An “Artificial Intelligence” Christmas Story
November 30, 2022 may well go down as a seminal moment in technology history. ChatGPT launched, an “AI” language model that can answer commands and questions with written responses. In basic terms, this has brought coding to the masses, because you can command the programme with written sentences or basic instructions, rather than complicated lines of code and the model will respond. How revolutionary is this? To find out, we put the machine to the test.
So what is ChatGPT? To answer this question, I asked the programme to explain to me how it worked. This was its response: “ChatGPT is a natural language processing (NLP) model that uses a combination of supervised and unsupervised learning to generate responses to user queries. It is a Transformer-based model that uses a deep learning architecture to process the input and generate the output. The model is trained on a large corpus of text and is capable of understanding context and generating relevant responses to user queries”.
Herein lies the curious thing about AI programmes like ChatGPT. The more information, or data, the programme is fed, the better it will become at identifying patterns, systems and what is a good or bad response to a command or question. One reason ChatGPT has been made available to the masses is because, by using it, we are feeding it data that will allow it to improve. How exponential these improvements will be is a matter for debate, but given over two million people have been playing with ChatGPT in its first couple of weeks, discovering that it can write scripts, essays (causing a problem for universities and schools), contracts, computer code, jokes, poems and appraisals (one of my colleagues found it gave him a very flattering review), it is being “fed” a large amount of varied input data. For context, it took Instagram over two months to get to a million users and Twitter almost two years.
The processing of inputs (“instructions”) and the generation of outputs (“the results”) is basically what AI is all about. This is something I learned when I completed an online course “AI For Everyone (deeplearning.ai)” in 2020, run by Professor Andre Ng of the Computer Science Department at Stanford University. As an example, large lines of code – inputs – are what makes a semi-autonomous car recognise people crossing a road and brake – the output.
What is so revolutionary about ChatGPT is that you can script it a bit like you would when coding, except that you can use natural language, everyday words and sentences, meaning you do not need to be fluent in coding languages like Python. As Phillip Isola, Associate Professor in EECS at MIT, noted “this means you can write out common tasks and attach them to command names. For example *simplify* X means rewrite X in simpler terms”:
All those English assignments trying to decipher Shakespeare just got a whole lot easier. Academics are already worrying. Dan Bebber, an Associate Professor of Ecology at Exeter University, tweeted: “Universities have a serious problem. ChatGPT can write perfectly good scientific essays, including reading and analysing scientific papers”.
Let’s take a step back before we get too worried about these immediate implications, because the answers to some of the problems AI might throw up appear quite medieval. If you are worried about students using AI tools to cheat on coursework, exam halls, pens and paper are probably a good solution.
If we can deal with the issues of AI models like ChatGPT, the opportunities are endless. If it can really analyse scientific papers, that is potentially revolutionary. Tom Hope, from the Allen Institute for AI in the US, pointed out that in biomedicine alone there are more than a million papers published each year: “humanity’s collective scientific knowledge is primarily documented in complex, unstructured texts in the scientific literature…think of an AI scientific assistant that could ‘read’ thousands of papers in your area of specialisation and help you get answers to questions and summarise multiple papers.”
Every single product or service could potentially be disrupted at some level by AI. Will we soon be able to shop, bank, buy services and find advice online just by typing in commands, without having to navigate through systems and websites? Or ask “please order some jeans that will suit me and go with my wardrobe” to a more sophisticated AI version of Siri or Alexa? Google Search appears vulnerable: the developer who invented Gmail predicted that the search giant may face “total disruption” within two years. And let’s not forget that we have also seen the first ever generative image programmes this year too (DallE, StableDiffusion and Midjourney). So really, 2022 will be seen as the year that generative AI emerged: the disruption and investment opportunities this will bring are only just getting started.
Of course, ChatGPT is still not perfect. It is a model very much in “training” mode, as Andrew Ng concluded: it is “sometimes amazing, and sometimes hilariously wrong”.
I put it to the test, asking it to write a Christmas story about the Montanaro Investment Team. My command is the first paragraph below and the text in green is what it came up with a few seconds later. A clever tool although I am not sure it got everything right, unless Yannis has something to share with us…
