Me On IT
Published on

Three Donations

Authors
  • Name
    Twitter

A German news magazine printed the following puzzle this week:

On the subway, people repeatedly get on and ask small donations. Steffi rarely gives anything, but today she is generous.

The first person who asks gets half of all the money she has on her, plus one euro.

Then, a second person asks for a donation. They get half of all the money Steffi still has, plus two euros.

Finally, a third person comes along. Steffi gives them half of all the money she still has, plus three euros.

When she leaves the subway, she has one euro left in her pocket.

How much money did Steffi have when she got on the subway?

The answer is 42, which can be easily derived with some basic school mathematics. The puzzle is not that exciting, nor is the solution particularly.

What is much more interesting, however, is what happens when you ask ChatGPT this question!

I tried exactly that. As expected, ChatGPT gives the wrong answer. One could now smugly proclaim: "Humans are still smarter than machines!" Ultimately, however, it's about finding out whether artificial intelligence might soon be able to solve such intelligence tasks despite its current weaknesses.

I tested this with the puzzle. The AI comes to a wrong conclusion and provides an incorrect solution. I then give the AI the right answer: 42. This leads to a new attempt at a solution, which also fails. Further probing results in the AI making a total of three new attempts, all of which also fail. However, the AI now independently checks the result and realizes that all the considerations it had previously made are wrong.

And then something amazing happens: The AI independently decides to try with a small Python program, and indeed -- the correct answer is calculated. This shows two things:

  1. The use of AI in software development will soon become a success factor. Even now, AI can write and deploy software. Exactly how and at what speed this ability will continue to develop is unknown.
  2. A LLM only realizes its potential when embedded into existing technical infrastructure. ChatGPT is an example of this. There, only the illusion is created that a conversation with a machine is taking place. In reality, an LLM is a simple I/O system. With each question, the AI receives the complete chat history, so it can deliver a new response that fits the context. The AI itself always generates an output in response to an input. That's all it can do!