Why “delve” is the most obvious sign of AI writing
🌈 Abstract
The article explores the reasons behind the prevalence of the word "delve" in AI-generated content, and how this is related to the training and fine-tuning of large language models (LLMs) using techniques like Reinforcement Learning from Human Feedback (RLHF).
🙋 Q&A
[01] How AI Language Models are Developed and Trained
1. What is the key difference between how humans and AI perceive and understand the world? Humans have a lived, experiential understanding of the world, while AI models are trained on vast amounts of data without true comprehension. AI uses a cluster of data points that are incomprehensible to humans, leading to fundamental differences in how they interpret information.
2. What is the significance of the "10:10" time display in AI-generated images? AI image generators tend to display the time as 10:10 because the training data it has ingested disproportionately represents clocks at that time, often due to an abundance of advertising images. AI cannot make the leap between cultural representations and the reality that clocks are not all set to 10:10.
3. How does the "chef in an alien kitchen" analogy explain AI's limitations in understanding the data it is trained on? The analogy suggests that AI is like a bewildered cook in an alien kitchen, using a vast cookbook it cannot comprehend. It can mix the data and ingredients, but has no real understanding of whether the resulting "meals" are edible or appropriate.
[02] Reinforcement Learning from Human Feedback (RLHF)
1. What is the purpose of using RLHF techniques in training AI language models? RLHF uses human preferences to guide the language model to refine its output, helping it align more to human tastes and less with its own algorithmic tendencies. Human reviewers rate the AI's outputs, and the model is rewarded accordingly.
2. How can RLHF lead to the overuse of certain words or phrases, like "delve", in AI-generated content? When the RLHF feedback loop is based on a limited sample of users who favor particular phrases or stylistic quirks, the AI learns to replicate these expressions, amplifying their usage in the generated text.
3. What is the potential impact of cultural and geographic biases in the RLHF process? The article suggests that outsourcing RLHF tasks to workers in the global South, who may use more formal or literary language due to their cultural and historical interactions with English, can lead to an overrepresentation of certain words like "delve" in the AI's output, which may not resonate universally.