The field of artificial intelligence has had many spurts of progress since it began in earnest in 1950, and also long periods of stagnation and pessimism. But in recent years we have seen the pace of research accelerate to surprising speed.
Large language models like GPT-4 have passed parts of the US Medical Licensing Examination, the Multistate Bar Examination and a test given to coding job candidates at Amazon. Image-generation models such as Midjourney are creating photorealistic images of famous figures that can fool thousands of people on social media. And highly specialised models like DeepMind’s AlphaFold predicted the structure of more than 200 million proteins, which researchers previously took years to solve one by one.
We have looked at this latest bump in AI ability in a special report. And here we have gathered five fundamental questions about artificial intelligence along with their answers.
Advertisement
What is artificial intelligence?
The term artificial intelligence was coined in 1956 by computer scientist John McCarthy. The context was a workshop at Dartmouth College in New Hampshire that attempted to “find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves”. The field has evolved since then, but AI is essentially still about creating machines that can do what we can, and more.
What is a neural network?
At the heart of current AI research are human brain-inspired computer programs called neural networks, each consisting of a vast number of connected digital neurons.
We receive input through our senses, but these AIs receive input as data, such as images, audio or text. In response to an input, our brains erupt with chaotic and complex sequences of firing synapses between neurons, whereas artificial neural networks process data more methodically.
A neural network can be trained to accomplish a task by passing huge amounts of data through it and then checking its output against a known expected result. For example, to create an AI that can transcribe handwriting, you could repeatedly show a neural network samples of writing and check its output against what the text is known to say. The values of the connections in the network would then be updated until the output edges nearer to the desired result. Do this for millions or billions of pieces of training data and you would find yourself with an AI that could transcribe handwriting it had never seen before.
The same approach powers the face-recognition feature on smartphones and the AI models that create photorealistic images from text prompts, as well as large language models (see main story). Almost all current AI research is based on neural networks, or specialised subsets of them that use different tricks and tactics to eke out better performance.
Why is ChatGPT so good?
OpenAI’s chatbot ChatGPT is powered by a type of neural network called a transformer, which was invented by Google in 2017. A transformer improves a standard neural network by attempting to pay attention to the most important parts of a data input while it is assessing the rest methodically. In a small way, it creates context.
AIs like ChatGPT are impressive, but their results are largely down to scale and vast sets of training data, says at New York University. “There’s no big, deep innovation, there’s not one big idea behind transformers. It’s just a particular way of fitting things together that happens to match really nicely with hardware and reliability in order to run really large-scale experiments.”
There are other approaches that could bring slight improvements, but there is currently no strong incentive for research teams to look elsewhere because transformers are working reliably and bringing results.
Are generative AIs conscious?
Some researchers, including those from big tech companies, have said that a few of the latest generative AIs show signs of consciousness. Others, however, strongly disagree.
Intelligence is a spectrum, rather than an end goal, says at the University of Birmingham, UK. Even simple objects like thermostats – which react in a predictable way to temperature – can be said to have rudimentary intelligence. GPT-4, the AI that powers ChatGPT, could therefore be thought to fall somewhere on that spectrum, but it certainly isn’t conscious, says Lee.
Will AI ever be truly intelligent?
There is “nothing magical” about how our brains operate that can’t be copied, says Lee. Yet he believes that human-like artificial general intelligence won’t arrive through the work of those hunting for it, but will emerge as a surprise from an unrelated project.
“We build the best weather simulation model in the world and it decides it doesn’t want to play with us. Or is lazy,” he says. “That’s when you have artificial general intelligence.”
Some researchers think it is just a matter of tweaks and time before current AI technology reaches this level. Others are exploring different avenues, such as altering neural networks so they more closely resemble the web of neurons in our brains and can better create contextual links.
Another approach involves studying animals such as fruit flies as a starting point and attempting to recreate their brains.
There could be many ways to create AI, but, for now, the human brain is the best model we have. The more that AIs reflect our brain, the better results we will get, says Lee.
