Unless you’re very good at ignoring both the news and social media you’ve probably heard a number of claims that the singularity is rapidly approaching, machines that are smarter, deadlier, and perhaps even sexier than us are coming and they are just years away from becoming reality. You might have even heard that robot overlords are going to kill us all.
It goes without saying that there’s a lot of info out there about the upcoming AI revolution, also referred to as the third wave of AI. Some of this info is hype and some based in scientific fact.
If you step back and take a look at the history of artificial intelligence alongside the latest developments and breakthroughs taking place in the AI field today, the upward trajectory of AI research becomes clear.
Still, the question remains.
Exactly how smart are machines right now and how long until there’s a super-intelligence that outsmarts us all? Are we just around the corner from reaching a level of artificial general intelligence that can match us at every turn?
The Waves of AI Development: Explained
Historically, development in artificial intelligence has gone through waves of machine intelligence capability which can be used as a benchmark for the level of artificial intelligence and its current capabilities.
Many experts believe we’re about to enter the third wave of AI but how did we get here?
The First Wave of AI
The first wave of artificial intelligence was based around ‘handcrafted’ knowledge. This enabled systems that could use logic to reason specific and narrowly defined problems or tasks. These systems required domain-specific rules that were first created by human experts in specific fields such as finance, logistics, or scientific applications and then input into a computer for it to process and comprehend.
Tax software is a good example of this type of rules-based approach. In this case, once the human experts (in this case, lawyers and accountants) established a definitive set of rules that must always be adhered to, the machine begins learning (but not actually learning) these rules using simple logical reasoning.
There were many issues with this kind of system, however.
For example, while first wave AI systems were great at reasoning grounded in logic-based rule following, they weren’t really good at anything else that would otherwise constitute any kind of real intelligence. Additionally, a system in the first wave of AI was unable to contextualize information based on what was happening in the wider world as its ability to perceive was non-existent.
When it came to tasks considered fundamental markers of intelligence, such as the ability to learn new knowledge or make abstractions based upon data acquired at an earlier time, first wave systems were not up to the task and didn’t cope well when it came to operating in the real world, beyond pre-programmed rules and scenarios.
First wave systems are good at:
- Applying logic-based reasoning to solve specific problems based upon rules provided by (human) domain experts
First wave systems are not so good at:
- Learning, abstraction, and perception – key indicators of intelligence.
A good example of the limitations of first wave AI could be seen in the DARPA Grand Challenge event of 2004 in which teams developing early versions of automated vehicles were tasked with driving 150 miles in the desert.
Of course, no vehicles completed the course.
This was largely due to the limitations of computer vision systems in the car which were unable to distinguish between natural features in the surrounding landscape. As a result, they continually made wrong decisions. Things move quickly in the world of AI development however, and just one year later in the 2005 Grand Challenge, a total of five vehicles successfully completed the entire course.
Their secret to success?
The Second Wave of AI
The core element distinguishing second wave AI systems from the first wave has been the application of statistical learning. This development allowed for unparalleled advances in everything from face recognition and speech recognition to machines that can (more or less) understand natural language and text.
The transition from the first to the second wave of AI can be placed somewhere around the early 2000s and over the subsequent years machine intelligence has steadily increased and it’s in this second wave of AI where we currently find ourselves. The current boom in artificial intelligence related developments, as well as the influx of intelligent consumer end-products from big tech companies, result from rapid and recursive progress made during the second wave of AI.
This era has predominantly been fueled by significant advances in the understanding, development, and application of areas such as advanced statistical learning and neural nets. From voice recognition assistants such as Alexa, Google’s language translator, self-driving cars, or computer programs that utilize deep neural networks to beat humans at one of the most complex games in the world, everything is built upon advances made in the second wave of AI.
The underlying driving forces behind this current level of AI progress are thanks to a combination of increasingly powerful computing ability, unprecedented and near-limitless amounts of available data, and most importantly, the implementation of advanced machine learning processes that enable systems to learn and adapt to increasingly complex situations.
It’s important to remember that even at this level of progress no machine is able to learn on its own. In fact, a considerable amount of work has to first be carried out by engineers and programmers to create models for computers to learn from before training them on huge data sets.
The second wave of AI systems are superior to those in the first wave in many of the areas we typically associate with intelligence, although they still continue to fall short in others.
The ability to understand or perceive the surrounding environment and natural world is one of the main areas in which the second wave of AI does particularly well. These kinds of system have the capability to distinguish between objects and faces (visual recognition) as well as from other inputs like sounds (e.g. voice recognition).
With the ability to train second wave systems on big datasets comes the ability for these systems to actually learn from the information they’re exposed to. This means that second wave AI systems are particularly good at both learning and adapting to different kinds of situations – great news if you’re designing an autonomous vehicle for driving across the desert.
As mentioned earlier however, systems from the second wave of AI are considerably more limited when it comes to carrying out logic based reasoning tasks. Similarly the capacity for abstracting knowledge from one area and applying it to another is something that second wave systems continue to struggle with.
Systems that fall within the second wave of AI are good at:
- Classifying and predicting the consequences of data thanks to the ability to learn and adapt based upon the data provided to the system.
These systems are not so good at:
- Logical reasoning or setting any kind of contextual basis in which events are taking place – second wave AI systems have minimal capacity to generalize.
There’s no doubt that second wave AI systems are now significantly more powerful than anything that’s come before and with more data available every second, the ability for these systems to recursively improve and get smarter isn’t going to slow down any time soon.
To get to the next level of intelligence however, extra steps are needed.
If we want to develop machines that are truly intelligent, we need systems that are able to contextualize the information they’re taking in to gain a deeper understanding of what’s going on.
You’re absolutely right increasingly smart AI – that is indeed a 25 year old woman that you’re currently looking at, but why is that the case and what does that actually mean in terms of context, and how did you come to all of these conclusions beyond mere statistics?
Systems in the second wave of AI are smart, but they’re by no means infallible and whether it’s trivial scenarios like the incorrect classification of an image of a cat or a potentially deadly mistake made by an autonomous vehicle system, the margin for error that exists in the current level of artificial intelligence is simply unacceptable for some real-world applications.
Getting machine intelligence to the stage where it’s able to adapt to context is the key to overcoming many of the existing weaknesses of second wave AI systems.
Once this is accomplished, the third wave of AI will be underway. And yes, it will be a game changer.
The Third Wave of AI
The third wave of AI is set to see the development of systems that can tackle the challenge of learning in a way that’s much closer to how humans learn.
Whereas second wave systems relied heavily on machines being provided with vast amounts of pre-labeled training data to learn from, third wave AI will begin to reduce the requirement of this kind of data ‘feeding’, learning instead from contextual models.
For example, rather than training the AI over hundreds of thousands of images of dogs (from which the machine will statistically evaluate millions of pixels), third wave AI systems should be able to learn from descriptive models that tell it the features to look for in order to identify a dog.
Through this approach, it wouldn’t matter whether the AI was told by text, spoken instruction, or through observing a visual image, all should result in the system being able to identify an object correctly.
This is different to the way machines currently learn and would advance machine intelligence significantly closer to our own as this is similar to the way in which humans learn to recognize objects.
Moreover, an intelligent machine from the third wave of AI should be able to explain why it was able to reach the conclusion that what it’s observed is a dog, rather than simply calculating the statistical probability of ‘dog’ over other possible outcomes.
Third wave AI systems would be able to explain that the observed object is indeed a dog as it has all of the characteristics that we would typically associate as being ‘dog-like’ such as a wet nose, four legs, a wagging tail, and any other specifically nuanced features or information.
While this kind of system will initially require huge amounts of data to train it (which of course, we have), the ace up the sleeve of third wave AI systems is that they will increasingly be able to apply contextual learning in the way that humans do.
This means that they will increasingly be able to use contextual models in order to understand how the world should be perceived and subsequently to recursively learn and improve itself using the model to reason and make logic based decisions.
As the third wave of AI develops further, it’s even possible that intelligent systems will be able to use this kind of contextual modelling to abstract and build upon data in a way that current AI systems can’t.
So what could the third wave of AI look like in reality and what kinds of impacts could it have on our everyday lives?
In theory, when machines are able to learn in this way it will open up opportunities in many areas currently out of reach by present AI.
This could be anything from contextually based (and therefore extremely natural sounding) language applications, AI researchers and workers that never get tired, and almost certainly countless consumer-facing and industrial uses that will make the smart machines of today seem pretty dumb.
Put simply: when we reach the third wave of AI, machine intelligence is going to feel noticeably more intelligent.
Looking Further Ahead: The Fourth Wave of AI
So what happens when we reach the end of the third wave of AI and the machines around us have mastered the art of contextual learning, adaptation, and self improvement?
The simple answer is that no one really knows for sure.
What we can be certain of however, is that when artificial intelligence becomes smart enough to learn through a truly adaptive and context-based approach to interactions with the natural world and the data rich environment in which we increasingly live, things will begin to move very quickly.
If the third wave of AI reaches a sufficient point in its evolution that machines are able to abstract data and take it further in recursive learning and improvement, it’s highly likely that it will only be a matter of time until some degree of human-level AI or artificial general intelligence (AGI) is reached.
At this point, the AI can be considered capable of performing any intellectual task that we can, making them as smart as (and very probably smarter than) humans.
If this level of AI development is reached, the potential for breakthroughs will increase drastically as the ability to think through problems that human brains currently can’t is removed as an obstacle. This could mean major scientific developments in everything from computing, to medicine, and even environmental and energy science.
It’s also highly likely that should we ever develop AGI, the possibility of reaching a level of artificial super intelligence (also known as the singularity) will increase significantly. At this point, the AI will have surpassed us intellectually and will potentially develop rapidly, utilizing previous experience and sophisticated contextual abstraction to improve itself.
It therefore goes without saying that by the time we get to this stage in the development of AI (assuming we do), it’s critically important that we’ve built in a system of rules, ethics, and a comprehensive moral compass that ensure any intelligent machines we end up sharing the world with have values that are aligned with, not against, ours.