Unmaking Sense

Episode 8.56: Fine-tuning GPT-3 to suit your particular interests, training it with customised data.

This episode is in two parts: in the first we discuss the nature of fine-tuning, how it’s done and why it’s necessary; in the second part we discuss how at we as end-users can use our own data to customise the whole system to suit and serve better the particular interests that we have in our amateur or professional lives. We briefly discuss the costs of this process and how to do it.

Episode 8.55: The personalities of chatbots are embodiments of their training, just as are ours.

In episode 54, we talked about propensities in human beings and in chatbots, and I suggested that the chatbots will embody to some extent certain characteristics that are reflected in the way they’ve been trained, on the choices that have been made by those training them. Extend that a little, and you can start to see how different chat bots from different parts of the world trained by different groups of people are likely to have different characteristics and so different personalities. So human beings, when they come to decide which chatbots to use, will be affected by those factors in just the same way that they are when they choose their own human partners, friends and collaborators. But the fact that we can wrap these chatbots with fine-tuning means that we can also tailor them to our own personal or corporate or collective interests in order to try to persuade them to give better responses to the things that we are interested in. But where does that start and where does it finish, and how safe are we from the misappropriation of chatbot technology by those who would wrap them in skins and give them personalities of a kind that we would deplore and justifiably find quite alarming, dangerous and frightening? Fine tuning will be in 8.56.

Episode 8.54: Can a chatbot have a personality?

How the way a chatbot is trained imputes opinions and values that mean that it will take on the characteristics of its culture through its training.

Episode 8.53: What matters is whether our AI produces useful output, not whether it models reality.

Our artificial intelligence models should not be thought of as if they are trying to reproduce whatever it was that produced the phenomena that they are modelling. So if they produce something that looks vaguely like Shakespeare, that does not mean that they have recreated Shakespeare’s brain. If they produce reliable predictions of earthquakes, it doesn’t mean that they have reproduced a working theory that is adequate to the seismology of the Earth, all that matters is that their predictions are reliable or interesting or entertaining. My experience with Andrew Karpathy’s GPT model from YouTube, “Let’s build GPT: from scratch, in code, spelled out.”

Episode 8.52: Attention is all you need: how tokens find matches that will come next.

The “Attention is All You Need” paper lies behind Andrew Karpathy’s excellent YouTube video “Let’s build GPT: from scratch, in code, spelled out”. We discuss some implications.

Episode 8.51: Understanding tensors in multidimensional chatGPT architecture.

This episode might well be skipped by anyone who doesn’t like mathematical things, particularly the idea of thinking involved in three dimensions but if you’re remotely interested in what tensors have got to do with machine learning and therefore with chatGPT and large language models, it may be of interest. It’s very imperfect, and I am conscious that it may well be a bit of a peculiar episode and in the context of this series. I will come back to something much more down-to-earth episode 52.

Episode 8.50: Learning new things when there are no precedents, and especially new words.

We consider the relationship between words, meanings and understanding in the circumstances where a word is new, and not one that we have met or used before. Perhaps even a circumstance when nobody has used it before, as happens when the human race invents a new word to deal with a new circumstance, such as the rise of the Internet. We consider the way we learn by using words even when we don’t understand them, and ask whether we are right to be dismissive of the idea that a chatbot doesn’t understand the words either when we consider all the ramifications of this example.

Episode 8.49: Parallels between human and chatGPT generative decoding in speech, writing and life.

We look in more detail at the parallels between the way generative decoders produces text and the way human beings speak and write and generally live their lives by putting one foot in front of the other and deciding now what without necessarily having a definite idea of a final destination.

Episode 8.48: Now what? Using generative decoders like chatGPT to decide what comes next in life.

At the end of episode 47 of the series, I made the observation that if we could apply similar technologies to life itself, we might be able to find answers to questions like what next or now what? We explore this idea in a little bit more detail and consider its implications for the question how much of the past we should allow ourselves to take into consideration as we decide on the immediate and longer-term future.

Episode 8.47: How chatGPT works to generate human-like conversation.

We describe the process of tokenisation, encoding, and decoding in a way that explains the operation of software like GPT-3 and chatGPT from the beginning to the end of the text generation process. More details on the HuggingFace website, to which this episode is indebted.