From Narrow to General to Super Intelligence

In Part 2, we saw that machine intelligence fails, sooner or later, because of the human-machine value misalignment issue. In Part 3, we see why addressing the human-machine value issue becomes imperative as we approach the possibility of reaching artificial general intelligence (AGI) and artificial superintelligence (ASI). 

Artificial Intelligence (AI) is often described as a series of broad development milestones that define the extent to which machine intelligence replicates human intelligence (Figure 1):

• Artificial Narrow Intelligence (ANI) describes machine intelligence that can outperform humans on single or defined tasks.
• Artificial General Intelligence (AGI) refers to machine intelligence capable of performing any intellectual task that a human can do.
• Artificial Super-Intelligence (ASI) is a hypothetical level of machine intelligence that exceeds cognitive performance in all areas of human intelligence.

Figure 1: AI is often described as a series of broad developmental milestones that define the extent to which machines replicate human intelligence. The distinction between ANI and AGI is today’s current juncture.

Artificial Narrow Intelligence (ANI)

ANI describes machine intelligence that can outperform humans on single- or narrow-defined tasks in specified contexts such as answering a question, correcting grammar in a document, predicting the fastest route, or suggesting related articles to read. ANIs are automated solutions to problems or tedious tasks that have the ability to learn and continue improving its results.

Today’s narrow systems are quite capable in that they can perceive (hear and see), respond using natural language, distinguish among possibilities, predict likely outcomes, and respond to preferences. Amazon’s Alexa and Apple’s Siri, for example, are voice-activated software assistants that observe and collect data in real-time and pull information from different sources to provide the information you ask for. Autonomous vehicles use ANI to make split-second decisions based on a detailed inspection of the environment. Apple’s Face ID uses ANI to analyze 30,000 invisible dots as they’re projected onto a face.

With the help of highly capable sensors, vision systems, natural language processing, and other technologies, ANIs seem more and more human-like these days; however, from an intelligence perspective, they’re still just assistive tools that have a number of limiters:

• They can only replicate the more tangible aspects of human intelligence domains: Representing knowledge, responding, perceiving, planning, reasoning, and communicating using natural language. They cannot replicate higher levels of human intelligence.
• They’re limited in the possible outcomes and contexts. They can fetch, differentiate, and predict but only within boundaries of a data set.
• They are not autonomous and require human programming.

AGI refers to machine intelligence capable of performing any intellectual task that a human can do. For this to be possible, AGI systems would work autonomously and replicate all aspects of human intelligence and apply them without limitation: Understanding, applying knowledge in new ways, making judgements, having and adhering to values, having and responding based on feelings, morals, and ethics, and more.

Machines with general intelligence would not be just tools, but instead, autonomous agents. Rather than being limited by what they’re programmed to do or limited to a particular intelligence domain, such systems could work with uncertainty, apply knowledge to new problems, think abstractly, and strategize. A machine with human-level intelligence could set its own goals, decide what to learn, deciding how to learn, and decide how to apply new knowledge. The possibilities are endless.

ASI is a hypothetical level of machine intelligence that exceeds cognitive performance in all areas of human intelligence. From science fiction, we imagine superintelligence as referring to a sentient, self-aware entity with all of the capabilities of the human mind, only faster, with perfect and upgradable memory, untiring performance, and ability to continuously self-improve. Such intelligence could be described as omnipotent superintelligence that has an infinite quality in its potential.

However, there are different possible levels or types of superintelligence imagined:
 

• Speed superintelligence could do all that a human intellect can do, only faster.
• Cognitive superintelligence refers to a system composed of smaller AIs, where the overall system outperforms humans. This intelligence would be useful for smaller tasks completed in parallel, with intelligence increased by adding more smaller units or reorganizing them for better efficiency.
• Quality superintelligence refers to intelligence that humans can’t attain because of physical limitations in our senses, nerves, and brain. A good analogy here is comparing limits in animal intelligence to human intelligence.

Reaching AGI almost certainly guarantees that ASI would be reached and likely in a short time. However, the methodical advancement of narrow intelligence and these categories of superintelligence also point toward reaching superintelligence in very defined ways. Take, for example, the fact that machine processing already outperforms humans in completing narrowly-defined tasks such as checking spelling, identifying best driving routes, or filing email.

Similarly, advances in machine learning have inched us closer to general intelligence as well. Whereas narrow systems previously required humans to set goals, data set parameters, and outcomes, some of today’s self-learning systems require only input data and use adaptive algorithms to identify patterns, determine possible goals, and infer meaning. Early successes have also been seen in systems that can apply intelligence to related data sets.

The potential of highly intelligent systems to transform the world is akin to the changes brought about by previous industrial revolutions. The question isn’t whether intelligent systems will continue to transform our lives; the question is in what ways and to what extent.

As we continue advancing AI and its corresponding technologies, we must consider, plan, and act on the potential implications of future AI engineering. In Part 4, we’ll explore three challenges in AI Safety Engineering: Unpredictability, unexplainability, and incomprehensibility.