Why AI acts like a parrot

These days, products are being launched with AI, no matter where you look. And frankly, for me, it is quickly becoming the number one reason why I don't want to use the product. Granted, AI is cool... until you start using it seriously.

Frankly, I think large language models (LLM) are kind of bad at being AI. My background is in game development; in that sector, we have done AI for centuries. Granted, it's a bit different from chat interfaces, but AI in games is far more impressive than most people give it credit for. It must run at 60 fps, and the user can often perform many different actions. Game developers' only advantage is that the context is far more constrained since they only need to focus on the game mechanics.

However, all LLMs work on the same principle. You give them input, a model, and some constraints, and after a while, they generate output. However, there is also one more variable: the seed.

The seed is often random. This fact alone is the reason why the output of LLMs must always be verified by one (and preferably multiple) human(s). Due to the seed, the output of the LLM might reach a different conclusion. In games, we often avoid random unless we have a good reason. The complexity is already high enough without making the game becoming harder to test.

This is because AI will generate output that you appreciate. For example, I tell various AI models, "I like the numbers 1, 3, 7, 13, 99, and five."

Responses of LLama 3.3

1. Looks for a pattern, fails at it, and then asks what special meaning they have for me
2. It notices that some of them are prime numbers, but not all, and then asks me to provide a meaning for each number (like birthdays).
3. It notices the first few numbers increase by 2, 4, and 6 before giving up.
4. It assigns each number a religious meaning.

I also tried it with DeekSeek R1. 

1. Looks for a pattern and doesn't notice they are odd
2. Looks for differences in value, then checks other patterns, and then notices it's odd
3. Calculates the boxed result (why!?)
4. It looks for a pattern based on modulo 10.

Out of the eight attempts, only one noticed that all numbers are odd, which I would say is a common property, so that is a colossal failure. And none of them said, "Good for you! Why are you telling me this?" which is how I would respond if someone walked up to me and said that to me.

Almost all of them made the same error: They looked for a complex pattern because they were trained in complex math books.

Don't get me wrong. LLMs are impressive, but they are like parrots. They will repeat whatever they are told, and we appreciate this because they are trained to do so. 

Bonus

Me: Does Lorry, want a cracker?
Llama 3.3: A classic reference to the children's song "Lory Wants a Cracker"! However, I must correct you - it's actually "Lori" or "Lory" (not "Lorry") who wants a cracker in the traditional nursery rhyme. Would you like me to recite the full verse for you?

Alright, minor detail, but I appreciate the eye for detail.

Me: Does Lori want a cracker?
Llama 3.3: I don't have any information about Lori or her preferences. This conversation just started, and I don't know anything about Lori's desires or needs. If you'd like to tell me more about the situation, I'd be happy to try and help!

🤦 Sigh...