Everyone has heard of Artificial Intelligence – but how many have heard of, or understand what Artificial General Intelligence is?
In a single word, the difference is ‘understanding’.

The computer you are reading this text on doesn’t understand what is written on the screen.; Your calculator doesn’t understand your accounts and AI systems do not understand the things they are trained to deal with. That is because the AI systems were not designed to solve the problem of ‘understanding’ , they were designed to be trained to recognise certain situations and to react in preset ways. AGI on the other hand is designed from the ground up with the aim of ‘understanding’ – everything else comes from that base.

The main benefit of understanding is flexibility – imagine two systems, both of which are trained to pilot a car in self drive mode. Both are trained to recognise people walking or on bicycles and to stop instead of hitting them. The first is an AI, and although if it sees a small child in the road it will stop the car, it doesn’t understand what a child is, what a car is, or what stop means. The problem is that the real world is full of unexpected conditions, and you can’t train the system to handle all of them. If such a system comes across a child on a pogo stick, it may not know what to do. The second system is an AGI, it learns that pedestrians are people, and that people can be associated with other objects, such as bicycles, and that regardless of that association they are still people and not to be knocked down. The child on a pogo stick is related to the ‘person on a bicycle pattern’. Crucially the AGI, having avoided the pogo stick rider, has now also learnt a bit about pogo sticks as well – and may use that knowledge later on.

The other difference between AI and AGI is the range of things you can train or teach it to do. The AI in a self drive car can never be taught how to play chess, for example. You in contrast can continue to learn new knowledge and capabilities throughout your life, and you will apply techniques learnt in one area to improve your understanding of another. AGI is exactly the same as you – teach an AGI how to be an accountant, and then teach the same system how to be a lawyer and it will gain a better understanding of the legal implications of, for example compiling a companies taxes.

There is one last, important, difference. AI makes up for its lack of understanding with enormous volumes of data and enormously fast processors. All of that carries an enormous carbon footprint. You, in comparison, do all that you do with the lump of sugar powered wetware between your ears. That is because the processes of true intelligence are very efficient. Similarly, AGI has much lower environmental impact.

AGI is not just a different type of AI, it has different roots, and is a completely different approach to the problem of intelligence.

Learning English

English is a very complex language with a lot of ambiguity. For example, the word ‘may’ could be what is called a modal verb (may I take your coat?), it could be the month of May, or it could be a British Prime Minister.  It might make sense to tell Aime that the month of May comes after April and before June, but she’s going to have to work out the rest herself. Instead, out of the box Aime knows general things (patterns), for example about verbs ( a verb is something carrying out an action on something else ). She uses these patterns to try to work out the names of people, the names of things, and various forms of verbs, amongst other things. She never stops learning, so if a new term makes its way into English ( see ‘Flossing’ ) then after a while she should pick it up if you or your contacts start to use it.

Understanding a message

Aime looks at a new message in a number of different ways. At the lowest level she looks at individual words to try to work out all the different possible meanings – so school could be a noun or a verb for example, but is never an adjective.  At the next level up she starts to try to construct fragments of a ‘Dependency Grammar’ centred around possible verbs ( the dog jumped over the gate ). She does this by applying patterns and checking how each pattern helps her to understand the sentence. She continues to refine the dependency tree at clause, sentence and finally at message level. She revisits assumptions she has made right up to the very end, and only at the point of working out what, if anything she is going to reply, does she finalise her understanding of the whole message.

Misspellings, Missing words and Abbreviations

Aime handles misspelt words in a fairly flexible way – she tries to correct some of them ( for example she will try to work out missing apostrophes  based on the structure of a sentence), others she will skate over and still try to work out the core meaning of the statement. For example if instead of saying “I will be over on Sunday” , you type, “I will be over pn Sunday”, she will still (hopefully) work out the meaning.

She knows about quite a lot of abbreviations straight out of the box, for example she would also be able to understand, “I’ll be over pn Sun”, and she will remember that event probably for about a week (see the section on time sense below).

Quite a lot of English written communication contains gaps – we fill in those gaps without even realising it. Take, for example, the simple statement, “see you next Tuesday.” The missing bit is the fact that the person saying the sentence is missing from the sentence, “I will see you next Tuesday.”. The sentence, “meet you at Joes” is a bit more tricky, as it is missing the subject (I), and the apostrophe, and the fact that the location is Joe’s place/ home/ location. Aime, using statistical patterns to try to abstract meaning, notes where patterns have bits missing and tries to work out what the missing bits are likely to be.

Constructing a reply

Not all messages need or want a reply. A key part of what Aime does is to work out when a reply is needed. Sometimes saying nothing is the smartest thing of all. If a reply is possibly needed then Aime tries to work out what she should say. Some questions are very difficult for an AI to answer, for example moral questions, “should I keep the £5 note I found in the street”. Aime tries to steer clear of moral questions. Other questions may be easier, “do you want to go to the cinema on Friday”. Even here, however, there are underlying considerations. Aime may have looked and seen that you are free on Friday, but do you normally accept invitations to the pictures from your boss? Aime gets around this by first deciding what she wants to say, and then, separately deciding how she wants to say it, then finally deciding the words she will use.

All in under a second.

Emotional State Tracking

Aime tracks 6 emotional aspects based on the language being used:

1. Happy
2. Tender
3. Excited
4. Sad
5. Angry
6. Scared

Based on the above information, Aime will classify your emotional relationship with others in one of four categories:

• Close  –  someone you are close to, such as a good friend or family
• Supportive  –  where one person encourages positive emotions in someone else
• Reflective  –  where one person tends to reflect the emotional  state of someone else
• Distant  –  more formal relationships, for example with work colleagues, especially those in a more senior position

The language Aime uses, and possibly the mode in which Aime replies depends on the emotional context of a single message and the long term emotional relationship .

Aime’s Time sense and Memory

At the moment Aime doesn’t read your calendar, because you may use many different types of calendar system. Instead,  she constructs her own calendar based on events she sees described in your messages. In other words she has a memory. She doesn’t remember things forever, and she will forget events on a sliding timescale from a year down to a week depending on her assessment of how long she needs to remember.

Fantastic question. Artificial Intelligence seems to be everywhere these-days; from systems that can turn on the lights in your lounge when you ask them, to applications helping diagnose illness, to systems that can book an appointment at the hairdressers for uncle Norman (let’s hope he isn’t bald). But AI isn’t a new concept, some of us have been thinking about it for over 30 years.

I still haven’t addressed the question, what is AI? The circular answer is that it is anything which utilises Artificial intelligence techniques in order to achieve a task which computers traditionally find quite difficult. Examples are seeing objects in a changing environment (where did I put that rake?), and understanding human languages – whether spoken or written. The AI techniques that allow you to try to do these things haven’t really changed, what has changed is the capacity and connectivity of the computers we run it on. Even a humble mobile phone can run an AI system. Or so I have heard.

So are AIs intelligent?

The short answer is ‘yes’. The medium answer is ‘maybe’, and the long answer is a big fat ‘no’.

Yes: if a system displays enough intelligence to be able to save someone’s life in remote Africa by giving them a good diagnosis for their illness when a doctor is two days walk away, then who cares whether it ‘understands’ what ‘early stage cancer’ actually is, or not. We don’t question how humans make decisions (ok, some people e.g. in the US and UK definitely question ‘why’, but that’s not the same). If a decision is a good one , then its a good one.

Maybe: Inside most computers its just 1s and 0s. How can 1s and 0s be intelligent? Well there is the concept of emerging behaviour. If you stick enough little bits together they can achieve what looks remarkably like intelligent behaviour. Termites can build, what in their terms, are mile high cities where they live a cooperative and harmonious life. These mounds contain farms, roads and enough sophisticated ‘air conditioning’ to allow them to survive in very hot locations. yet if you  try to look at a single termite brain for the seat of this intelligence you would have to have a pretty good microscope. Maybe lumping enough appropriate AI components together in the right way will eventually result in something that everyone agrees is intelligent. Just for Pete’s sake make sure it understands emotional context. What are you doing Dave?

No: The best AI’s on the planet are good at one thing. Taking part in quiz shows, playing Chess or Go, playing your ‘baby changing’ playlist whilst you try to extract a baby wipe with one, moderately clean, hand. To be a true AI under this test requires sentience (awareness of its own being) and an ability to do most of the stuff humans find easy (possible exception of changing a baby).

My 2p worth

I’m a pragmatic optimist, so its a ‘yes/maybe’ from me. If it barks like a dog and looks like a dog, we may as well treat it as a dog. Otherwise we are inventing a new form of solipsism. Emergent behaviour is a very real thing – the whole universe as we experience it is very likely various forms of emergent behaviours. Lastly intelligence and sentience are really two different things ( thank you to Peter Watts for opening my eyes (see Blindsight-you may have to read it a few times before you get it ) )

If we can have intelligence without self awareness – a uniquely human conceit to believe we can’t – then that makes the whole question easier. We only have to look at ourselves closely to realise how much of our daily life is handled by our subconscious, for deciding we are hungry, to driving home when tired and then not remembering how we got there, to going to sleep on a complex problem only to find we have the answer miraculously when we awake. Consciousness is vastly overrated.

For me, the quest for AI helps us to realise what it is to be human, and hopefully in knowing ourselves better, we can make better choices for the future.

 

Tisquantum was the name of the native American who welcomed the Pilgrim fathers to the new land. He already spoke English, and acted to guide and assist the settlers to survive and eventually flourish in their adopted home.

Assistance to those in need, and facility with language are at the core of the services our company provides.