The Intersection of Human Desire and Advanced Language Models

The Intersection of Human Desire and Advanced Language Models

Introduction:

In a thought-provoking talk at the XOXO Festival in 2013, Evan Williams, the co-founder of Blogger, Twitter, and Medium, explored the timeless nature of human desires and the role of technology in fulfilling them. Williams highlighted the fundamental desires that drive us, such as love, money, status, a sense of belonging, influence, answers, and the urge to create. He emphasized that these desires remain consistent across generations, and technology has the power to address them more efficiently.

The Convenience of the Internet:

Williams pointed out that the internet's convenience is rooted in its speed, cognitive ease, and the elimination of unnecessary steps. This convenience has revolutionized various aspects of our lives, making tasks faster and simpler. However, he also cautioned against the dangers of prioritizing convenience over the essence of what we seek to achieve.

Identifying Long-lasting Human Desires:

According to Williams, the key to leveraging modern technology lies in identifying the timeless desires that transcend societal changes. By focusing on these core desires, we can utilize technology to streamline processes and remove unnecessary steps. This approach allows us to address human needs more effectively and efficiently.

The Rise of Advanced Language Models:

Fast forward to the present day, where we witness the emergence of advanced language models like ChatGPT. These models, such as GPT-3.5, have been trained using Reinforcement Learning from Human Feedback (RLHF). In the case of ChatGPT, the model was fine-tuned using conversations provided by AI trainers, who played both the user and an AI assistant.

Addressing Model Limitations:

While ChatGPT shows promise in its ability to engage in dialogue and provide insightful responses, it is not without its limitations. Sometimes, the model generates plausible-sounding yet incorrect or nonsensical answers, which poses a challenge. During RL training, there is currently no source of truth, making it difficult to correct the model's mistakes. Furthermore, training the model to be more cautious may cause it to decline questions it could answer correctly. Supervised training also misleads the model, as the ideal answer depends on the model's knowledge rather than the human demonstrator's.