Four key ingredients of modern AI | Deeper dive

Each of these four key ingredients for modern, post 2017 AI, from the previous video on large language models (LLMs) will be addressed individually below. 

1) Large amounts of data: Illustratively, we saw in the video that it would take well over 2000 years for a human to read the amount of text required to train ChatGPT3. That is the equivalent of over 1000 original CD ROMs (as seen in the image of Bill Gates below). If Bill Gates were to sit on a stack of paper high enough to encompass this information, he would be at the cruising altitude of a jetliner. Note too, this is now the training of an AI system that took place several years ago, and ChatGPT has progressed significantly since then.

Image by Louie Psihoyos from National Geographic

2) Suitable algorithm/mathematical function: These data need to be processed in some way; in the case of a LLM, this processing occurs through a Transformer (a sequence of mathematical functions and the ‘T’ in ChatGPT). We will avoid the details of this here, however it is actually this key ingredient that was really the last to the show, and is the reason that we are seeing the massive expansion of AI today. It is largely the advent of the Transformer architecture that has allowed for the seemingly disparate applications of AI to make advancements in parallel. If you have ever heard the term ‘black box’ when people speak about AI systems, it is this algorithm (or equivalent) that is being referenced.

TASK: Prompt an AI system to, “Explain simply in fairy-tale language the ‘black box’ people reference when talking about AI”.

What kind of response do you get? Is this satisfactory to you in helping you understand the term algorithm?

3) Extensive computational resources: To train ChatGPT3, you would need to have hundreds of the best computer chips running for one year without interruption. This manifests in costs well into the millions of pounds just to facilitate computation, not even considering research and development. This places the development of AI systems exclusively in the hands of those with large amounts of capital, creating an inherent power imbalance. We will be discussing these aspects in greater detail in the AI and Environment module of this course (Week 3).

4) Reinforcement Learning with Human Feedback (RLHF): This is a subtle concept with many faces. To familiarise yourself with the idea, try playing around with an LLM. 

TASK: Prompt your chosen AI system to “Simulate a short dialogue between an AI and a human discussing why each reinforcement learning with human feedback is used in the development of LLMs. Provide a concise example of RLHF, where I can see how an AI would give a response and the human would interact with the AI.”

Reflect on the overall title of this course, “AI Ethics, Inclusion & Society” as you read the AI response.

An excellent, in depth and accessible compliment to our above learning is the podcast “Preparing for the Intelligence Explosion” by MacAskill and Moorhouse, published online by Forethought in 2025.