Today in AI...I'm using LLMs to provide personalized perfume recommendations
Or, guiding AI to create my unique 'scent profile'
I recently started getting back into perfumes (or parfums, as the French say). With a gentle nudge from my significant other, I ordered some Discovery Kits from parfumeries like Le Labo and Parfums d’Marly (I haven’t had the courage to order yet another one from Maison Francis Kurkdjian).
There’s something about looking good, feeling good, and smelling good to give you a boost of confidence throughout the day when you are crushing bugs creating amazing product visions. Yet parfums (even the Discovery Set) cost around $40-$50 each. And the bottles of eau de parfum cost around $200 (Layton costs $490 for 200ml!!! You can argue in the comments that this is robbery, but…i dunno, supply and demand I guess..).
Given that I want to keep costs down, I kept it to two sample sets. Here is what I got:
Le Labo
Parfums de Marly
Test parameters
Ok, so at first I figured I would just describe the one or two perfumes I really like, and then have the AI just find more like that. But that would be too easy. So I’m going to see if it can do a few more things:
I’m going to rate each perfume on a numerical scale from 1-10 (1 being ‘hate it’,10 being ‘love it and gotta have it’)
I want the AI to describe the head, heart, and base notes of each parfum
I want the AI to classify which season the parfum is best suited for (either fall, winter, spring, summer, all seasons)
Based on my preferences, I want the AI to create a ‘scent profile’ for me (I’m going to leave it up to the AI to figure this one out…)
Lastly, since the easiest thing in the world for me is to just buy the actual bottle of the parfum I like the best, I’m going to ask the AI to:
Recommend 2-3 scents (eau de parfums, not eau de toilettes) that will complement but not duplicate the parfum(s) I like the best, and
Diversify the seasonality from the ones I like the best (for instance, if my favorite ones are worn most during fall and spring, I want it to recommend parfums that are best worn during winter and summer)
Then I will either order either samples or, barring that, the actual bottles of the ones that the AI that I think best accomplishes the exercise. I’ll report back and see if it got it right!
I’m going to use Claude, Bard, and ChatGPT4 for this test. Let’s see how they do.
Parfum ratings
Given that I don’t have official (word here), I just tore up pieces of paper, sprayed the parfum, smelled the paper, and recorded the rating. Here’s what I came up with. Once I do this, I will have all the data I need to start the experiment .
Here are the ratings:
Le Labo
Gaiac 10 - 9/10
Cedrat 37 - 8/10
The Noir 29 - 4/10
Matcha 26 - 2/10
Parfums de Marly
Kalan - 3/10
Godolphin - 6.5/10
Percival - 6/10
Pegasus - 5.5/1
Layton - 8/10
Herod - 7/10
Greenley - 7/10
Carlisle - 7.5/10
Haltane - 6/10
ChatGPT4
Let’s get started with ChatGPT 4 (w/ Webpilot plugin) first
Claude.ai
Now onto Claude
Bard
Now onto Bard!
Scent Profile
I used ChatGPT4’s Advanced Data Analysis setting to come up with a visualization of the scent profile. I used a typical ‘perfume wheel’ as my visualization style. After a number of iterations I sort of gave up, since this was the best the AI could produce (I manually added a mask in post-processing):
I guess this might be helpful…somewhere? Sometime? I’m not even sure.
Next steps
I had to eliminate Claude’s responses since 2 of them were eau de cologne (I specified eau de parfum, which last longer). Therefore, it’s a face-off between ChatGPT and Bard. However, they both provided the same results! So, in the end, it’s pretty much a tie, with a slight edge to ChatGPT simply because it provided more information about when to wear the suggestions. I had to further clarify the recommendations to see which one came out ahead so I modified the request:
Needs to be for men (not unisex or female)
Bard
ChatGPT
They are both so similar in their explanation of why these were chosen. ChatGPT seems to just provide a little more information than Bard, but in reality it’s a tie. I then decided to test their recommendations out by getting samples of the perfums they recommended.
The order and the verdict
ChatGPT recommendations
Dunhill Icon - 5/10
Royal Oud - 5/10
Bard recommendations
Luna Rossa Black - 3/10
Explorer - 7/10
Hacivat - 7.5/10
Even though Bard gave me a dud, the other two perfumes it recommended were winners. (And in fact the dud wasn’t entirely a dud - it was just so faint smelling to be not worth even considering). Bard provided overall better perfumes that I enjoyed. Overall, I am going to use Bard to give me recommendations for perfumes going forward.
Reflections
Overall, I’m not sure how helpful this study was, especially given that I don’t have funds to just order endless samples of perfume. While I appreciate that Bard might produce better recommendations for a limited data set, maybe the biggest takeaway is that I tried to assign this much agency to an AI for something as personal as choice of personal scent. Some of this is simply a matter of taste, and can’t always be quantified by data. Or maybe it can, and I just need to ‘scale up’ the study? Either way, there is some science behind how the scents one choose extend beyond simple ‘masking effects’.