BerserkerDay is filled with wonderful offers and exciting new wines. Would you like a sneak peak on those new wines to see if you might like them based on your past preferences?
We are working on a new feature for paid subscribers that uses your personal consumption and tasting data to make a prediction if you will like a wine. It is being slowly released to subscribers as we gather feedback and refine it. For BerserkerDay, we’d like to give you a chance to jump to the head of queue and check this out!
If you are a current paid CellarTracker subscriber and would like join the testing group (also featuring food pairing suggestions which are even cooler, tbh), please send a DM with your CellarTracker username and I will add you. Please note this is only available from the new app, so make sure to update to a recent version or download it now.
As somebody who works on recommendation systems and is passionate about wine, I’m genuinely curious about this feature and how it is implemented. I don’t think anybody has ever solved this problem very well. But if anybody can do it, CT certainly is sitting on the dataset for it. Looking forward to trying it out.
As a current tester, I love where this is going. Generative AI applied over the extensive dataset of CT reviews (both personal reviews and community) is very powerful. So much opportunity here and the first two use cases work quite well. Hoping to see some open ended prompting abilities too in the future
This feature is incredibly impressive. I feel like it’s easy to think my ratings are all over the place as I like all different styles of wine, but “Will I Like This” gets me! It even said I probably wouldn’t like Caymus or Bevan Ontogeny (I don’t care for Caymus and Bevan, while not like Caymus, has never really wowed me), but said I will probably like Myriad To Kalon, and while Myriad doesn’t excite me like it used to, I still enjoy them when I want a modern Napa Cab.
I guess maybe it depends on your TN practices. I rarely use the “do not like” button and the feature seems to think it’s “very likely” I’ll like just about any BDX because I’ve rated a few bottles highly. Yes, it gets the obvious things like I’m unlikely to enjoy Caymus, but I haven’t found that it seems to be able to differentiate between similar labels that are very different wines.
If it’s calibrated based on the like/do not like button, I wonder if it would be possible to indicate a points threshold for like/do not like. E.g., I usually use 88 as an average so 87 and below would be “do not like” and 89 and above would be “like” for me (others use the 100-pt scale differently so being able to customize your threshold would be great).
It most likely just takes your most recent reviews, or at least your scores, and just feeds it into ChatGPT, along with the target wine. While that can work in some cases, it obviously has some flaws compared to more traditional collaborative filtering and recommendation system approaches. For instance, it said that I would like a cognac because I like white wine (most likely it got confused by “Grande Champagne” which is a cru for cognac).
I definitely like the direction that this is going, but I don’t think that just bolting ChatGPT on to the app is necessarily the best way to do it.
Sort of but there is more under the hood that looks at consumption and other data versus just notes. Honestly, reviews are the hardest. Parsing out whether ‘oak’ was used as an epithet or praise, for example.
We don’t really support spirits. You will probably always be able to trick it with non-wine edge cases.
The next iteration will be more recommendation engine and using tools like collaborative filtering with some AI boost. 'WILTW" is a first step and test case for both interest in and useful feedback about.
I’ve searched several different wine with different styles and each time they’ve nailed the recommendation. I intentionally put in wines I have had that I dis not enjoy but never wrote notes on. Each time it told me I probably wouldn’t enjoy the wine and explains why.
It searches your cellar, consumption and notes and even notices palate shift. I like it. This will be useful when dealing with offers where I have no history and wonder if this white burg will be oaky or steely. Or will this champagne disagree with my palate.
One thing I did notice is that it doesn’t seem like it’s able to pull in a user’s complete review history. For instance, I drank a wine a few years ago and wrote a positive note. I then looked up a more recent vintage of the wine and asked CT if I would like it and it gave an inconclusive answer and said that I didn’t have any tasting notes that would indicate whether I would like the wine or not. So perhaps the LLM has a pretty short context window to include the user’s review history currently? (I submitted feedback through the app for this instance in case that helps with debugging)
We do use like/not and scores as well as consumption history. We are always working on ways to make these kind of features useful for people who are not interested in public notes.
Yes. I don’t remember what it is. It is a balancing act as too much information often gets contradictory or just stale. (And early iteration was putting too much weight on way past notes in my case.) These are dials we can tinker with.
Speculation here : I also think that among people who write a lot of tasting notes across a category, there is a tendency for vintage variation to get more prominence beyond stylistic matters. There is also a level where this is intended for cases of being utterly or mostly ignorant of the wine versus getting finer grained vintage information.
Let me just quickly say, this feature is really great. I continue to use it to see what the model thinks, and 9 times out of 10 I’m really impressed. Love where you’re taking this, and super excited to see even more smart ways of leveraging the personal + community data to do new things.
One little thing I found interesting it that it appears to be vintage-specific. I entered the 2023 vintage (most recent in the market) for a wine and the outcome was that I may “potentially” like it. It was a producer and varietal I had not consumed before but the CT reviews on the 2022 vintage were right in my wheelhouse. So when I entered the 2022 vintage, the app told me it would be “very likely” that I may like it. There have not yet been any reviews for the 2023 vintage so maybe that’s a limitation.
I’d love to give a try. My username is Zerklaires. I don’t really write reviews so it will be interesting to see how it works. To be honest, I’ve never seen a like/not like button on the new app. Will check for it!