Breeze Ripple

“Intelligence is the ability to adapt to change” Stephan Hawking

“I know that I am intelligent because I know that I know nothing.”  Socrate

Intelligence is the ability to solve problems,  learn and adapt to our environment.  Artificial intelligence is when machines can do that.

People believe AI is a machine which thinks like a human. It is not. AI is not a precise simulation of the human mind. And how can it be?! Human brain is still one of the biggest mysteries in science. Still, AI is an advanced technology which may surprise humans. Unlike traditional programming where the logic is implemented by the programmer, AI algorithms learn and adapt by themselves. They may produce outcomes no one expects and abstract the rules without being programmed.

We differentiate between two types of artificial intelligence—weak/narrow AI vs. strong/general AI. 

Narrow AI (ANI) is the AI we see, use and develop today. It is trained to perform specific tasks. It is the AI behind Siri by Apple, Alexa by Amazon, Watson by IBM, self-driving cars, etc. Strong AI, on the other hand,  is dominant in sci-fi movies and novels. It is also known as Artificial General Intelligence (AGI). In that form of AI a machine’s intelligence equals human intelligence; it would be self aware and conscious. Yet it exists only in theory. 

Although Artificial Intelligence sounds very fancy, it is actually quite an old discipline. It was a theoretical field at first, as computers were not powerful enough to support the AI algorithms. Alan Turing is considered the founder of AI. In the 1950s, after creating a machine to crack Enigma’s codes during WWII, he assumed the possibility that a machine could be intelligent. Turing asked the fundamental question “Can machines think?” and subsequently came up with a test to validate machine’s intelligence – the Turing test or the imitation game.  This brilliant mathematician proposed an idea much ahead of his time – that someday machines can resemble the human mind, even teach and learn from one another! 

With the increasing computational power, computing speed and the data available today, AI managed to get out of the academic papers and research labs. Today it is available to everyone, everywhere. People interact with it regularly through face recognition technologies, recommendation engines, chatbots and personal assistants.

AI has evolved from a concept in the lab through a psychotherapist chatbot (ELIZA), board games master(DeepBlue, Watson, AlphaGo), personal assistant(Siri) to a writer and designer(MidJourney, ChatGPT). 

My belief is that it is our chance to solve some of the world’s toughest problems in climate change and healthcare. AI or also known as ‘cognitive computing’ for me is fascinating. Below I will walk you through some of the most prominent examples of AI software and how they work. 

ELIZA – the first chatbot ever. The software was very simple even for its time –  referring to a list of keywords it used pattern matching to substitute words in its response to make it look plausible.  The funny thing here is that it did not have any level of understanding and still many people believed it did. Even its creator  – Joseph Weizenbaum – wasn’t able to convince the society in the 60s that Eliza is not that sophisticated and does not understand them. ELIZA is an important part of AI history although it is not an example of AI. There isn’t any learning or adaptation involved. But it shows people’s tendency to prescribe consciousness and intelligence to a piece of software. 

Deep Blue (IBM,) – the chess master. It beated Garry Kasparov in 1997. Back in those days AI was less based on machine learning, it used a symbolic approach. The good old-fashioned AI as symbolic AI is also called was dominant at that time due to computing power constraints. Symbolic AI is based on rules and if-then statements, the human knowledge is encoded as symbols. As you can guess, such software would be extremely specific.  In contrast, modern AI based on machine learning would use raw data and the machine would look for the knowledge it needs to solve the problem. 

Watson (IBM, 2011) -the “Jeopardy!” TV game winner is such software. It uses early machine learning algorithms. It is fed with a lot of data, for example from Wikipedia, to gather knowledge about the world. Jeopardy! requires not only a breadth of knowledge but a mastery in human communication.. The clues in the game are often vague and metaphoric, so Watson must make sense of complex nuanced English, context and syntax and generate a quick response. Watson is deployed across several IBM servers and runs parallel processing to achieve speed. In its core it is a question-answering technology based on severe data analysis and statistics. Although some of its  skills would rarely be useful outside Jeopardy! It has applications in healthcare(digging through medical records and assisting doctors with answers), entertainment, retail, finance.  

AlphaGo (DeepMind, 2015) – the game of Go champion.  AlphaGo relies heavily on machine learning –  deep learning, convolutional neural networks, decision tree search, supervised and reinforcement learning. Go is considered to be one of the most complex games ever devised by humans. More complex than chess. Why? The number of possible moves and combinations is immense. Alpha Go uses two neural nets – one to predict the next best move and another to evaluate performance. The first network called “the policy network” learns from a combo of human and computer generated data.  It goes through supervised training looking at numerous human games and plays mimicking human moves. Then plays against versions of itself and learns from its mistakes aiming to choose the best moves which would lead to a win. Such learning where the goal is to maximise the reward is called reinforcement learning. The other net, the “value” network, estimated the winning probability without going through exhaustive search – it evaluates if there is a chance to win or not and if not, the software skips that game. This speeds up the training process. AlphaGo shows that computer power and a smart algorithm can achieve a lot.

AlphaGo Zero(DeepMind, ) Unlike AlphaGo, AlphaGo Zero isn’t trained with any human data . It only learns by self-play. And that is what makes it one of the most important advancements in AI – it is machine learning which needs zero human knowledge as input to learn. AlphaGo Zero uses one neural network which relies entirely on reinforcement learning.  Fed only with the basic rules of the game it generates data by playing against versions of itself. It also uses a simpler search tree than Monte Carlo. As a result, AlhaGo Zero needs less computing power than Alpha Go. AI is no longer constrained by the limits of human knowledge. AlphaGo Zero shows that computational power and data is not all that is important in machine learning, the brilliance of an algorithm is.

Alpha Zero (DeepMind, ) Alpha Zero is a step towards more general AI. It can play not only Go, but also chess , shogi (japanese chess) and Atari from scratch. Again, it doesn’t need any human data as input and it is entirely trained with reinforcement learning. It is not taught any rules and learns from experience. 

AlphaFold (DeepMind, ) One of the most fascinating examples of AI for me is AlphaFold. It solves an essential and old problem in biology – protein folding. AlphaFold and its upgraded version AlphaFold 2 (2020) predict protein structure taking the sequence of amino acids as an input. 

For those who are not familiar with the protein folding problem: At first proteins are just a sequence of amino-acids. At some point they fold to take up their real role in the body – be it form a muscle, an organ such as pancreas, or the immune system. The amino-acid sequence determines the path the protein will fold and the form it will take. In short, protein’s structure defines its function. Knowing how proteins will fold would immensely accelerate drug discovery, medical research, our understanding of diseases and of the natural world as a whole. The challenge is that there are way too many possibilities and time will not be enough for even the fastest machine to compute those. Brute force and extensive search is not the answer. But where human power failed, machines succeeded.

AlphaFold 2 relies on deep learning techniques. The algorithm is trained with data from Worldwide Protein Data Bank which stores all known protein structures. The amino acid sequence of our protein in question is fed into AlphaFold which then looks through Multiple Sequence Alignments (MSA) all similar sequences among species in all available databases with known proteins. It then produces a 3D form of our protein and assigns a confidence score to its predictions. AlphaFold 2 contributed to expanding protein structure knowledge from 17% to 98%! Just note that AlphaFold2 can be only run in a lab as it requires immense computing power.

Honestly, I would like to hear more on how the Alpha Fold 2 application is going, how it contributes to healthcare and biology. Please, share with me if you have any news on the topic. 

Society is now divided regarding AI development – some look at it with the greatest hope and believe it may solve the world’s major problems in areas such as climate change, healthcare, agriculture, etc. Others see it as a threat to humanity – leaving people jobless, leading to unthinkable scams.. I guess both of these scenarios are possible but we should work hard towards the optimistic one while keeping in mind and protecting ourselves from the more pessimistic possibility. 

What lies in the future is how to ensure responsible use of generative AI. Companies are now teaming up to build regulations although critics doubt the objectivity of such regulations. Most probably they will favour the big players as the cookie policy and ads did. But still some are better than none. And it is up to society and every single person to stay aware.