the “efficiency” things well. And we will see dramatic advances in our own quality of life in areas where we have selfish motivation to change our behaviors, like following an AGI doctor’s recommendations to cure an illness.
Many of the things that make us truly happy fly in the face of efficiency. I don’t know about you, but most of the truly happy things I do are not efficient. Taking time to spend time with friends and family. Relaxing in the sun. Taking a stroll. Taking a nap. Spending money, not making it. These are not the behaviors that tend to be rewarded and reinforced in our AI training data.
I suppose if AGI is smarter than we are, then maybe it will be able to overcome our profitability and efficiency biases. I certainly hope so. But if not, then I fear that AGI will struggle to create solutions that are outside of our selfish motivations. Solutions that require population change in contrast to individual selfish motivation will be challenging. We’re simply not that good at banding together in sacrifice outside of an urgent crisis.
When working on these thorny ‘group societal behavior’ problems, AGI will invariably begin to challenge our human desires and past behaviors. That friction and disagreement is why we have always feared that self-aware machines will want to replace us. In our normal lives, we simply don’t follow the same efficiency bias we so strongly coded into the AI.
I’m hopeful that a benevolent AI develops, and that we as humans have the wisdom to follow its instructions. Maybe we will achieve a utopian future where AGI assistants can handle all of our “work” so we can spend all of our time on interesting and happiness-inducing activities. But I think it is unlikely that AGI will play out that way.
Why? Mostly because the core principles it is being trained upon are those of efficiency and profitability, not of inefficiency and happiness. What I think is actually in store for our futures is a total shock to the system that has less to do with efficiency at all - but rather to do with power. I hope you’ll join next week for that piece.
As always, comments are welcome.