A Dilemma in AI Development

Recently, I attended an AI gathering with university students to discuss ideas and innovations. The event was enlightening until an incident sparked some deep thoughts.

An electrical engineering student proposed an AI secretary for CEOs and managers. The idea was that when you call a manager, this AI would answer, allowing you to interact, make appointments, inquire about free time, and handle other secretarial tasks.

The student’s explanation of the implementation involved buzzwords like deep learning and machine learning, clearly demonstrating a lack of understanding. While the proposal was flawed, it made me ponder two significant issues:

1. The Feasibility Paradox

You can always propose using various AI techniques - gradient boosting trees for audio processing, thousands of estimators, etc. But at what point can we definitively say an idea is unfeasible without actually implementing it?

What if implementation begins, and we enter an endless cycle of excuses?

• “We need more computational power.”
• “We require more data.”
• “We must add more layers and estimators.”

Each failure leads to another attempt, ad infinitum.

The philosophical question arises: How far back must we step to see that the puzzle pieces don’t fit? At what level does the violation occur?

• The atoms are good
• The molecules are good
• The material is good
• The structure is good
• …and so on

2. The Knowledge-Application Gap

When you’re familiar with a concept (like Gradient Boosted Decision Trees in machine learning) and suggest applying it to a problem, there’s often a fundamental issue in the deduction process.

The critical question is: When and how do you discover this flaw without actual implementation?

These reflections highlight the complexities in AI development and the importance of critical thinking in the face of technological hype.