What is generative AI and how does it work? – The Turing Lectures with Mirella Lapata

The speaker begins by introducing the audience to generative artificial intelligence (AI), explaining it as a combination of AI and generative capabilities. AI is described as the ability to make a computer program perform tasks typically done by humans, while the generative aspect involves creating new content based on patterns the computer has learned. Examples of generative AI include audio, computer code, images, text, and video. The speaker focuses on text due to their expertise in natural language processing. They aim to demystify the technology and present it as a tool. The talk is structured around the past, present, and future of AI, with a historical perspective provided by a quote from Alice Morse Earle, emphasizing the importance of the present.

The speaker discusses the evolution of generative AI, noting that it is not a new concept, with examples like Google Translate and Siri being early forms of the technology. Google Translate, launched in 2006, and Siri, launched in 2011, are highlighted as significant milestones. The speaker also mentions the auto-completion features in smartphones and search engines as everyday instances of generative AI. The lecture then shifts to the more sophisticated capabilities of AI, exemplified by OpenAI's GPT-4, which is claimed to perform exceptionally well on standardized tests like the SAT and in various professional exams. The speaker also presents a comparison of user adoption rates between Google Translate, TikTok, and ChatGPT, illustrating the rapid growth in the use of generative AI.

The speaker delves into the technology behind generative AI, focusing on language modelling. Language models are based on the principle of predicting the next word in a sequence given a context. The speaker explains that these models have evolved from simple counting mechanisms to sophisticated neural networks that can learn and predict word sequences. The process of training a language model involves using a large corpus of text, from which the model learns by predicting words that have been removed from sentences. The speaker also introduces the concept of a neural network, describing it as a series of interconnected nodes that process input data to produce output. The complexity of these networks is highlighted by the number of parameters they contain, with larger networks having more parameters and thus being more capable.

The speaker provides a deeper look into neural networks, particularly transformers, which are the foundation of models like GPT. Transformers are composed of blocks, each containing smaller neural networks, and are designed to process sequences of data, such as text. The speaker explains that transformers use a process called self-supervised learning, where they predict truncated parts of sentences from a large corpus of text. The lecture also touches on the importance of fine-tuning these pre-trained models for specific tasks, such as medical diagnosis or writing reports. The speaker emphasizes the significance of the number of parameters in determining the capabilities and complexity of a neural network model.

The speaker discusses the impact of scaling up language models, highlighting the increase in model sizes since 2018. They present a graph showing the exponential growth in the number of parameters, from GPT-1 to GPT-4, which boasts one trillion parameters. The speaker compares the size of these models to the human brain and discusses the amount of text that these models have been trained on. They also address the cost of training such large models, mentioning that GPT-4 cost $100 million to develop. The lecture concludes with a video illustrating the effect of scaling on the capabilities of language models, showing how increased parameters enable the model to perform more tasks.

The speaker explores the concept of fine-tuning in language models, explaining how models like GPT are initially trained on a vast amount of data and then fine-tuned for specific tasks. They discuss the importance of aligning AI behavior with human expectations, using the HHH framework—helpful, honest, and harmless—as a guideline. The speaker also addresses the challenges of fine-tuning, including the need for human involvement to provide preferences and the associated costs. They demonstrate the process with examples of how the model can be trained to provide helpful, accurate, and safe responses.

The speaker engages in a live demonstration with the audience, asking the AI questions and highlighting its capabilities and limitations. They showcase the AI's ability to provide detailed answers, generate creative content like poetry, and its struggles with brevity and relevance. The demonstration includes questions about historical figures, jokes, and even a request for a short song about relativity. The speaker emphasizes the AI's tendency to be verbose and the ongoing challenges in aligning its outputs with user expectations.

The speaker addresses the societal and ethical implications of generative AI, discussing the potential for job displacement and the risk of creating fake content. They provide examples of deep fakes and fake news generated by AI, highlighting the importance of regulation and societal awareness. The speaker also touches on the environmental impact of AI, noting the high energy consumption and carbon emissions associated with training and deploying large AI models. They conclude by emphasizing the need for responsible development and use of AI technologies.

In the final part of the lecture, the speaker reflects on the future of generative AI, acknowledging the uncertainty but expressing optimism. They quote Sir Tim Berners-Lee, the inventor of the internet, who suggests that while super intelligent AI is not yet a reality, the potential for both beneficial and harmful AI exists. The speaker argues that regulation and societal oversight are crucial in mitigating risks. They conclude by emphasizing the importance of considering the benefits and risks of AI in the context of other global challenges, such as climate change, and the need for responsible stewardship of AI technologies.

The speaker concludes the lecture by summarizing the key points and inviting questions from the audience. They reiterate the importance of understanding and regulating AI technologies and the need for ongoing dialogue about their impact on society. The speaker expresses gratitude for the audience's engagement and looks forward to a thoughtful Q&A session, highlighting the interactive nature of the lecture and the importance of community involvement in AI discourse.