Citation: Michael Webb, “2.3. A summary of key capabilities, limitations, and concerns around ChatGPT and other Large Language Models,” A Generative AI Primer, National Centre for AI, last modified May 22, 2023, https://nationalcentreforai.jiscinvolve.org/wp/2023/05/11/generative-ai-primer/. [Small adjustments were made to the text for clarity.]
Size of the training dataset.
The models underlying these tools generate their next-word probabilities based on patterns found in billions of words within a collection of preselected texts known as the training data.
Complexity of the next-word determination.
Rather than simply referencing a probability table, large language models like OpenAI’s GPT-3.5 and GPT-4 first perform billions of calculations using parameters that were determined from the training data in order to transform the initial prompt into a prediction about what word might come next. These are often referred to as “neural networks”. They also analyze the semantic structure of the sentences, which factors into their calculations.
Reliance on humans for training.
Without proper guardrails in place, large language models have a tendency to generate toxic content. In order for these tools to provide polite and safe responses, they go through a process called “reinforcement learning from human feedback.” This process requires human workers to manually review AI-generated content and provide feedback, which is then used to further refine the model. Notably, many of these multibillion dollar US-based tech companies use international contract workers, who are exposed to traumatizing content with low financial compensation and minimal mental health support. (See Time‘s article “OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic.”)
[From the AI Pedagogy Project @ metaLAB @ Harvard https://aipedagogy.org/guide/starter/ ]