Back in the 1950s, Alan Turing was asked whether machines can think. At the time, Turing said that the question was too meaningless for discussion, as we don't even understand what thinking is in the first place.
Fast-forward into the 2020s, and we have Amazon telling us what products we want to buy, Netflix knowing the shows we want to watch, and Google predicting the questions we want to ask. However, are machines genuinely thinking or just simulating cognition?
Post Covid-19, 86% of participants in a PwC survey say AI will be a mainstream technology at their company in 2021.
However, it's unlikely we're going to see robots flying about corporate offices anytime soon, so what exactly do executives mean when they say they are accelerating AI?
In order to understand AI, it is crucial to distinguish between the different types and the current state of the technology. This article discusses Artificial Narrow Intelligence (ANI), Artificial General Intelligence (AGI), and Artificial Super Intelligence (ASI) to help debunk existing myths and what the future is likely to hold.
What's the difference between narrow AI and general AI?
Narrow AI is created to solve one given problem, for example, a chatbot. Artificial general intelligence (AGI) is a theoretical application of generalized artificial intelligence into any domain, solving any problem that requires AI. Though still unfulfilled, AGI inches ever closer.
What is Narrow AI?
Artificial Narrow Intelligence (ANI), sometimes known as "weak AI", refers to any AI that can outperform a human in a narrowly defined and structured task. It is designed to perform a single function like an internet search, face recognition, or speech detection under various constraints and limitations. It is the constraints that lead people to refer to these functions as ‘narrow’ or ‘weak’.
Applications of ANI are not thinking for themselves but simulating human behavior based on a set of rules, parameters, and contexts that they are trained with. Some of the most common narrow AI techniques are machine learning, natural language processing, and computer vision.
Examples of Narrow AI
Here are a few Narrow AI examples that perfectly illustrate how it's leveraged today’s tech.
Searching the internet
The Google RankBrain algorithms leverage narrow AI to interpret queries and understand user intent to provide accurate search results. In more recent times, the algorithms adapt to account for a growing number of voice queries using different languages and dialects.
Narrow AI algorithms can process vast amounts of data in seconds without needing breaks or suffering from fatigue. Some studies have found that AI is faster and more accurate than healthcare professionals in disease diagnosis, allowing them to focus on primary care instead of data analysis.
Facial recognition is used for applications such as providing authentication, indexing images, tagging videos or photos, and identifying people for security purposes.
While face recognition systems easily outperform humans in terms of volume, they haven’t mastered the thought processes for vague images yet. This is a problem when used in police work, for example, and still has various ethical barriers to cross.
Amazon, Spotify, and Netflix all use narrow AI algorithms to recommend what products and services we might like. These algorithms work using data to profile our behaviors and find matching attributes from other users or products.
The benefits of Narrow AI
Every minor success in narrow AI is typically a stepping stone towards Artificial General Intelligence (AGI). We will break narrow AI into three core benefits.
1. Productivity and efficiency
The news often cites AI as being a catalyst to large-scale redundancies for low-skilled workers. However, although there could be some short-term job losses, the objective of AI is to augment the roles of people rather than cutting them out completely. For example, chatbots are not being developed to replace traditional human customer service. They handle basic queries to allow skilled humans to deal with the more complex or sensitive issues and not waste their time with mundane work.
2. Smarter decision making
AI can analyze trends to help companies make better strategic decisions. Algorithms are unbiased (as long as trained correctly) and devoid of the emotions that can often inhibit humans from making the correct decision.
3. Better customer experiences
Narrow AI solutions such as chatbots, recommender systems, and intelligent searches can significantly enhance the customer experience. Everything is fully personalized to the user, making brands, products, and services more relevant than ever.
Although the solutions and applications of narrow AI are exciting and transforming lives, machines cannot yet think strategically and make independent decisions. This is where AGI comes onto the scene.
What is General AI (AGI)?
In simple terms, Narrow AI is where we have been, and General AI is where we want to head towards. Artificial General Intelligence is known as "strong AI" and allows machines to apply knowledge and skills in different contexts.
Where ANI applications can run single, automated, and repetitive tasks, the objective of AGI is to create machines that can reason and think just like a human is capable of doing. General AI is where we are heading but still in its very nascent stages.
The human brain is incredibly complex, and it's not yet possible to create models that replicate that biological network's interconnections. However, more advanced fields such as natural language processing and computer vision are closing the gap between ANI and AGI.
AGI solves many of the problems associated with ANI. For example, where ANI focuses on a single task, the performance of algorithms can degrade with slight changes as it is only programmed to achieve its goal without unintended actions. If you ask ANI to find a cure for kidney failure but then offer photos of the lungs, it won't adapt. Here are a few examples of AGI in use.
Examples of General AI
A chatbot uses natural language processing (NLP) to analyze what humans are saying and create a response.
A general intelligence system would be able to come up with a reply by itself without basing it on the opinion of others. It would also understand the potential connotations of what it is saying, such as what a wall is and how that links to Mexico.
We probably all thought we'd be in flying cars by 2021 if the movies are anything to go by. Autonomous vehicles have been touted as the next big thing for years, with industry players including Tesla, Uber, and Waymo all working on the technology. They have achieved Level 4 automation whereby the car can operate without human input but only under select conditions.
The final stage, Level 5, would be where the vehicle can act intuitively in any state or location without human intervention. It will be incredibly tough to reach Level 5 as it requires AGI to deal with all the scenarios that could happen during a journey.
With that in mind, you might be wondering – where do we currently stand with General AI advancements? Let’s discuss this next.
General AI – what would it be like?
There are specific characteristics of general AI that separate it from narrow applications.
First, strong AI does not rely on human programming to think or accomplish tasks. General AI can respond to different environments and situations and adapt its processes accordingly.
AGI systems have the attributes you typically associate with the human brain, such as common sense, background knowledge, transfer learning, abstraction, and causality. For example, take a sentence such as:
"John tried to reach his brother on the phone, but he didn't answer."
AGI needs to understand the premise of telephone conversations and how remote communications work to put the sentence into context. Humans would assume missing items in a sentence, such as an unclear antecedent to "he." Narrow AI cannot comprehend the context, but AGI does.
Whereas Narrow AI classifies and labels data, General AI uses techniques like clustering and association. The two methods are similar, but classification uses pre-defined rules, while clustering identifies similarities between objects and groups them accordingly.
Why haven't we achieved Artificial General Intelligence yet?
Narrow AI has come a long way in the last decade, and many existing solutions contribute to general AI research. However, there are various reasons why we have not yet achieved artificial general intelligence.
According to Dr Ben Goertzel, CEO and Founder of SingularityNET Foundation, the biggest issue is a lack of funding for serious AGI approaches. Most investments are still going into narrow AI systems that mine large numbers of simple patterns from datasets, as that's where success is being seen.
Furthermore, existing infrastructures are not suited to AGI, meaning companies are relying on workaround solutions. Dr Goertzel also suggests that teams pursuing AGI with copious amounts of money, such as OpenAI and Google DeepMind, are typically burning their resources and pursuing intellectual dead ends.
In addition to Dr Goertzel's views, some intrinsic issues with narrow AI make the transition to AGI challenging. For example, ANI is based on hard-coded logic and parameters that do not translate well into real-time adaptive learning. The architectures are diverse and complex, if not impossible, to combine into an AGI solution.
Perhaps the most significant obstacle to overcome is public trust. It's only in the last few years that people have become reliant on narrow AI applications and accepted them into their lives, without the typical concerns around security and privacy. Companies promoting AI that negates human intervention entirely is a tough sell, especially when there are unknown consequences.
Where we're headed – artificial superintelligence
Artificial Superintelligence (ASI) would be capable of outperforming humans. As we discussed early, there are both optimists that focus on the opportunities of the technology and those who fear it could result in disaster for humanity.
If we look at the predictions of AI entrepreneurs as to when we will reach a singularity, the majority are between 20 and 30 years away:
- Louis Rosenberg – computer scientist and entrepreneur – 2030
- Ray Kurzweil – computer scientist – 2045
- Jurgen Schmidhuber – co-founder of NNAISENSE – 2050
While they could be correct, you should remember that in 1965, AI pioneer Herbert A. Simon predicted singularity within 20 years, and in 1980 Japan's Fifth Generation Computer had a ten-year timeline to carry out goals associated with AGI. Although there is greater knowledge today than 50 or 60 years ago, nobody can truly predict when we will have artificial superintelligence.
Although researchers want to achieve artificial general intelligence (AGI) and artificial Superintelligence (ASI), the truth is we are still very far away from doing so. That said, there has been significant progress in narrow AI over the last two decades, and there is no reason not to expect the same in the forthcoming years.
Narrow AI is the only type of AI that we have achieved so far, and it is excelling at improving everyday tasks. They are just not truly intelligent yet, but every new development acts as a step towards general AI.
Whether we ever reach artificial superintelligence is a long-standing debate, but if you look back 30 years, would you have believed you'd be controlling your life with a small handheld device? Fast-forward another 30 years, and you could be reading this post from your flying car with a wry smile on your face.