A Plea for Sober AI

🌈 Abstract

The article discusses the hype surrounding the recent AI product announcements from OpenAI and Google, and argues for a more sober and practical approach to AI development and deployment.

🙋 Q&A

[01] The Hype Around AI Announcements

1. What are the key issues with the hype around AI announcements?

• The hype is so loud that it drowns out the true magic and practical capabilities of the AI products
• The demos and announcements from companies like OpenAI and Google set unrealistic expectations that are not met by the actual products
• There is a disconnect between the grand claims made about AI and the reality that these models still have limitations and make mistakes

2. How does the author contrast the hype with the actual capabilities of AI models?

• The author cites an example where a staff prompt engineer at Scale.ai quickly embarrassed GPT-4 once it was released, showing it did not live up to the expectations set by OpenAI's demo
• The author also points out that Google's AI tools came with small disclaimers acknowledging their limitations, which the author argues the companies do not think is a problem

3. What is the author's view on the current state of AI development?

• The author believes we are not close to achieving Artificial General Intelligence (AGI) anytime soon, and that the leap from GPT-3 to GPT-4 is not as significant as the leap from GPT-2 to GPT-3
• However, the author argues that a "quiet revolution" is happening with "Sober AI" - practical applications of large language models to solve specific problems, rather than aiming for all-singing, all-dancing models

[02] The Need for a More Sober Approach to AI

1. What is the author's main argument for a more sober approach to AI?

• The constant hype and overpromising of AI capabilities has led to unrealistic expectations and disappointment, even when the AI models can produce "decent first drafts" in seconds
• This hype trains users to either blindly trust whatever AI is presented to them, or to dismiss the entire field of AI

2. How does the author contrast the hype-driven approach with the "Sober AI" approach?

• The hype-driven approach focuses on flawless demos and general models that claim to handle anything, while hiding the difficulties and challenges of prompting and the reality of hallucinations
• In contrast, the "Sober AI" approach prioritizes consistency, efficiency, and practical use of large language models to solve specific problems, rather than aiming for AGI

3. What is the author's view on the future of AI development?

• The author believes that if we dial back the hype, the "Sober AI" approach will steadily and quietly remake many industries and applications, and that this more pragmatic path may actually get us to advanced AI capabilities faster than the hype-driven approach.