Probably best to start at the semi-beginning as I’m sure you’ve already heard the term ‘AI’. If you ask anyone on the street, you’d most likely get some vague reference to the Terminator movies or some ‘tech that took over the world’ story that left us all in a grave state of dysphoria. On the other hand, though, an AI researcher will tell you that it’s a set of algorithms that produces results without having to be programmed specifically to give that particular answer to a given scenario. Both interestingly enough are correct answers but let’s get it cleaned up a bit to make it easier to follow along ok.
A clearer definition of AI
In basic terms, Artificial Intelligence (AI) is:
- An intelligent set of algorithms created by humans, designed to mimic human thinking or human thinking patterns
- Fully capable (relatively speaking) of performing tasks intelligently without explicitly being told what to do
- Capable of rationally and humanly (the appearance of) thinking and acting on its own
How can we know that really happens?
We use a few tests to verify that AI is actually functioning. Without going too deep into the details, it would help to understand a bit of what a couple of the major ones are:
Turing Test: Simply put, the Turing Test would tell us that the AI entity used to speak with a human would not be noticeable to the human as an AI entity. It would appear to be human itself to the human agent.
Cognitive Modelling: The AI model needs to be based on human thinking and must ‘distil the essence’ of the human mind using introspection (how we think), psychological experiments (how we behave) and brain imaging (what our thinking looks like).
Using the Laws of Thought: This is a rather large list of logical statements that basically control the operation of our mind. The goal is to codify these laws and apply them to AI. This has its challenges as solving a problem in principle does not always translate well to problem solving in real life. Too many nuances make this a tough dragon to slay. For instance, we will take actions where we are not 100% certain of the outcome (how forceful do we turn an old rusty key to open a lock for example).
Using Rational Agents: This goal is simply to be the best possible outcome in its present circumstances. Unlike the laws of thought approach where only logical statements rule the outcome, there are instances where there is no logical right thing to do with multiple outcomes involving different outcomes and corresponding compromises. This process is more dynamic and adaptable, accounting as best as possible for the shortfalls in the Laws of Thought.
So now that we’ve seen how we determine the existence of AI, we will naturally progress how best to categorize AI by its levels of intelligence. Of this, we can say that AI comes in three levels:
- ANI – Narrow Artificial Intelligence
- AGI – General Artificial Intelligence
- ASI – Super Artificial Intelligence
Type #1 – Artificial Narrow Intelligence (ANI)
Also known as Weak AI, this is when AI mimics human intelligence and/or behavior but remains isolated to a narrow range of parameters and contexts.
For the record, today, all existing AI are ANI.
Very important in this statement above is that we do not confuse narrow intelligence with low intelligence. The AI that Google is working with to rank pages is not stupid. It is very sophisticated technology worth billions, but it cannot do much more than just rank pages. It can’t write code for example, for that you’d need a codebot.
ANI 2.0
Using the above example of Codebots at Google, one way to expand the capabilities of AI would be to combine separate technologies.
The Codebots platform used at Google is an ecosystem consisting of individual Codebots (Lampbot and Javabot). These bots are both hooked up to several technologies, most important of them for this purpose is natural-language processing (NLP).
These Codebots reside in the background. They will interpret models and instructions and in-turn output code. The chatbots are the interface through which all this happens.
Type #2 – Artificial General Intelligence (AGI)
To have an AI’s ability to mimic human intelligence and/or behavior to the point it becomes indistinguishable from that of a human is known as AGI: also known as Strong AI or Deep AI.
Experts are in agreement that this is possible but not necessarily in agreement as to when this may come to fruition. If we look at one of the most advanced supercomputers built today (Fujitsu’s K computer), for a single second of neural activity to be actioned, it took about 40 minutes. It will be quite a bit of time before we see exponential growth in this area.
Type #3 – Artificial Super Intelligence (ASI)
When an AI not only mimics human intelligence or behavior but in effect surpasses it, it’s becomes an ASI.
All we can do today is take an educated guess on what that looks like. ASI would surpass our knowledge in the fields of math, medicine, writing, and much, much more. All things we do will be done at such speeds it would seem incomprehensible to us how much knowledge an ASI system would contain as it would always be learning from both external and internal resources.
So, given what we just read, the question would be, is it even possible? To formally count as an ASI though, to us, this AI would need to be capable of things we believe we will always do better than bots. This would include things like relationships and the arts for instance. Even the most optimistic experts believe AGI, let alone ASI still requires decades and decades of advanced research. It is not uncommon for experts to state it may take centuries to meet a fully functioning AGI system.
Final Thoughts
For now (and maybe that’s a good thing), AGI and ASI are just Sci-Fi and all existing AI is still ANI. Siri, Google and Alexa are great to have handy and while they can handle most of the basic heavy lifting of daily planning and research, the remain nowhere near leaving the ANI neighborhood any time soon. Less time working and more time living remains the goal but that will not be happening in ASI mode for a very, very long time….and maybe for now, until we understand it better, that may be a good thing.