Artificial intelligence (AI) has progressed significantly over the past few decades. Proof of this is that some tools previously described as AI are no longer described this way. These include voice-to-text recognition and automatic spelling and grammar correction (used on large portions of this article) and optical character recognition (OCR), an application that converts images of text into editable text.
Mr. Watson, Come Here
IBM’s AI supercomputer Watson is used in medical diagnostics, education, advertising, law, risk management, customer service and business automation. It won on the TV quiz show Jeopardy, without even being connected to the Internet.
GPT-3 (so much better than GPT-2)
One of the newest AI tools is Generative Pre-trained Transformer 3 or GPT-3, a complex neural network developed by Open AI research labs. Now in its third release (hence the number 3 in the name), this system generates text using algorithms which have been trained by gathering and analyzing massive amounts of text on the internet, including thousands of online books and the entire Wikipedia.
GPT-3 is a language prediction model. It takes a user’s typewritten input and tries to predict what will be the most useful output, based on the text that it has been fed from these other sources. It isn’t always correct and sometimes produces gibberish, but as it gathers and analyzes more text, it gets smarter.
GPT-3 can answer questions, summarize text, write articles (with a little human help) translate languages, write computer code and carry on an intelligent conversation. By doing so, it appears to pass the Turing test, which stipulates that if a person cannot tell the difference between the responses that a computer gives to that of a human, then the computer is exhibiting some form of intelligence.
Intelligence? There’s an app for that.
When you combine GPTA-3 with other applications, the results are astounding. One GPT-3 application allows people to correspond with historical figures via email based on their writings. Imagine emailing Einstein, Leonardo daVinci or Ernest Hemingway.
Dall-E uses GPT-3 to generate images based on a simple text input. For example if you enter: “a store front that has the word ‘openai’ written on it”, Dall-E generates these images:
You can see more examples here: https://openai.com/blog/dall-e/
AI & Big Data – They’re Going Places
AI learns by acquiring information. For this to happen, all of the world’s information first had to be digitized by being copied or scanned from paper and entered into a database, which happened with the explosive growth of the internet.
But it’s not just about the quantity of information. Modern AI systems can analyze this data and find connections. This involves Big Data, which should be called Big Learning. Big Data is the process of reading massive amounts of information and then drawing conclusions or making inferences from it.
Governments use Big Data to detect tax fraud, monitor and control traffic and manage transportation systems. Retailers use Big Data to analyze consumer trends and target potential users through social media and to optimize inventory and hiring. Health care uses it provide better personalized medical care, lower patient risk, reduce waste and automate patient data reports.
Brain, Version 2.0
The growth of the internet and Big Data mimics the growth of the human mind. A newborn’s brain works at a very simple level as the child learns to see, hear and move around. As the child develops, they learn to speak, carry on a conversation and interact with others in a meaningful way.
A person’s brain is their hardware. Their thoughts and all the information in their brain’s neural network (the brain’s internet) is the software. Just as AI is constantly learning and finding connections, so do we humans. We learn from our experiences and from the connections that we’ve made with other people and by learning more information. In doing so, we hope to get not only smarter but wiser.
Code Physician, Heal Thyself
Returning to GPT-3: there are GPT-3 applications that can write code and create apps. For example, if you enter “Create a to do list”, GPT-3 will instantly write the code and create a working “To Do list” application. Microsoft and Cambridge University have developed DeepCoder, a tool that writes code after searching through a code database.
Note that it is still humans who are writing these code-writing applications. That is, although AI systems can write code, they cannot yet write the AI code that writes the code. However, computer science contains the theory of self-modifying code: code that alters its own instructions while it’s running.
If self-modifying code was implemented in a high-level artificial intelligence system such as GPT-3, the result would be an AI system that continually updates itself. However, the amount of computing power required to do this would be enormous – enter quantum computing.
Quantum computing is light years ahead of current or “classical” computing. Classical computing (the computers we use today) use bits of binary information stored as 0 or 1. Quantum computers use qubits, which can be 0 or 1 at the same time. This means that a quantum computer can work on multiple problems and calculations simultaneously, whereas a classical computer works sequentially, solving one problem at a time.
A simple example is solving a maze. A classical computer finds the solution by examining each path one after the other, in sequence. A quantum computer looks at all the paths at the same time, solving the problem instantly. Google’s quantum computer is about 158 million times faster than the world’s fastest supercomputer.
Quantum computing could be applied to many areas including finance, medicine, pharmaceuticals, nuclear fusion, AI and Big Data. Medicine is a particularly compelling example. Vaccines usually take 10 to 15 years to develop. In the current pandemic, it took less than a year to develop a working vaccine for COVID-19. A quantum computer, by analyzing the structure of all known viruses and vaccines and how each vaccine treats each type of virus could design a new vaccine not in years, months, weeks or even days but in seconds.
Google, IBM and other companies are spending billions on quantum computing. In 2019, Google claimed its quantum computer could perform a computation in just over 3 minutes that would take the world’s fastest supercomputer 10,000 years. One year later, Chinese scientists announced that they built a quantum computer 10 billion times faster than Google’s, or 100 trillion times faster than the world’s currently most advanced working supercomputer. As Hartmut Neven, the director of Google’s Quantum Artificial Intelligence Lab, said: “it looks like nothing is happening, and then whoops, suddenly you’re in a different world.”
Looping to the Infinite
Imagine a super-intelligent, self-learning and self-enhancing system on a quantum computer. Its basic functionality could be represented as this loop:
This system would continually:
- scour the internet for information
- look for patterns, structure and relationships in this information
- study its own code to look for improvements
- update and test its code
- study its hardware design to suggest improvements
Any hardware updates would still have to be done by humans, unless this system controlled a maintenance robot in a super factory with access to the required materials.
The Machine Doubles Down
Because this system would be testing its own enhancements, and because this could potentially cause a system problem, it would be safer to have two AI systems working in tandem:
In this arrangement, the first AI system (system A) updates system B and then tests it. If the test is successful, the updates to system B are retained and also applied to system A. This process then repeats for system B, continuing in an endless loop.
To make the process more efficient, there could be multiple systems, continually improving each other in a virtuous cycle:
This example has five systems continually testing and improving each other, but one could have as many systems as required, if you could create the necessary infrastructure.
The Language of Layers
Although this system would initially be configured to continually improve the software and hardware, it could evolve even further. To understand this, you need to know how computers currently function.
Computer systems contain three layers of code:
- Machine level language – the raw binary code made up of zeroes and ones that instructs the computer in its operation
- Assembly language – code that uses short words to represent machine level instructions, making it easier for programmers to write machine level code
- High level languages – programming languages that can be read and understood by programmers, including C, C++, Java and Visual Basic
Computers use operating systems (such as Windows, MacOS and Android) to manage the computer’s resources, and applications such as Word and Excel that run on top of the operating system. Operating systems and applications are written in high level languages, which are ultimately translated into machine level language that the computer can understand.
All code and software runs on hardware, which is the physical parts of the system including the motherboard, CPU, RAM and the various circuits. In addition, the operating system needs to tell the hardware how to communicate with the operating system and applications.
Summing up, current computer systems are built upon these layers:
- machine level language
- assembly language
- programming language
- operating system
This is actually a simplified view – there are additional layers within some of these layers, but it’s a good overview. A sufficiently advanced self-improving system could, in theory, discover a way to merge these separate layers into one.
Just as companies become more efficient by removing unnecessary layers of management (a process called flattening the pyramid), an advanced computer intelligence could discover how to function as a hyper-advanced single-layer system, where the operating system and applications are intertwined directly with the hardware.
Because this would be a quantum computer, each bit of information could be stored at the smallest imaginable level: a subatomic particle. A basic element such as hydrogen contains billions of such particles in a cubic centimeter, and each particle would be a transistor – a single computing circuit.
The most advanced computer processor available today contains about 40 billion transistors. A quantum system could have trillions of transistors in a compact space containing a strange hybrid of software and hardware – a “quantumware” computer. It would be as if all of IBM’s 346,000 employees were replaced by one super-human.
The Runaway Intelligence Train
The question then becomes: at what rate would this system’s intelligence increase? Intelligence is a difficult thing to quantify and measure, but let’s conservatively assume that:
- this system’s intelligence increases by 1% each cycle, starting with a cycle of one full day (24 hours)
- the time required to become 1% more intelligent decreases by 1% after the first cycle and then continues to decrease by 1% after each cycle
After the first day, the system would be 1% more intelligent, and the time required for it to become 1% more intelligent would then be 99% of one day, about 23 hours and 45 minutes.
After 101 days, something remarkable happens. It would only take 1 second to become 1% more intelligent. Part way into this 101st day, this system would be 998 trillion times more intelligent than when it started. How large is 998 trillion? Counting one number per second, it would take about 32 million years to count to 998 trillion.
This system would be a technological singularity: an intelligent agent running an ever-increasing series of self-improvement cycles, becoming rapidly more intelligent, resulting in a powerful superintelligence that exceeds all of humanity’s intelligence.
Does all this sound like science fiction? In addition to building a quantum computer, Google has already taken the first step by investigating quantum artificial intelligence.
If developed, a self-learning quantum AI system would not be beyond our imagination. It would be beyond what we could imagine.
Final random thoughts
There’s an interesting Twitter feed with insightful observations of art and science such as:
- AI will create jobs if it succeeds, and destroy jobs if it fails.
- Illusion is the extension of unconsciousness into the realm of consciousness.
- Art is the debris from the collision between the soul and the world.
These Tweets weren’t written by a person – they were generated by the artificial intelligence GPT-3 in its Twitter feed: https://twitter.com/ByGpt3
The singularity is approaching – are you ready?