A Short Guide to AI for Laymen
What’s all the hype?
--
The advent of AI is gradual, though for the masses it seems to occur in discontinuous leaps. In the 90s a human chess champion lost a game to Watson, a chess program from IBM, and perhaps the next thing you know is another match between AlphaGo and Lee Sedol hitting the headline. The increasingly staggering complexity that modern AI systems exhibit should concern us, as everyone, willy-nilly, will be affected by this irrevocable trend.
Structures
This article aims to inform people about artificial intelligence in general. It requires little domain specific knowledge to understand.
The content is structured as follows. First, a brief overview of AI systems will be given. Then I’ll dive into the notion of intelligence itself. Finally comes the part where we look at the prospects of AI, and what is still lacking in these systems. I try to make these parts independent to each other, so feel free to skim or skip uninteresting sections.
Overview
In his book Human Compatible*, Stuart Russell, professor of Computer Science at UC Berkeley, defined artificial intelligence as
Machines are beneficial to the extent that their actions can be expected to achieve our objectives.
Intelligence — the Essence of AI
Natural intelligence
For the sake of simplicity, we define* an intelligent entity to the extent that what it does is likely to achieve what it wants, given what it has perceived. The most ancient form of intelligence, as far as we know, came from good old evolution. An E.coli roams randomly in our lower bowels, and follows strict glucose-seeking rules that its genes have planned out for it. Simplistic and trivial creature, but it bears the basis for modern intelligence. It does this to maximize the reward from the environment so that the genes have a higher chance to survive and be propagated.
Such hard-coded behavior is not flexible. Mother nature cannot forecast every single detail of the environment, so the ability to learn after the phenotype has been grown is a huge boon to survival.
To further define intelligence, let’s consider the most simple case. For one rational single agent that is also intelligent, it aims to maximize a utility value. (Of course, humans are notorious for not being consistent in being rational, but that’s another story. For now, we can ignore occasional irrationality.) Such an agent aims to do anything in a given environment that is going to benefit its own existence, concisely represented as a utility value. This is nice and all, but the real world wouldn’t be just one and only agent. Thus it’s natural to extend this model. When 2 agents are involved, things get tricky. We cannot use probabilities to work out the best strategy, because how the other is going to behave is opaque. (This is a subject commonly known as Game Theory, one of the most well-known examples of which you may have heard of is the prisoner dilemma.) It turns out that Nash Equilibrium does give a solution, so long as both agents are rational. So far for biology.
Plain computers
Artificial machines initially were cumbersome and slow. Few viewed it as intelligent. Alan Turing, the famous British mathematician, defined the Universal Turing Machine that formed the basis of all modern computers. A core concept of Computer Science is algorithm. And software can run numerous algorithms as subroutines, building up layers of complexity. On the flip side, hardware has been observed to obey Moore’s Law, which predicts steady and exponentially better performance.
There must be a limit to it, however, as running a wrong algorithm on a faster machine gets you, well, the wrong answer, only more quickly. So a godlike machine, in an intelligent sense, is just as useless as a feeble one. Additionally, not all problems are computable in the first place. This means, for some problems a general computer may fail to spit out the answer — it never halts to return the solution. Consequently, problems that are structured in a way such that a Turing Machine(the essential prototype of all modern computers) can give a yes/no assertion are called decidable languages. Inside the realm of decidable languages, the fastest algorithms discovered so far for some problems sadly require exponential time to solve, with respect to the size of the input. And we all know of the explosive boost of exponential function as the input size gets larger.
These formidable obstacles seem to block our way into building an AI. Fortunately, they are not fatal. We know that a human cannot solve exponentially complex problem quickly either. AI, with more computational power, can provide decent suboptimal solution in reasonable time, and that’s already something worth celebrating.
Artificial intelligence
After 2 failures attempting to build AI systems before 2000, modern AI now taps into probability theory, utility theory, statistics, control theory, etc. And it seems to go surprisingly well this time.
The intelligence of AI is represented by an intelligent agent. An intelligent agent is something that perceives and acts intelligently. It can be specified to solve an chess match, or analyze the sentiments of a piece of text, both of which need totally different objectives, rules and design. The kinds of problems are numerous, and each requires a unique set of parameters and tweaks to work. Yet the ultimate goal of AI research is that specifics should not matter. We could just tell the agent to do something, and it’ll pick up all the relevant knowledge and skills along the way, ask for help if necessary and go about executing that objective.
When problems are small, a definite answer can be effectively searched. Upon completion, a successful agent will output a function which will tell us which what to do in every possible state. For example, for the problem of searching for a shortest path from A to B in a graph, an algorithm can print out the full shortest path, and we can follow this instruction for execution. However a large portion of problems is too vast to be traversed entirely. So far another kind of algorithm performs rather well on these problems. It is called Reinforcement Learning, in which the agent tries to maximize some reward function that’s given from the environment and learns as it iterates through the states.
Future — Expectations on AI
As mentioned in the beginning, AI progresses similarly to other research areas: it’s spiral and continuous, though sometimes into some cul-de-sac. One significant leap is the reap of probably decades of research and study. Despite great advancement, current AI is still a far cry from the ideal superintelligence.
When will superintelligence arrive?
This is a worrisome question but is preferred to not answer at all. 3 reasons were given by Stuart*:
- Wrong predictions in history. At least twice from respectable researchers. It’s prudent to not give these assertions at all.
- No clear-cut threshold. The systems are not either intelligent or stupid. In some specific areas, machines have exceeded human performance, and others are shamed by just a newborn baby.
- Intrinsically unpredictable. The development of AI or science by and large has a lot of uncertainty. Just a few hours after Rutherford published the article claiming the impossibility of harnessing nuclear power, Szilard refuted and presented his invention of nuclear chain reaction. The same applies to AI research as well.
What else needs to be done?
We are still unclear about the very nature of a superintelligence. Yet to be able to reach superintelligence, some major breakthroughs have to be made. This might not be an exhaustive list, but to overcome these challenges is necessary.
- Language and common sense
As a hallmark of human, language has been intensively studied but still baffles machines. Natural Language Processing(NLP) technology today:
can extract simple information from clearly stated facts but cannot build complex knowledge structures from text; nor can they answer questions that require extensive chains of reasoning with information from multiple sources.
As a fundamental means of communicating with humans, NLP is crucial to build superintelligence that serves us.
- Bootstrapping, recursion and reflection
This is one of the most powerful paradigms in Computer Science. The basic idea is that we can construct a datum that refers to itself, or a function that during its execution, calls itself. This recursive relationship is non-trivial and worth awe, in that the entity we are building right now has the ability to build upon its existing self. By the same token, we want our AI system to be able to gather some knowledge, process it and use that to construct and understand higher level concepts. It’s very much like writing a C programming language compiler without any language more expressive than C: people would write some assembly code to simply some tasks or routines, and iteratively use that to further develop and extend this compiler, until a full-fledged compiler is constructed. This is a typical process of bootstrapping.
- Cumulative learning of concepts and theories
Newton’s quote “If I have seen further than others, it is by standing upon the shoulders of giants” is an exemplar. Knowledge, however miniscule or grandiose, can be passed down to next generations. Unfortunately, AI systems still cannot accumulate knowledge. Very much like bootstrapping, intelligent machines have to cumulate generations of new concepts in various fields, in an order children might be taught at school.
- Discovering actions
Simply go to Google and search any keyword. Results from billions of web pages will be returned. This is inconceivable even just a few decades ago. Consider the computer network hierarchy(see figure above). Each time you type in google.com in the browser, oodles of protocols and operations, which may involve dozens of machines, are performed just to deliver the website to you. As users, all we need to know is how to type those keystrokes and the machines will run numerous networking algorithms that have been developed over decades. This feat, for human, is a convenient abstraction. We want AI to be able to learn this abstraction by itself, such that it can build layers of abstractions, akin to human. How this can be done still remains an enigma.
Are there some limitations to AI?
Indeed, imagine a superintelligence being able to consume all books ever written within a day, and you will get a chill down your spine. However, it’s safe to say that such an AI system cannot foretell a perfect picture of the real world. It is also, reassuringly, bound by the speed at which the world runs. That is, it cannot know whether a new drug would be effective unless some experiments or simulations have been made, the latter option is even more computationally expensive than a real experiment itself. Finally, superintelligence is not human, which imposes some difficulties when predicting and understanding actions/feelings that are native to human. For instance, the feeling evoked by scratching a blackboard with nails is immediately uncomfortable for us, but this has to be run in a complicated simulation for AI to comprehend.
How could AI benefit us?
In Cloud Computing, we distinguish PaaS(Platform as a service) and IaaS(Infrastructure as a Service). If a superintelligence exists, we shall gain huge advantages that it provides us with, which can be conveniently abbreviated as EaaS(Everything as a Service). Everyone benefits from the vastness of general AI. Education is personalized to fit one’s characteristics, health status can be calculated with exceeding precision, and the average living standards will be lifted by a notch.
Conclusion
While you may be still letting your imagination go berserk, this is the end of this article. We have seen how natural intelligence inspired artificial intelligence, what modern AI is about and what key pieces are missing in these systems. I hope you learned some key ideas behind computer science and artificial intelligence. The fact that AI is quickly permeating every corner of our life urges us to have a basic understanding of what AI is. For the nitty-gritty details, you may well refer to the references I listed below to dive into any field that piques your interest.
References and Citations
Tanenbaum, A., & Wetherall, D. (2010). Computer Networks (5th Edition) (5th ed.). Pearson.
Russell, S. (2019). Human Compatible: AI and the Problem of Control. Allen Lane.
Sipser, M. (2012). Introduction to the Theory of Computation (3rd ed.). Cengage Learning.
MIT. (2021). Introduction to Deep Learning. http://introtodeeplearning.com/
