The Turing Lectures: The future of generative AI

The speaker, Hari Sood, welcomes the audience to a special Turing Lecture, the last in the series of 2023. He introduces himself as a research application manager at the Turing Institute, focusing on real-world applications of the institute's research. The event is a hybrid lecture and discourse, marking a new tradition. Sood engages the audience with a quick show of hands, recognizing both newcomers and returning attendees. He highlights the significance of the Turing Lectures, which have hosted world-leading experts in data science and AI since 2016. The institute is named after Alan Turing, a prominent British mathematician known for his role in cracking the Enigma code during World War II. The lecture series aims to advance data science and AI research for societal benefit. Sood also outlines the format of the evening, including a Q&A session and the use of social media for audience engagement, and sets the stage for a discussion on generative AI.

The lecture delves into the history of AI, noting its slow progress until the 21st century. It emphasizes the pivotal role of machine learning in advancing AI, particularly after 2005. Machine learning, despite its misleading name, does not involve self-study but rather the use of training data to teach AI systems. The talk uses the example of facial recognition to illustrate supervised learning, where a system is shown input-output pairs to learn to recognize faces. The importance of training data is underscored, with social media contributing significantly to this data pool. The lecture also touches on the risks and potential future of generative AI, which is the focus of the lecture series.

The discussion shifts to the biological inspiration behind neural networks, which mimic the human brain's vast network of neurons. Each neuron performs a simple pattern recognition task, and when a pattern is detected, it signals other connected neurons. This mechanism is harnessed in AI to recognize complex patterns, such as images. The process of training a neural network involves adjusting the network's connections using training data until it can reproduce the desired output. The lecture also highlights the importance of data in training these networks and the significant advancements made possible by the availability of big data and increased computing power.

The lecture explores the concept of big data and its impact on AI, particularly in the training of neural networks. The rise of deep learning, the abundance of data, and affordable computing power are credited with the significant progress in AI. Training a neural network involves adjusting its structure to produce the desired output when presented with training data. The lecture also discusses the role of GPUs in training large neural networks and the commercial interest from Silicon Valley, leading to substantial investments in AI technology.

The lecture describes the process of training AI systems using vast amounts of data from the internet. The training data for a system like GPT3 consists of 500 billion words from various sources, including web pages, PDF documents, and more. This data is used to train the AI in language tasks, such as predicting the next word in a sequence. The lecture also touches on the legal and ethical considerations surrounding the use of copyrighted material in AI training data.

The speaker addresses the limitations of AI, including its tendency to produce incorrect information that is presented in a plausible and fluent manner. The lecture discusses the need for fact-checking and the ethical implications of AI's use, particularly concerning bias, toxicity, and copyright issues. The speaker also highlights the challenges AI faces in understanding and reasoning about complex or abstract concepts, emphasizing that AI lacks the mental processes and consciousness of humans.

The lecture contemplates the concept of general AI, which would possess a broad range of cognitive abilities similar to humans. It outlines various levels of general AI, from a system capable of any intellectual task to one that can perform any language-based task. The speaker expresses skepticism about the imminence of full general AI, citing the limitations of current technology and the complexities of tasks like loading a dishwasher. The lecture also considers the potential for AI to become 'superhuman' and the ethical considerations this raises.

The lecture concludes with a discussion on the future of AI, suggesting that while large language models are powerful tools, they do not provide deep insights into human nature or mental processes. The speaker also addresses the potential for AI to augment human capabilities, the importance of understanding the brain's functional organization, and the ongoing research into integrating neural and symbolic AI approaches. The lecture ends with a reminder of the Turing Lecture series' upcoming Christmas Lecture and an invitation to follow the Turing Institute for future events.