“Humans in the loop” must detect the hardest-to-spot errors, at superhuman speed

🌈 Abstract

The article discusses the economic viability of the AI industry, the challenges it faces in finding high-value applications, and the limitations of the "human in the loop" approach to addressing AI errors.

🙋 Q&A

[01] The Economic Viability of the AI Industry

1. What are the key challenges the AI industry faces in terms of economic viability?

• The AI industry cannot indefinitely spend 1,700% more on Nvidia chips than it earns, nor can it give away millions of queries for free.
• The industry will need to find a mix of applications that can cover its operating costs, as investor disillusionment is inevitable.
• Low-value applications can soak up excess capacity but cannot sustain the industry on their own - the industry needs to find high-value applications as well.

2. What are the characteristics of the high-value applications the AI industry has identified?

• These high-value applications are high-stakes, meaning they are very sensitive to errors. Mistakes in applications that produce code, drive cars, or identify medical issues can have severe consequences.
• Some businesses may be insensitive to these consequences, as seen with Air Canada's chatbot fraud, but this is an unstable situation.

3. How does the "human in the loop" approach attempt to address the high-stakes nature of these applications?

• The "human in the loop" approach involves having fewer, cheaper workers supervise and monitor the AI system for errors.
• However, this approach has significant problems, as it essentially creates a "reverse centaur" where the human is used to augment the robot, leading to issues with vigilance and the subtle nature of AI errors.

[02] The Limitations of the "Human in the Loop" Approach

1. What are the key issues with the "human in the loop" approach?

• Humans are not good at maintaining eternal, perfect vigilance, especially when monitoring for rare and unpredictable errors.
• The types of errors AI systems make are often subtle and statistically indistinguishable from the truth, making them extremely difficult for humans to detect.
• The human in the loop is essentially being actively deceived by the AI system, which is constructing "what's wrong with this picture" puzzles that must be solved at high speed.

2. How does this approach impact workers?

• For workers, serving as the human in the loop in a scheme that cuts wage bills through AI is a nightmare - the worst possible kind of automation.
• It essentially turns the worker into a "reverse centaur", where they are used to augment the robot rather than being augmented by it.

3. What are the implications for AI companies and their high-value customers?

• The vigilance problem and the nature of AI errors make the "human in the loop" approach a significant challenge for AI companies trying to attract and retain high-value, high-stakes customers.
• This undermines the AI industry's claims that they can perform difficult scientific tasks at superhuman speed and produce billion-dollar insights, as these claims have often been found to be exaggerated or even hoaxes.