Generative AI: time to learn a whole new vocabulary

I have no idea how to talk about sport. This was a disadvantage when growing up as a teenager at an all boy’s school. I felt as if I’d missed an important lesson, or failed to read the manual that the other boys had been issued at an early age. How else did everybody else have a vocabulary, a set of concepts, a whole language, that was opaque to me?

I initially felt the same way when attempting to learn in public about generative AI, the set of solutions such as ChatGPT and DALL-E which are receiving a lot of attention right now.

This was the week when I was supposed to read a few more detailed papers, to find a couple of books, and to go deep enough to get to grips with the main concepts. However, I found that, unlike my similar experiment with quantum computing, it was hard to find accessible entry points. Perhaps this is because, despite rapid developments in recent years, the ideas behind quantum computing have been around for a long while - long enough for experts to write introductions for curious laypeople like me. By contrast, most of the material describing generative AI technologies was quite new, and either so high level that it told me little I didn’t already know (and much that I had reason to be sceptical of), or dived so deep that I was as baffled as if listening to the dissection of a football match. No-one has had time to write the accessible introduction yet.

However, with some perseverance, and a few helpful links (shared at the end of this article), I was able to figure some things out. The main lessons I learnt were:

There are a few important terms . . .

I saw the same few terms over and over again, and derived the following meaning:

• Large Language Model (LLM): a model with a very high number of parameters, trained on huge quantities of data (i.e. the majority of text available on the Internet).
• Transformer: a multi-layer machine learning architecture in which an input (for example, a prompt) is encoded into a mathematical representation, then used to predict a valid output, then decoded back into a linguistic form.
• Attention: a way of processing the outputs from hidden steps in the architecture to figure out which parts of a piece of text are most relevant, and to overcome the problem that many language models forget what they are talking about partway through.

. . . but it’s all still machine learning doing prediction . . .

Despite these new terms, the technology is, at root, still machine learning using neural networks. Networks are trained using a combination of supervised and unsupervised learning. Then they are used to make a prediction: that this output is an acceptable response to this input.

. . . the difference is architecture and scale.

The difference between these new generative AI solutions and other solutions is therefore in architecture (different models and layers have been assembled to do different jobs in interesting ways) and scale (the models are huge and are trained on huge quantities of data).

We’ll come back to these points in a second: they have implications for the comprehensibility of generative AI solutions.

The models deliver plausibility, not accuracy or creativity . . .

Because, despite their architecture and scale, generative AI solutions are just machine learning models trained on large but limited data sets, they have the characteristics of all such models.

First, they are designed to suit a particular purpose, and that purpose is to provide an output which the recipient perceives as plausible: readable text in the case of ChatGPT; a recognisable image in the case of DALL-E. Plausibility is not the same as accuracy: ChatGPT and similar solutions are notoriously prone to producing text which sounds credible but which is fabricated.

Second, they are constrained by the training data. This means that these solutions are not creative: they may produce sentences which have never been written before, but they are limited to the total body of data which they have been fed. Furthermore, they are subject to all of the biases inherent in the data set.

. . . and it’s really hard to understand exactly what’s going on.

There are two layers of obscurity in these technologies.

First, they are complex specialisms which take effort to understand - a lot of effort. Unlike sport, technology is my field. And, as it’s a fast moving field, I often find myself running to catch up, and having to learn quickly. With this set of technologies, though, the field is developing so fast, and that the specialism is deepening so quickly, that I found machine learning experts struggling to explain the concepts to each other.

Second, machine learning models are notoriously hard to understand, as training results in a set of weights in a neural network which mean nothing to human beings. The size of LLMs put them far beyond the limits of human comprehension, resulting in us probing at the models through the interface (the prompt) to understand how they really work.

So, a few weeks of reading have led me to the conclusion that: generative AI is a type of machine learning built using large models, trained with huge datasets and a complex architecture, to predict plausible text or recognisable images based on a prompt. It is dependent on pre-existing data, and is neither creative nor reliably accurate. It is easy to imagine many practical applications, but it is hard to understand both the technology and the models.

I started this series by saying that I didn’t know how I felt about generative AI, but felt that I ought to feel something. I now know roughly how I feel: excited and impressed, but also concerned that we are creating a new set of practical technologies which are difficult to understand, which are often misrepresented, but which are so useful that they are already becoming part of our lives. I frequently argue that technologists have a duty to explain: in this case, I think that we also have an urgent duty to understand.

In my next article I will explore that duty.

In the meantime, here are a few links which helped my get to grips with this topic:

• Jay Alammar ’s great visual introduction to Transformers with Hugging Face
• Jay Alammar’s blog (worth reading many entries)
• A forthright and entertaining perspective on LLMs from Valliappa Lakshmanan
• The always wonderful AI Weirdness blog by Janelle Shane

(Views in this article are my own.)