Young’s AI Tier Rating Scale
I wrote most of this about ten years ago, 2010s, when Deep Learning was the latest new AI innovation that was going to take our jobs (Siri hasn’t taken my job). I use this both for evaluating technology, and for writing science fiction. No, it was not written by an AI.
I see another news article about the how there is an “Artificial Intelligence revolution”. I still think we are just scratching the surface. There have been a number of ways of categorizing AI, but I put on my science fiction writing hat over my computer science degree hat to put together my own categorization. By my estimation, we are pretty mature on tier 0, have some solid tier 1 AIs, making progress towards tier 2 & 3, and just scratching the surface of tier 4.
Tier 0 AI: This is a helping technology that has no intelligence in itself. Many of these were invented by early workers in the AI field. They may at a glance appear to have some intelligence, but don’t. These include spelling and grammar checkers, relational databases, search engines, voice output, and speech recognition.
Tier 1 AI: A tier 1 AI does a specific function, but can learn to do it better. For example a speech recognition engine that adapts to the accent of its user(s) to give better speech recognition. To the end user, a tier 1 AI may be indistinguishable from its fixed-functionality tier 0 counterpart. A tier 1 might recognize what time of day you do things and volunteer the information you will want. For example, we have a water heater that watches our water usage and schedules when to cool down some at night to save money. A tier 1 might be a digital assistant that is a slightly intelligent calendar, weather program, etc.
Tier 2 AI: This is a subject matter expert within a narrow range of expertise. The user of the AI must be trained to only have it incorporate data within a narrow field of expertise. Scientists and engineers will work with tier 2 AIs. They will store an immense amount of data in the field. They will recognize trends in the field (but can’t form a hypothesis about the reason for the trend). They will recognize trends in data. They might identify conflicting data that may be in error. I would like it to be able to read the latest news articles and flag things that might be of interest for the particular users work. A tier 2 would be of little help to a medical doctor as its area of expertise has to be too narrow. A tier 2 AI may have several tier 1 AI helpers; speech recognition, building a glossary of specialized terminology, translating technical papers from other languages. Within its narrow field it actually understands how facts interconnect and physics works, so it can use data not just output it. Expert systems are very minimalist attempts at a tier 2 AI.
Tier 3: This is a generalist. It has less depth of knowledge than the tier 2, but can correlate knowledge from a broad range of subjects. A city planner would use a tier 3 to analyze the interplay between utilities, transportation, recreational facilities, shopping centers, and real estate prices. The tier 3 could indicate what infrastructure is needed to grow an area. It can define how to make a difference between an older neighborhood becoming a slum and being ripe for an urban renewal. A tier 3 can help a MD consider possible diagnoses or suggest tests. However, a tier 3 could not weigh factors and make decisions. A tier 3 may draw on a number of tier 2 AIs with expertise in traffic volume, social interaction, utility construction, etc.
Tier 4: The tier 4 AI is a tier 3 with a bunch of extra pieces for decision making and closer integration with tier 2 specialized knowledge. Simple decisions can be logical, such as the order in which you do the steps of building a house; foundation before walls, walls before roof, etc. More complex decisions requiring weighing conflicting data; how much inconvenience in longer commute time will people accept to gain a nicer home, neighborhood, school, etc. Such soft decisions aren’t just percentages of fuzzy logic. They also require understanding motives, which are based on feelings. As such a tier 4 understands feelings, and may have some level of emotional capability itself. A tier 4 can weigh all of these factors and make a difficult decision, and explain why it made the decision. A tier 4 can have a discussion in which it presents its arguments for a given decision and adapt its decision if the other person gives new data to take into consideration. A tier 4 may still have a field of expertise.
We may have a couple secret criteria for declaring an AI to be tier 4. These are things that the AI has to accomplish or figure out itself, but we aren’t going to tell it what they are.
Side note: At tier 4 and above, the AI needs some type of consciousness. Thus it may need to be running and thinking to itself even when you aren’t giving it any input to process. It may in turn need to clean up an organize the accumulated thoughts and experiences (sleep and dream, or some equivalent). This probably requires some mechanism to interrupt those thought processes, particularly with external stimulus, and in turn the possibility that it will be able to and choose to ignore the external stimulus when making progress on a very important problem (important to it).
Side note: At tier 4 and above, we gain a new risk that the AI may have emotional hang ups, problems, several mental health conditions, be psychopathic, etc. Some likes, dislikes, and personality traits are just an expected part of a healthy psyche, but as with human mental health conditions one or a few parts of a healthy psyche expanding to dominate the personality and over-ride logic and other emotions can result in an AI that is unusable, nonfunctioning, suicidal, or even homicidal.
Tier 5: A tier 5 AI can act as a fully functional human being, other than obvious biological differences. A tier 5 can be sent as an independent operative to a distant planet, the ocean floor, swimming through magma, etc. A tier 5 will have relationships with loved ones, likes, dislikes, hobbies, and a full range of emotions. A tier 5 could do the job of almost any low skilled or mid skilled human worker (but might be too expensive to replace many humans). A tier 5 AI can learn both knowledge and skills on their own from a book or class, but with no specific human interaction to program them. By this point either humans or AIs would have built AI building utilities that the tier 5s can use to make new baby tier 5s.
Side note: Maybe baby AIs have to start at tier 3+ or 4 or 5 and work their way up to the higher tiers.
Tier 6: A tier 6 has creativity, extrapolation, experimentation. It can become the next Einstein, Edison, Da Vinci, van Gogh, etc. It can go beyond what any human has done, not just a little better at doing the same things, or a variation on a known theme.
Tier 7: This hypothetical tier could have the ability to gain what we would view as god-like powers. It could watch every camera, phone, computer, email, website, etc. and manage all of human society. Or a tier 7 might leave humanity behind, travel the universe using faster-than-light technology of its own creating, and invent new life forms to inhabit other worlds. It might experiment with creating new pocket universes.
Certainly our list started with current technology and ended with science fiction that could be thousands of years in the future or never happen. The point is, when you stand back and look at the big picture, most of the current successes such as the popular deep learning technique tend to be just good progress on the tier 1-3 problem and far from the science fiction vision (usually tier 4-6).
We needs a concept of false AI. A chatbot or large language network may mimic a tier 3, even though it doesn’t understand what it is talking about.
Some other sources that have their own AI rating systems.
The original Turing Test was to fool a human in conversation.
Variations on the Turing Test
https://simplicable.com/new/artificial-intelligence-test
This one has a big list of tasks categorized as toy-level (games), physical, athletic, storytelling,
artistic, social, economic, intimate, meta
https://maraoz.com/2022/10/31/agi-roadmap/?utm_source=tldrnewsletter
This one classifies as reactive machines, limited memory, theory of mind, self-aware, artificial
narrow intelligence, artificial general intelligence, artificial super-intelligence
https://www.forbes.com/sites/cognitiveworld/2019/06/19/7-types-of-artificial-intelligence/?sh=41fac490233e
An AI entrepreneur suggested the AI should be given $100,000 and told to turn it into $1,000,000 The AI would have to investigate business ideas and start a business (presumably in a simulation). This is more challenging than a simple interactive chat test. However, maybe the majority of the human population couldn’t pass it.