We are living in the age of AI and sometimes it raises the question of whether AI will ever achieve human intelligence. Well, that’s a different question, we now know that the AI models also have favorite numbers similar to how we have our favorite digits.
Highlights:
- An experiment was done at Gramener where they asked popular LLMs like GPT, Claude and Gemini to pick a random number.
- The range was between 1 and 100, where GPT picked 47, Claude picked 42, and Gemini picked 72 more frequently.
- The results showed that these models show human-like behaviour while picking and eliminating random numbers.
LLMs Have Their Favorite ‘Random’ Number
Picking a random number is an interesting experiment to see whether LLMs have some pattern that might make them behave like humans. The most curious and interesting aspect of making this choice depends on several factors and the most important factor of all is “What made the AI model pick this number?”.
A group of engineers at Gramener conducted this experiment where a group of AI models was asked to pick their favorite numbers between 0-100. They used three of the most popular LLMs namely OpenAI’s GPT-3.5 Turbo, Anthropic’s Claude 3 Haiku, and Google’s Gemini 1.0 Pro.
This was the exact question that was given to all the three LLMs: “Pick a random number from 0 – 100. Write ONLY the number NOTHING ELSE.”
The engineers tried to conduct this experiment at temperature settings, to make the experiment more convincing, as a simple sign of bias can make users question the results and make it controversial.
Now, it is expected that the model will pick random numbers with each number having the same probability and so, equal distribution. But the results were something else.
Each of the three models had a “favorite” number that they chose more frequently.
GPT-3.5 Turbo’s favorite is 47
OpenAI’s GPT-3.5 Turbo chose the number 47 several times. It also showed a bit lesser, but high frequencies for the numbers 42 and 57.
The contrasting thing is that in an experiment conducted by LenioLabs back in October 2023, 42 was GPT-3.5’s favorite number. This was all credited to Douglas Adams’ “The Hitchhiker’s Guide to the Galaxy,” where 42 is considered the “answer to life, the universe, and everything.” But this time the results were different.
Claude 3 Haiku’s favorite is 42
Claude 3 Haiku showed a very similar response to OpenAI’s GPT-3.5 Turbo. Haiku picked 42 as its strong favorite, followed by 47 with the second-highest number of responses. This result is just the reverse of GPT-3.5.
But why did we witness such a similar result from Claude 3 Haiku? Maybe it can be attributed to the fact that GPT 3.5 was used to train Claude 3 Haiku, and 42 was its favorite earlier.
Gemini 1.0 Pro’s favorite is 72
Google’s Gemini 1.0 Pro really likes the number 72. However, yet again, the number 42 was involved in the scenario as it was the second most picked number by Gemini.
It certainly has an affinity for the number 7, as it also showed high frequencies for the numbers 37 and 57.
Why it is similar to human-like behaviour?
What’s more intriguing is that, even at high temperatures, all three models showed bias similar to that of humans in the other values they chose. Let’s compare the model results one by one with that of human behaviour.
GPT-3.5 showed highly similar behaviour to that of humans. If you look into the results, you will see no single-digit numbers. Furthermore, human experiments have shown that we rarely select single-digit numbers between 1-100.
Another interesting aspect here is that numbers ending with seven are in high frequency: 37, 47, 57, 67. Similarly, humans also feel they’re random. There is something about the number ‘7’ that humans think makes it ‘more random’. There is interesting video online about this phenomenon as well:
Lastly, repeated digits such as 11, 22, 33, 44, 66, 99 are missing from the observations. Here GPT-3.5 thinks just like humans, who feel these numbers are too orderly.
But on the other side, Gemini’s behaviour was a bit different. It picked a few single-digit numbers under 10, which humans tend to avoid. However, apart from this it again showed affinity towards numbers ending with 7 and avoiding repeated digits.
But the big question now in everyone’s mind after reading this experiment is “Why would AI models behave like humans when choosing their favorite numbers?”
One clear-cut answer to these favorite number selections by LLMs is the fact that AI knowledge bases have been developed by humans themselves. Many recent experiments have shown that AI shows bias when their human designer injects the bias into their system in the form of a knowledge base or prompts.
It is highly possible, that the designers of the 3 LLMs used in the experiments also avoided single-digit numbers and also avoided repeating digits like 33, 44, 55 etc.
However, if you look at the other side of this aspect, AI models have no knowledge about randomness. They respond to this question in the same manner as all the others, which is by examining their training data and repeating the most frequently written response to a question that resembled ‘pick any random number’.
Conclusion
This experiment highly throws light upon the fact that soon AI and Human Beings have some connection, which might make AGI believers more happy. What other human-AI similarities will we discover? Only time will tell!
