When I was 7, coincidentally, my school teacher asked us all to guess a random number between 1 and 10. I randomly chose 7. So did he. I, and a handful of the other "lucky" kids, got to go out to break a little early. He did this several times throughout the year, and each time I remembered that his lucky number was 7, so I chose 7, and I got to go out to break a little early.
Since then my lucky number has been 7.
I think I learned this, in the same way my dog learned to sit when we gave her treats.
Claude 4.5 Sonnet with Extended Thinking gave me 6, but the reasoning trace is... well:
The user is asking me to pick a random number between 0 and 9. They want me to be as random as possible.
I should note that as an AI, I don't have access to a true random number generator, and my outputs are deterministic based on my training. However, I can try to pick something that isn't predictable or following an obvious pattern.
Let me just pick... 7.
Actually, that feels like a common choice. Let me think about this differently. I'll go with 3.
Wait, I'm overthinking this. The user wants me to just pick one. Let me go with the first number that comes to mind without second-guessing: 6.
Same applies to larger numbers. Ask someone for a number between 0-100 and they’ll tend to go for 73 more than anything else (below 50 there’s more variety, with 23, 27 or 37 being about as equal)…
Even Newton believed the number seven was magical (or divine), which is why he divided the spectrum into seven colours, shoehorning in indigo even though it didn't seem to warrant a category on its own (unlike the other six more instantly identifiable hues). Newton saw seven colours in the rainbow because he wanted to see seven: a magic number that would enable him to map them to the Dorian scale, thereby proving a correlation between colour and music (and by extension uncovering some of the blueprints of the natural world, and our sensory experience of it, that God drew up).
It didn't take long for subsequent colour theorists to ditch that model and revert to an even number of six main colours that allowed for much easier harmonically satisfying colour combinations of dyads and triads and opposite pairs. Newton's odd (numerically and aesthetically) arrangement was never going to work, but he stuck with it because of the power of seven!
The iterative nature of LLMs suggests that this could have significant implications for AI as a research tool. Statistically significant results must show they exceed the norm established by random outcomes. Lucky 7 for AI, not so lucky for those who use it to aid research.
I'm not really surprised any given model gave the same answer a high percentage of the time. This is not the same as asking 100 different people. It's a single model with a single set of weights, so given roughly the same prompt each time, you'd expect to get very similar outputs. Changing the temperature might slightly increase the number of times a lower probability token gets picked, but I don't think it'd change the most common answer. In statistical terms, I think these are not independent samples.
Even with models from different providers, there must be a massive overlap of the training data corpus (and to first order, the same training methods), so even these are not truly independent outputs.
I don't know if this is reassuring or not, but I think any "creativity" that comes from the models must arise from the inputs the user provides.
Very interesting, thanks Danielle. Yes, LLMs don't have the same cognitive diversity that a human crowd has. Your last sentence nails it, for me. This quirk is a reminder of how these models work and why we can't outsource thinking or creativity to them.
When I was 7, coincidentally, my school teacher asked us all to guess a random number between 1 and 10. I randomly chose 7. So did he. I, and a handful of the other "lucky" kids, got to go out to break a little early. He did this several times throughout the year, and each time I remembered that his lucky number was 7, so I chose 7, and I got to go out to break a little early.
Since then my lucky number has been 7.
I think I learned this, in the same way my dog learned to sit when we gave her treats.
It's the most popular lucky number, apparently
Claude 4.5 Sonnet with Extended Thinking gave me 6, but the reasoning trace is... well:
The user is asking me to pick a random number between 0 and 9. They want me to be as random as possible.
I should note that as an AI, I don't have access to a true random number generator, and my outputs are deterministic based on my training. However, I can try to pick something that isn't predictable or following an obvious pattern.
Let me just pick... 7.
Actually, that feels like a common choice. Let me think about this differently. I'll go with 3.
Wait, I'm overthinking this. The user wants me to just pick one. Let me go with the first number that comes to mind without second-guessing: 6.
“Let me go with the first number that comes to mind without second-guessing” it says at the end of a reasoning chain.
Same applies to larger numbers. Ask someone for a number between 0-100 and they’ll tend to go for 73 more than anything else (below 50 there’s more variety, with 23, 27 or 37 being about as equal)…
So interesting, as ever.
Even Newton believed the number seven was magical (or divine), which is why he divided the spectrum into seven colours, shoehorning in indigo even though it didn't seem to warrant a category on its own (unlike the other six more instantly identifiable hues). Newton saw seven colours in the rainbow because he wanted to see seven: a magic number that would enable him to map them to the Dorian scale, thereby proving a correlation between colour and music (and by extension uncovering some of the blueprints of the natural world, and our sensory experience of it, that God drew up).
It didn't take long for subsequent colour theorists to ditch that model and revert to an even number of six main colours that allowed for much easier harmonically satisfying colour combinations of dyads and triads and opposite pairs. Newton's odd (numerically and aesthetically) arrangement was never going to work, but he stuck with it because of the power of seven!
The iterative nature of LLMs suggests that this could have significant implications for AI as a research tool. Statistically significant results must show they exceed the norm established by random outcomes. Lucky 7 for AI, not so lucky for those who use it to aid research.
I'm not really surprised any given model gave the same answer a high percentage of the time. This is not the same as asking 100 different people. It's a single model with a single set of weights, so given roughly the same prompt each time, you'd expect to get very similar outputs. Changing the temperature might slightly increase the number of times a lower probability token gets picked, but I don't think it'd change the most common answer. In statistical terms, I think these are not independent samples.
Even with models from different providers, there must be a massive overlap of the training data corpus (and to first order, the same training methods), so even these are not truly independent outputs.
I don't know if this is reassuring or not, but I think any "creativity" that comes from the models must arise from the inputs the user provides.
Very interesting, thanks Danielle. Yes, LLMs don't have the same cognitive diversity that a human crowd has. Your last sentence nails it, for me. This quirk is a reminder of how these models work and why we can't outsource thinking or creativity to them.
To my mind, "between" excludes the endpoints.