Artificial Intelligence History, Toby Walsh, 2025

Featured Image: An artificial intelligence ChatGPT created, Ghibli Studio styled representation, of an old bushwalking photo taken at Square Rock ACT. The request took seconds.

  Breadtag Sagas ©: Author Tony, 22 September 2025

This is article two in a series on AI. The first article in the series is  1 Artificial Intelligence Explained.

Toby Walsh The Shortest History of AI 

Reflections on The Shortest History of AI by Toby Walsh, 29 April 2025

Main Points

• Introduction to Toby Walsh
• Pre-history of AI to 1987
• The Symbolic Era 1956 to 2011
  • Searching for Answers
  • Making the Best Move
  • Following Rules
• The Learning Era 2012 to Now
  • Artificial Brains and LLMs
  • Learning by Reinforcement
  • Uncertainty and Navigation Revisited
• Conclusion

1 Preamble

I attended a meet the author discussion with Toby Walsh on his book The Shortest History of AI, April 2025, with Andrew Leigh (an ACT Labor MP and author of The Shortest History of Economics, February 2024).

Andrew Leigh in his book says one reason the industrial revolution lasted such a long time is that it took almost a century for people to properly harness coal.

The same pattern can be seen with other general-purpose technologies. …

When it comes to coal power, electric motors and computers, general-purpose technologies underwhelm in the short run and dazzle in the long run.

This is true of artificial intelligence which has had its kick-starts and AI winters or downturns. I first became excited by AI in the 1980s with the expert systems boom. And, because of this (nostalgia perhaps) I am still disappointed that the potential of expert systems to improve our lives has barely been explored.

2 Introduction

My first article Artificial Intelligence (AI) explained was my attempt to simply overview the field, provide some history, definitions and explanations.

This second article is focused on Toby Walsh a contemporary pioneer of AI in Australia, who was asked to write a book in the shortest history series.

I mentioned another Australian researcher Rodney Brooks in my first article. Brooks headed up the Computer Science and Artificial Intelligence Lab at MIT, became a world-renowned robotics innovator and helped found the incredibly popular commercial robot vacuum cleaner manufacturer iRobot.

Toby Walsh has had a similarly brilliant career in the once small world of artificial intelligence.

Talking about his career beginning in the late 1980s Toby Walsh says in self-deprecation:

It was at this point I graduated from university and entered the field of artificial intelligence. Lucky me! This second AI winter lasted two decades. Lucky me! But eventually the sun came out again around 2012, as the machine learning era began.

2.1 Toby Walsh

Toby Walsh is a British computer scientist in AI, with a BA from Cambridge, an MSc and PhD from the University of Edinburgh, and is domiciled in Australia. He is one of the world’s leading researchers in AI.

Toby Walsh 2008, by Avobronte Wikimedia

Toby Walsh knew some of the pioneers of the field personally, for example, John McCarthy, but as he says it was a very small field until quite recently.

The purpose of the current article is to provide a summary of the AI history and experience outlined in Artificial Intelligence (AI) explained through the eyes of an insider, who lived through it over the last forty years.

These two parallel articles have some overlap, but a much different focus and provide a synergy that will provide more enlightenment on something that is taking over our lives for good and ill, and with a very human lack of planning and control.

3 Pre-history of AI to the second AI winter in 1987

Toby Walsh’s timeline of AI resembles mine in the article mentioned above, but he emphasises different aspects and events. I am copying his table of events below (without permission). If you enjoy this article, I’d strongly recommend Walsh’s book for further reading.

Toby Walsh Table 1A

Toby Walsh Table 1B

Some things missed in my article Artificial Intelligence (AI) explained, which are useful in this one are:

3.1 A chess program was written by Alan Turing and his friend David Champernowne in 1948 called Turochamp. Turochamp only looked two moves ahead, but was not capable of being played by a computer for years, until the technology had developed sufficiently. Turing used a pen, paper and manually calculated the moves.

Alan Turing 1951, photo Elliot and Fry, NPG London

3.2 The Logical Theorist by Allen Newell and Herbert Simon 1955 was a very early mathematical AI program designed to prove mathematical theorems. In fact it proved 38 of the first 52 theorems in Principia Mathematica by Bertrand Russell and Alfred North Whitehead. This brilliant program was first showcased at the 1956 Dartmouth Workshop.

3.3 AI began on 18 June 1956, the beginning of an eight-week long workshop at Dartmouth College whose goal was to build intelligent machines and was organised by John McCarthy.

3.4 DENDRAL 1965, the first expert system, was designed by a talented and visionary multidisciplinary team at Stanford University. DENDRAL was designed to suggest possible chemical structures from mass spectra. A mass spectrograph defines the mass of various chemical elements making up a compound.

3.5 In 1968 Shakey the robot and robot navigation were created (also at Stanford). In 1968 A* search: was the algorithm to navigate Shakey. In 1971 the STRIPS algorithm was developed to solve goals for Shakey. In 1986 SLAM, simultaneous location and mapping method, was developed to enable any robot to map and locate in its environment, a much more advanced approach.

Let’s move on to how Toby Walsh outlines the field.

Eras and Ideas

Toby Walsh divides AI history from 1956 into two eras and six ideas. The first era the Symbolic Era from 1956 to 2011 is covered under three ideas. The second era the Learning Era from 2012 to now is also covered under three ideas.

The second era is more complex for us — the layperson — to understand.

4 The Symbolic Era 1956-2011

We’ll skip the detail and briefly summarise the ideas.

4.1 Idea #1 Searching for Answers

You can reduce many problems to searching for an answer. This seems ridiculously simple. But in AI or computer terms, it can mean searching its internal representation of the world for the symbol representing a starting state to the symbol representing a goal state. This isn’t new. It’s called navigation.

A* search was invented in 1968 to navigate a robot called Shakey (see 3.5). Shakey was aptly named. It was a robot with a camera to see the world, a microphone for instructions, wheels and motors to move about and an onboard computer to make decisions.

Shakey the Robot 1969, Computer History Museum, Wikimedia

Don’t dismiss A* search. Every time you use your phone for directions using GPS there’s a little AI program using A* search to find the shortest route.

The Logical Theorist (3.2) can be taken up a level from the original idea of Shakey navigation using much bigger (logical) maps. The Logical Theorist didn’t just prove theorems from Principia Mathematica, it actually found new and in some cases shorter proofs for some of these theorems.

STRIPS (3.5) inspired by The Logical Theorist stripped problems down in clever ways to form logical propositions, which could then be built upon.

4.2 Idea #2 Making the Best Move

In competitive settings, you should choose your best possible move, assuming your opponent chooses their best possible move.

Noughts and crosses (tic-tac-toe), checkers, chess. The first examples were very simple games with a perfect strategy defined by an algorithm (noughts and crosses). After checkers, chess was the first complex game to be mastered by computers.

From the 1980s as computer memory and processing speed grew, so did sophisticated chess playing programs. These programs initially used strategic techniques such as game trees, evaluation functions and the like, and later vast databases of games.

IBM’s Big Blue eventually beat Garry Kasparov the world champion in 1997 but more by brute force (processing power) than by new techniques. More current programs post the millennium (e.g. Stockfish, a free open-source chess engine) use neural networks that make them immensely more powerful.

4.3 Idea #3 Following Rules

In narrow domains you can simulate human expertise by following rules. This can be labelled expert systems. DENDRAL was immensely powerful in assisting to define the structure of complex chemical compounds from mass spectra. A potential later, but much more difficult development would be to define the shape of complex compounds such as proteins. A much more complex AI affirmation of DENDRAL type programs is AlphaFold 2018 designed to elucidate the folding structure of proteins.

Expert systems unfortunately hit bottlenecks in the late 1980s leading to the second AI winter.

The Learning Era 2012 to Now

5 Artificial Brains and LLMs

5.1 Idea #4 Artificial Brains

You can copy the human brain with a network of artificial neurons that learn from experience. The idea had been around for a long time, since Turing. An artificial neuron theory called perceptrons was developed. Complex issues arose around the perceptron theory. A book by AI pioneers Minsky and Papert in 1969 on perceptrons showed the impossibility of just one single layer learning a concept as simple as even or odd. Unintentionally perhaps, the book had a negative and undue influence on the field.

Geoffrey Hinton 2024, Nobel Prize Laureate in Physics, by Arthur Petron

One maverick researcher Geoffrey Hinton continued doggedly on the topic through the second bitter AI winter. He did this despite negativity through various institutions and countries.

Hinton showed that an intermediate or ‘hidden layer’ could solve the problem of odd and even above and also other more complicated functions. But Hinton didn’t stop there. He popularised the idea of deep learning using neural networks with many layers.

The next issue Hinton addressed was the speed of this learning. He and others developed backpropogation as a smart way to increase speed. Backpropogation updates the weights (weightings) in a deep network with many layers quickly. Thinking in multi-dimensions was also useful.

Toby Walsh says amusingly of Hinton that if you wanted to imagine fourteen dimensions, Hinton’s advice was to imagine three dimensions and then say to yourself repeatedly fourteen in a loud voice.

Simplified Scematic of a Neural Network

A chance discovery in 2012 by one of Hinton’s PhD students found that GPUs (graphic processors) used to render 3D video were ideally suited to perform the matrix calculations needed to compute gradients and to do backpropogation.

The discovery would quickly add over a trillion dollars to the market capitalisation of Nvidia Corporation, the company with an 80% share of the GPU market.

Hinton needed a third and final ingredient to get deep learning to work well and that was training data. Fortunately, another researcher Fei-Fei Li had come to the same conclusion in 2006. She wanted a computer to identify objects, e.g. an apple, a banana, a bicycle. She obtained funding to pay people to label images and developed ImageNet. By 2009 ImageNet had over 3 million images and was released to the world. It has grown to over 14 million (legal) images, all helpfully labelled into one of 21,000 categories.

5.2 Transformers

To get to AI chatbots like ChatGPT (a large language model) one needs another ingredient. The ‘T’ in GPT stands for transformer. Transformer is another clever way to simplify the programming effort. Unlike clever ways to handle the two-dimensional nature of images, transformers work especially well with one dimensional or sequential data such as text, weather, stock prices, protein sequences, music and speech.

Let’s stick with text. In a transformer step one is to turn words into numbers or tokens, which can be understood as sequences in a sentence. But, certain words have complex relationships unrelated to sentence structure. For example ‘queen’ is near ‘king’ but ‘queen’ is also near ‘bee’ and ‘bee’ is near ‘beetle’. Transformers uses word vectors to identify these more complex relationships.

The second step is called encoding turning the numbers into vectors, that is, something with many potential dimensions.

The third step adds more magic by taking the multi-dimensional space the vectors have created, but now also taking into account the relationship between the words to uncover the connection between words in a body of text.

The fourth and final step is decoding. This transforms vectors back into tokens. If the decoder simply inverts the encoder, we get another English sentence; but if the decoder is for Russian, we get a translation of the English sentence into Russian.

(Toby Walsh tries to explain this in more detail, but we only need a general sense of what is going on.)

5.3 General Purpose Technologies: ChatGPT or LLMS

Open AI the developer of ChatGPT was founded in 2015 (as a not-for-profit). In its early affirmations, it and other LLMs (large language models) existed in the shadow of Deep Mind (see 6 below). Open AI was intent on scaling LLMs with increasingly large sets of data.

Sam Altman of Open AI 2019, by Steve Jennings for TechCrunch

The first LLM in the GPT family was GPT-1 released in June 2018. The neural network had an impressive 17 million weights (or, as they tend to be called now parameters). GPT-1 was trained on the Toronto Book Corpus, a dataset of around 7000 self-published books (with permission).

GPT-2 was released in February 2019 with 1.5 billion parameters (2 orders of magnitude larger) on a dataset of 8 million web pages. GPT-3 was released in June 2020 with 175 billion parameters (another 2 orders of magnitude larger) and was trained on over half a trillion tokens scraped from the internet. Open AI’s fourth model GPT-4 was released in March 2023 with about 1.76 trillion parameters.

After some dramatic and controversial organisational changes, part of Open AI became a for profit organisation. Open AI is no longer open about the details of its discoveries, supposedly for commercial reasons.

GPT stands for Generative Pre-trained Transformer (an ugly name). Toby Walsh suggests that the name was also perhaps consciously chosen to remind us that they are trying to build a general-purpose technology. Walsh says: general-purpose technologies is a technical term in economics that impacts the whole economy, with wide ranging financial and societal consequences (as mentioned in Section 1).

Open AI says GPT-4 cost $100 million to train. In November 2022, Open AI made a master move according to Toby Walsh and built a chatbot called ChatGPT over the GPT family and released it to the public.

ChatGPT has been an incredible and popular success, but there has been some dissension in the ranks in Open AI concerning the direction of the company. As with Google (‘Don’t be Evil’), the early aspirations of Open AI have been hard to live up to. Certainly, Open AI requires massive injections of cash to keep progressing and investors aren’t altruists.

5.4 Technical Issues with LLMs

In AI, scaling (as in the massive increases in scale from ChatGPT 1 to ChatGPT 4) is usually a problem, with researchers frequently running into combinatorial bottlenecks. Walsh says we are used to scale making problems worse, not solving our problems. Fortuitously with GPT and other LLM applications the larger the dataset, the better the performance.

Nevertheless, despite the marvellous performance there are still issues and these are not solvable at present.

Toby Walsh says: The overall idea behind ChatGPT is simple. It’s like auto-complete on your phone but on steroids…

The limitation he says is that LLMs are statistical models. They work on probabilities. They’re not understanding the language, as you or I do. They have no underlying model of the world.

He goes on to outline some simple problems. Most models he says struggle with reasoning. He gives an example of where ChatGPT can’t provide sentences with words of one vowel. When asked, it repeatedly produces sentences containing some two-vowel words.

When asked, what was the first elephant to swim across the English Channel, ChatGPT responds with plausible nonsense. When Toby asks it about a book he wrote, it is even worse. This is again bullshit! he says.

Friends who use Google search often comment that the AI summary is frequently off the mark, either vague or rubbish.

The fact that ChatGPT hallucinates is no surprise, it relates to the probabilities and it is not going to go away. It’s even something you wouldn’t want to disappear completely, because creativity in ChatGPT depends upon it.

Racism and other negative aspects of LLMs is widely covered in the press. These are an example of biased data and hallucination, but they require fixes to remove them, which may create worse problems.

In this context, it was a very brave move for Open AI to launch ChatGPT knowing full-well it would make things up. There have been other terrible examples of premature launches. Walsh cites Meta’s Galactica LLM trained on a high quality, curated corpus of humanity’s scientific knowledge, which was released on 15 November 2022 and taken down two days later because it was wrong and biased, but it always sounded authoritative. (Perhaps, an inability to reason was the problem. There is a real danger with accepting AI uncritically and as humans we have a predisposition to treat authoritative statements with respect.)

6 Learning by Reinforcement

6.1 Idea#5 Rewarding Success

This involves reinforcement. You can learn from experience, rewarding success and penalising failure.

The idea is more complex than it sounds. How do you assign credit in a chess game? You may win, but which moves led to success. There may have also been poor moves that impeded it?

6.1.1 MENACE

Donald Mitchie, a colleague of Alan Turing at Bletchely Park, in 1960 built MENACE (machine educable noughts and crosses engine) to win a bet with a colleague that it was possible to build a learning machine. He used matchboxes, beads and grids drawn on paper.

MENACE was made up of 304 matchboxes, each representing one unique arrangement of the noughts and crosses grid of which 304 states were possible for the starting player. Each square on the grid had a different colour. Each matchbox contained nine coloured beads.

Recreation of MENACE built in 2015, Wikimedia

A bead would first be removed at random from the matchbox representing the current state of the grid. The colour would determine where MENACE would place its next nought.

At the end of the game, if MENACE had won, the coloured beads were returned to their respective boxes with three bonus beads of the same colour to reward the successful play. If MENACE had lost, no beads were returned at all, penalising the moves that had led to the loss. After a few hundred games MENACE played noughts and crosses perfectly.

(You don’t have to understand in detail. I can’t visualise it completely.)

6.2 Deep Mind

6.2.1 Atari Games

Three geniuses founded the next step a company called Deep Mind. Two Demis Hassabis and Mustafa Suleyman met at grammar school in North London. Shane Legg a New Zealander met Hassabis at University College, London. They set up Deep Mind in 2010. The company had an ambitious but uncomplicated mission: to solve artificial intelligence and use it to solve humanity’s problems.

The first project of Deep Mind was learning to play games using reinforcement learning. They began with 49 different Atari video games such as Pong, Breakout and Space Invaders.

Demis Hassabis in 2025, by Christopher Michel

Deep Mind’s AI program was not given any background knowledge of the games, just access to the score and to the raw pixels. It didn’t know about bats, balls or walls. After several hours of playing itself, it learned everything from first principles.

Deep Mind’s AI was able to play most games at the level of a human player and in a dozen cases it played at superhuman level.

The results stunned the AI research community. As one expert said: That’s both impressive and scary… And, it is!

6.2.2 AlphaGo

The second breakthrough was with the game of Go as outlined in Artificial Intelligence (AI) explained. AlphaGo beat the world champion Lee Sedol in March 2016. Unlike IBM’s Deep Blue, Alpha Go began learning Go from first principles as in the Atari games. Many pundits had previously said that a computer would never be able to master the game of Go.

Lee Se Dol 2016, photo LG Electronics, Wikimedia

The breakthrough had many ramifications.

Move 37 in game two was a move no human would ever make and seemed premature, a mistake. However, many moves later it became apparent that it was a nearly perfect move. (Eerily, something similar thing happened in the second Game Gary Kasparov had with Deep Blue at move 37, but it was a computer glitch.)

For those in the know, the AlphaGo result was even more scary than the Atari result. Alien winds were blowing strongly.

The result stimulated China’s long-term ambitions in AI with a massive investment plan over 20 years. Other countries responded in kind, the USA, UK, Germany and even India. Sadly the Australian government didn’t do anything.

In January 2014, Google purchased Deep Mind for around $500 million.

Hassabis and Suleyman now head AI research at Google and Microsoft, respectively. Legg works for Google too.

6.2.3 AlphaFold

AlphaFold (mentioned briefly earlier) may not be as sexy, but is an equally brilliant use of learning reinforcement. Proteins are complex molecules that are the basic building blocks of all life. Biochemists spent decades determining the structure and folding patterns of proteins, as shape determines function. Nobel prizes and careers were determined by elucidating the structure of single important human proteins.

Mustafa Suleyman in 2018, by Christopher Wilson

Protein structure prediction was a good choice of problem for Deep Mind to tackle, as biologists had turned it into a game. Every two years there was a world championship competition. Before 2018, computational methods had limited accuracy. Deep Mind entered the competition beating 97 other entrants, with the most accurate structure for 25 out of 43 proteins. Second place scored 3 out of 43. In 2020 Deep Mind did even better.

Reinforcement learning plays an important part in AlphaFold. In 2022 Deep Mind did not enter the competition. Everyone was using open-sourced AlphaFold. Deep Mind turned its energies to serious prediction. In July 2023, it released a database of 200 million proteins for scientists to use freely. The beginning of a new era for science.

7 Uncertainty and Navigation Revisited

7.1 Idea #6 Reasoning About Beliefs

The world is uncertain, AIs need to deal with uncertainty. Fortunately, 250 years ago the Reverend Thomas Bayes developed a method of dealing with uncertainty. A radical method that has had mixed popularity over time is called Baysian statistics.

A chicken-and-egg problem is a good example of the wide-ranging application of Bayes idea. SLAM (simultaneous localisation and mapping) is an AI technique applicable to building a self-driving car.

You can put an expensive Lidar sensor on the car (a light-based version of radar), which will help to identify where obstacles are like pedestrians, bicycles and other cars. You use that and other sensors to map the world around the car, but you can’t locate yourself without an up-to-date map of the world (GPS, Google maps). You can’t do both at once. SLAM solves the problem by using Bayes idea to update the probable location both of the car and the objects in its map of the world.

Self-driving cars rely heavily on high definition maps and GPS. But, maps are always out of date and GPS can be blocked by tunnels and urban canyons; not to mention objects which may be unpredicted or unpredictable. On a Level of 1 to 5, self-driving cars are at level 3-4. But, to achieve level 5, full-automation is a major challenge despite promises. Self-driving taxis being piloted in US cities are only at level 4 (high automation) but the next step is enormous.

8 Conclusion

Toby Walsh is an expert AI researcher and has been involved in the field since the 1980s. His approach to the development of AI is quite different from my layman’s approach in Artificial Intelligence (AI) explained. I suggest both are equally valid and balance one’s perspective, but for different purposes.

Toby’s is an excellent insider perspective and insight into what is important in the field, from someone who has spent his adult life researching it. I was particularly struck by the distinction between LLMs (large language models) like ChatGPT (follow the rules) and Deep Mind (learning by reinforcement), which are so fundamentally different.

8.1 LLMs

We haven’t yet explored how LLMs can also manipulate images, create deep fakes or write computer code among other amazing things. At the moment let’s keep these at the back of our minds for later, but always remember that LLMs are amazingly versatile.

Yet, we also need to be always conscious that LLM AIs have no model of the world and no understanding of how language operates. This explains the hallucinations, mistakes and outright ‘bullshit’ that is sometimes produced.

There are other considerations also. It is frightening to think that LLMs are being introduced into government and other applications without comprehensive risk assessments.

I have stopped using LLMs for much writing. It was an interesting experiment but the general prose they come up with is always somehow lacking. However, for summaries, specific points or repetitive information, for paragraphs about and analysis of key issues in larger chunks of text, even for research into new topics they can be amazingly time saving and sometimes insightful. But, one should never take the information at face value. Everything needs to be checked.

I was surprised when I first checked Perplexity‘s citations, which I had thought an improvement on ChatGPT. I found that some of them did not refer at all to the information cited.

I can remember when librarians, teachers and others were overly concerned with veracity in Wikipedia. I have always used Wikipedia as my first point of call and mostly expected it to be accurate. I still do. But, up until now Wikipedia has always been written by humans and its editors (many of them volunteers) try their best to verify information.

ChatGPT and other LLMs may seem human (particularly LLM companions) but they aren’t. AIs are alien. They have no resemblance to us, or the way we think. They have no model of the world to date, but even when they do, it will be nothing like ours.

8.2 Deep Mind

Deep Mind (Atari games and Alpha Go) learned to play games from first principles, often lacking basic information. AI through reinforcement learning, rewarding success and penalising failure, has been shown to be frighteningly impressive but scary. There is no way to miss the alien winds blowing here.

Alpha Go was especially impressive. First, because both Go experts and computer professionals did not think Go would yield to computation power in the same way that chess had. And second, because of move 37. As Toby Walsh says: Conventional wisdom from thousands of years of human play is that you play on the fourth line — not the fifth — at such an early point in the game.

Despite the frightening potential, AlphaFold shows that AI can transcend hundreds, perhaps potentially thousands of years, of human effort in an incredibly short time and be amazingly efficient and useful.

There is always talk of productivity and how to improve it in analysis of countries’ economies. How would macro-economists rate AlphaFold’s productivity in this context?

8.3 Final Comment

LLMs and ChatGPT in particular are also thought to be a great improver of human productivity. Negative consequences tend to be ignored.

But, governments, legal systems and ethicists are so far behind in the chase that the use of AI will continue to grow without restrictions and probably without assessments of risk.

Many people have commented on how OpenAI through ChatGPT has bulldozed through copyright and other intellectual property issues without permission. But, has the horse bolted…. and does anyone know what the stable door looks like?

We live in a brave new world. Perhaps, there will be some backlash. Perhaps, some governments may develop regulations for the use of AI, but others won’t.

Key Words: AI, artificial intelligence, history, Toby Walsh, The Shortest History of AI, Symbolic Era, searching for answers, making the best move, following rules, Learning Era, artificial brains, learning by reinforcement, uncertainty, navigation, general-purpose technologies, AI winter, Alan Turing, David Champernowne, Turochamp, The Logical Theorist, Allen Newell, Herbert Simon, Dartmouth Workshop, John McCarthy, DENDRAL, Shakey, robot, A* search, STRIPS, IBM’s Big Blue, neuron, perceptron, Geoffrey Hinton, deep learning, neural networks, deep layers, backpropogation, Fei-Fei Li, ImageNet, training data, ChatGPT, LLM, large language model, Sam Altman, transformer, Generative Pre-trained Transformer, token, encoding, vector, multi-dimensions, decoding, scaling, statistical models, probabilities, no model of the world, hallucinate, Donald Mitchie, MENACE, Deep Mind, Demis Hassabis, Mustafa Suleyman, Shane Legg, Alpha Go, AlphaFold, Baysian statistics, SLAM, simultaneous localisation and mapping, self-driving car, Lidar sensor, risk assessment

Further Information

The source material for this article has all come from Toby Walsh The Shortest History of AI April 2025. If you think the material above is good, then I’d heartily recommend that you purchase and read Toby’s book. As the title suggests, it is short — 186 pages of text. It is vey readable and, apart from a few short sections, not particularly technical.

I have tried to summarise and paraphrase Toby’s book — any gaps, mistakes or errors are mine. I have also only used quotations, where they are personal comments or for other reasons.

I have made no attempt to rewrite Toby’s text and have shamelessly plagiarised it. It’s all his work not mine, and to have turned the bulk of the text into quotations would have just made it difficult to read.

I haven’t included Walsh’s short Part 3 on the future (for a later article perhaps).

Posted in Canberra.