This is a great article, and I had a lot of thoughts throughout, but, for now, as far as AI goes, I have the feeling that we can't compete with AI trying to do AI-kinds of things. My mindset about this is not that much different than, say, a calculator. Why try to out-calculator a calculator?
Instead, we can stick to doing what we do best: be as human as humanly possible. I'm sure that we'll eventually navigate this change.
Thanks Catania! And totally agree, when it comes to storing, quickly processing and analyzing huge amounts data there's nothing we can do to best these models. But the human element when it comes to social interaction and the kinship of shared sensory experience, not sure that can ever be replaced!
If a FOH manager couldn’t find a Somm, had a reasonably curated wine selection in the cellar, and had some knowledge of AI…Could the FOH and AI pick the pre-selected wine pairing for a multi-course tasting menu? AI could do that now, if queried properly and given access to a wine list with reasonable depth, the AI might come up with a perfectly good wine paired menu. Input the menu with ingredients, the wine list, pricing constraints, a reasonable query. Yep, I think that could happen, may actually be happening today.
In fact, I did it in the background with Notebook AI, using the publicly available wine list from a Michelin Two Star restaurant in San Francisco, along with their public seasonal tasting menu. It took three sets of prompts for me to narrow it down but it actually isn’t a bad pairing menu. I even had it show me costs, using only current BTG wine and I kept the bottle prices to under $125, guessing a 4x markup, so as to have options. There were 2,018 total by the bottle selections available on the current wine list. I can’t keep that many in my head or remember it…but it was easy for the AI to summarize for me. This all took about 5 minutes to get done, btw.
Getting back to the premise, Can AI replace Somms? Insofar as guiding someone’s choices, on the spot, in a restaurant with a tasting menu where the guests don’t wanna go with the pre-selected wine pairing? That’s an interesting dialogue between the guests at the table and the Somm. No, this dialogue, this interaction should and will continue, same as iPads with wine lists didn’t replace Somms either. Are guests going to start doing prompt engineering at the table? Perish the thought.
Would/Should a Somm at a Michelin restaurant with over 2000 wine choices on their menu use AI to help develop pairing suggestions? I would think they would be today. The Somm could then use their own tasting knowledge to refine, curate the list to the specific tasting menu at hand or the substitutions that come in last minute, sea bass didn’t show, we’re going with halibut kind of thing.
Scary how quickly it's gotten here, but it's definitely an incredible tool we have rapidly growing under our fingertips! And I think you're right, the key distinction currently is and will continue to be that one AI take over on matters of analyzing huge amounts of data effectively and quickly, while the human side will be much more difficult to derail. I don't really see any way that AI can replace human social interaction, and for the foreseeable future there will inveitably be a need for first hand human sensory experience somewhere in the process (be that in providing the initial data to feed the models, or curating the list late in the process). But if these blind tasting robots win enough competitions, who knows, maybe they'll be the palates people trust the most. Maybe I'm being blindly optimistic, but I have a feeling that even if they get that advanced there will remain a natural human instinct to retain more trust in actual people when seeking insights on matters of sensory experience.
My dudes, did you just use an AI-generated image as the marquee for an anti-AI post?!?? I gotta give you some sh*t for that. And push you use those a little less in the future, too. AI art is one of *the* most ethically iffy offerings of current AI, so at least minimize the use of it!
Now, great post to get us started here, though I'm somewhat shocked to realize...I think I might disagree with literally every single point you argued! And I'm not remotely pro-AI!
I have to save most of it for my own upcoming post, but I'll comment on one statement here:
"LLMs can’t tell you what a particular wine actually tastes like. They can often tell you what specific critics who have tasted the wine have said about it, or what the producers have said about it, or they can give an averaged out generalization of notes that have been publicly released."
My response to that: *how are most human somms any different than this?*
Humans can tell you what they think they taste, which may or may not be affected by what they recently ate, drank, breathed, how well they slept, etc., and even putting that aside thet can't say what any other person is going to taste.
And even beyond this, human group think is real: most somms are going to describe a wine based on a pre-taught set of flavor notes they're programmeed to think of and use, especially in regards to any specific type of wine. Whether it's an AI cobbling together a description based on past writings or a human essentially doing the same, neither can be trusted to have determined a wine's taste based on anything less than programming given to the by other humans. Other humans have a similar success ratio when trying to describe a wine to me as any official write up of a wine does in matching my own take on the wine. And even that "official write-up", at the end of the day, was written by a human describing how the wine tasted to them. I'm not sure I'm seeing where AI and human somms differ enough on this front.
A lot to unpack here! Okay, first of all, this isn't an "anti-AI" post, it's part of an examination of AI and its role in the world of wine, and respectfully disagree that ironically using an AI image as the cover is problematic. As far as your perspective on the quoted part of the post I think this is a bad faith interpretation of what we wrote. The idea that humans don't have honest personal opinions and subjective sensory experiences is insane. I won't argue that there isn't bias -- I think the best or most original tasting notes often come from people who know little to nothing about wine or the classic tasting notes given to them. But that doesn't change the fact that these are people who can actually experience the wines themselves, and sure there are probably people who just regurgitate other people's notes, but you've got to seek out the voices that you trust. You can find people who you agree with or disagree with the tastes of, and they can be a reliable resource, like finding movie or book critics you respect or understand. They'll give you consistent perspectives and whether you disagree or agree with them, you get a directional sense of where you'll land because you trust the perspective to be honest. You can't say the same of an AI unless it's quoting specific people.
And as for this "Whether it's an AI cobbling together a description based on past writings or a human essentially doing the same" like I am genuinely baffled that you think humans just amass averaged opinions from other sources and don't have their own. Just using our blog for example, we are only writing tasting notes on bottles we've tasted, we aren't scraping the internet for other people's opinions who, by your logic, did the same. Where's the source material in that loop? At any rate, appreciate you taking the time to comment, looking forward to seeing your deeper analysis in the days to come!
AI art will always be problematic, whether ironic in its use or not. You may be able to argue that it isn't ironic (though I still don't buy that, either - this is a lot of pushback on two line items for a "not an anti-AI post") but that still doesn't argue against the issue of AI art on the whole.
"The idea that humans don't have honest personal opinions and subjective sensory experiences is insane."
That's not what I wrote, speaking of bad faith interpretations - or perhaps simply faulty interpretations, let's not conflate those two thing, I think we are both doing this in good faith. I literally wrote "Humans can tell you what they taste..." Humans of course have a subjective experience, but transforming that into language is rarely subjective or particularly personal. If it was, wine tasting notes wouldn't be as homogenous as they are. Have you ever tasted "garrigue"? How about a raw gooseberry? Or fresh churned earth? Certainly some human writers have written more personal tasting notes than others, but not the majority. And then consider even if they did: how would this be helpful? Sometimes I see tasting notes about very specific species of cherry or orange or mushroom that I've never even heard of. Is that helpful? And would I peronally taste the same thing even if I was familiar?
We're not talking about if a person can slowly cobble together their own palate translations to another's - we're talking about whether an AI in a somm position is more or less useful than a human in that same role. When it comes to giving tasting notes in a moment of service, I have to say I'm not seeing the difference between the two, in this specific context.
"I am genuinely baffled that you think humans just amass averaged opinions from other sources and don't have their own."
Humans are social creatures and very, *very* vulnerable to group think, mob mentality, and social modes of thinking at any given time in history. Sociology is pretty clear on this topic. There will always be outliers, unlike with AI, but outliers alone don't create a major enough distinction when considering somms-in-service as a whole. And when you couple this with the fact that the majority of humanity lacks any significant understanding of wine to begin with (something that is key), thn whether a somm states something like the a god-given truth or an AI does, the guest isn't likely to distinguish the difference between the two. And, again, we're talking about whether AI can replace somms. And I think somms as they currently exist - there's a LOT that AI can replace, given where we currently are.
Anywho, I was trying to keep it simple for the comment section, but there is much to unpack here, as you said.
Maybe it's not what you meant by it, but saying that somms are cobbling together a description based on past writings implies that they aren't using subjective experience to provide honest personal opinions on a case by case basis. There's undoubtedly descriptors that "trained palates" are prone to using, but as you noted, there's a clear justification for the use of familiar descriptors so that people do understand them, and I strongly disagree that translating that subjective experience into language isn't personal.
As you sort of touch on here, homogeneity is a byproduct of familiarity, but it's true that getting more specific has its own pitfalls (though, I'm not sure what kinds of oranges are throwing you off 😂). Still, I think the differences are often nuanced with things like that – I don't think you lose someone when you get more specific within a familiar category for them; even if they can't tell you from memory the difference between a satsuma and a blood orange, it doesn't render the distinction meaningless and someone who has tasted a "standard issue orange" will still get a sense of what that means. Personally I think if you are being more specific in a way that makes sense to you and is honest, then as long as it isn't that rare flower you and 2 other people in existence smelled that one time you were on the top of Mt Fuji, specificity is usually a net positive.
As for the descriptors you called out, I totally get the point you're making - garrigue and gooseberry (which I had to order a can of online to familiarize myself with) are much more obscure to people in the US (though I think "fresh churned earth" is a pretty familiar scent to most people who have used a shovel). I will say that I refuse to use sensory descriptors that I am not familiar with and I go out of my way to familiarize myself with the ones that are commonly used in wine notes. It doesn't make those descriptors more useful though, and I tend to shy away from using them because I am aware of that level of unfamiliarity. This is something that Zach and I disagree on sometimes, but one thing we are aligned on is never providing notes that don't reflect what we are actually getting from a bottle. Something that is only possible if you've actually tasted the bottle in question.
Another key aspect I think you're overlooking here is that descriptors come up repeatedly for certain wines because there are patterns that exist. It's the reason people are able to guess wines blind in the first place. Undoubtedly that contributes to similar notes coming up again and again for different bottles from the same varietal or region, and it doesn't suggest some conspiratorial movement in the wine community to pretend those aromatics or flavors are there. It comes full circle back to using notes that people understand when you notice that the variety is lacking, but again, you have to balance familiarity with specificity as artfully as you can.
Ultimately what we were trying to hammer home in our post is that sensory experience is still uniquely human. The very important distinction between a human and an AI when describing a specific bottle is that the AI has never tasted it. It disregards the importance of individual specificity and trust of an individuals honest conveyance of their personal tastes. And I think saying that we're not talking about cobbling together our "own palate translations to another's - we're talking about whether an AI in a somm position is more or less useful than a human in that same role" misses the core point we were making, which is that an AI simply cannot describe a specific bottle of wine if humans haven't already done so. We aren't saying that getting a collated group of tasting notes from human sources provided by an AI won't be useful, we're saying that without people doing it first, it's literally impossible.
I'm *not* saying - nor have I said - that humans are cobbling together a description based on previous writings - I'm saying/asking: "what's the difference?" in the context of a somm service position. Questioning the difference of the result, especially as all AI writing is the result of something another human wrote. Whether it's that hiuman or another human standing in front of you - what's the difference in getting tasting notes culled from a stranger? And especially especially as our tasting notes, when shaped into words, are funneled through a cultural group think that is rarely personal or unique. People not into wine writing/reading or education will always surprise with notes that veer far away from the norm, and yet giving them these norms are somehow more uniquely human than an AI compliation of other humans? How? In what way?
I also never used a word or idea like "conspiratorial", and quite consciously so - if you need to reframe an argument as an illogical extreme in order to counter it, you're not actually countering it. Human being are gonna human, and sociological effects of group think are well documented, and "industries" are littered with them moreso than other groups. I'm equally as baffled as to how you think this is not a basic reality of human thinking and interactions in social situations, of which somms and the guests they cater to very much are.
Will we still need humans tasting and writing up wines in some way, somewhere, to some extenet? At least insofar as AI currently exists, that would be necessary, but that in no way = AI cannot replace somms, at least to a significant degree. And if the AI tech that allows them to breakdown aroma and taste like a human, that could very easily replace humans needing to do that in the first place.
Which, again, is not the same consideration as how much we do or shoiuld value the difference of a human vs. an AI. It's just looking very soberly at whether it CAN be done, and how much generations that AI will be native to and/or people who know so little about wine in the first place that they themsleves could never appreciate the difference, would ever care. And I'm thinking they won't, unless we give them a hell of a lot more reason to. If we don't, the replaceement isn't going to be hard, at least in many contexts.
This is a great article, and I had a lot of thoughts throughout, but, for now, as far as AI goes, I have the feeling that we can't compete with AI trying to do AI-kinds of things. My mindset about this is not that much different than, say, a calculator. Why try to out-calculator a calculator?
Instead, we can stick to doing what we do best: be as human as humanly possible. I'm sure that we'll eventually navigate this change.
Thanks Catania! And totally agree, when it comes to storing, quickly processing and analyzing huge amounts data there's nothing we can do to best these models. But the human element when it comes to social interaction and the kinship of shared sensory experience, not sure that can ever be replaced!
No.
😂 Well put! Cheers!🥂
Can AI replace Somms?
If a FOH manager couldn’t find a Somm, had a reasonably curated wine selection in the cellar, and had some knowledge of AI…Could the FOH and AI pick the pre-selected wine pairing for a multi-course tasting menu? AI could do that now, if queried properly and given access to a wine list with reasonable depth, the AI might come up with a perfectly good wine paired menu. Input the menu with ingredients, the wine list, pricing constraints, a reasonable query. Yep, I think that could happen, may actually be happening today.
In fact, I did it in the background with Notebook AI, using the publicly available wine list from a Michelin Two Star restaurant in San Francisco, along with their public seasonal tasting menu. It took three sets of prompts for me to narrow it down but it actually isn’t a bad pairing menu. I even had it show me costs, using only current BTG wine and I kept the bottle prices to under $125, guessing a 4x markup, so as to have options. There were 2,018 total by the bottle selections available on the current wine list. I can’t keep that many in my head or remember it…but it was easy for the AI to summarize for me. This all took about 5 minutes to get done, btw.
Getting back to the premise, Can AI replace Somms? Insofar as guiding someone’s choices, on the spot, in a restaurant with a tasting menu where the guests don’t wanna go with the pre-selected wine pairing? That’s an interesting dialogue between the guests at the table and the Somm. No, this dialogue, this interaction should and will continue, same as iPads with wine lists didn’t replace Somms either. Are guests going to start doing prompt engineering at the table? Perish the thought.
Would/Should a Somm at a Michelin restaurant with over 2000 wine choices on their menu use AI to help develop pairing suggestions? I would think they would be today. The Somm could then use their own tasting knowledge to refine, curate the list to the specific tasting menu at hand or the substitutions that come in last minute, sea bass didn’t show, we’re going with halibut kind of thing.
Scary how quickly it's gotten here, but it's definitely an incredible tool we have rapidly growing under our fingertips! And I think you're right, the key distinction currently is and will continue to be that one AI take over on matters of analyzing huge amounts of data effectively and quickly, while the human side will be much more difficult to derail. I don't really see any way that AI can replace human social interaction, and for the foreseeable future there will inveitably be a need for first hand human sensory experience somewhere in the process (be that in providing the initial data to feed the models, or curating the list late in the process). But if these blind tasting robots win enough competitions, who knows, maybe they'll be the palates people trust the most. Maybe I'm being blindly optimistic, but I have a feeling that even if they get that advanced there will remain a natural human instinct to retain more trust in actual people when seeking insights on matters of sensory experience.
My dudes, did you just use an AI-generated image as the marquee for an anti-AI post?!?? I gotta give you some sh*t for that. And push you use those a little less in the future, too. AI art is one of *the* most ethically iffy offerings of current AI, so at least minimize the use of it!
Now, great post to get us started here, though I'm somewhat shocked to realize...I think I might disagree with literally every single point you argued! And I'm not remotely pro-AI!
I have to save most of it for my own upcoming post, but I'll comment on one statement here:
"LLMs can’t tell you what a particular wine actually tastes like. They can often tell you what specific critics who have tasted the wine have said about it, or what the producers have said about it, or they can give an averaged out generalization of notes that have been publicly released."
My response to that: *how are most human somms any different than this?*
Humans can tell you what they think they taste, which may or may not be affected by what they recently ate, drank, breathed, how well they slept, etc., and even putting that aside thet can't say what any other person is going to taste.
And even beyond this, human group think is real: most somms are going to describe a wine based on a pre-taught set of flavor notes they're programmeed to think of and use, especially in regards to any specific type of wine. Whether it's an AI cobbling together a description based on past writings or a human essentially doing the same, neither can be trusted to have determined a wine's taste based on anything less than programming given to the by other humans. Other humans have a similar success ratio when trying to describe a wine to me as any official write up of a wine does in matching my own take on the wine. And even that "official write-up", at the end of the day, was written by a human describing how the wine tasted to them. I'm not sure I'm seeing where AI and human somms differ enough on this front.
A lot to unpack here! Okay, first of all, this isn't an "anti-AI" post, it's part of an examination of AI and its role in the world of wine, and respectfully disagree that ironically using an AI image as the cover is problematic. As far as your perspective on the quoted part of the post I think this is a bad faith interpretation of what we wrote. The idea that humans don't have honest personal opinions and subjective sensory experiences is insane. I won't argue that there isn't bias -- I think the best or most original tasting notes often come from people who know little to nothing about wine or the classic tasting notes given to them. But that doesn't change the fact that these are people who can actually experience the wines themselves, and sure there are probably people who just regurgitate other people's notes, but you've got to seek out the voices that you trust. You can find people who you agree with or disagree with the tastes of, and they can be a reliable resource, like finding movie or book critics you respect or understand. They'll give you consistent perspectives and whether you disagree or agree with them, you get a directional sense of where you'll land because you trust the perspective to be honest. You can't say the same of an AI unless it's quoting specific people.
And as for this "Whether it's an AI cobbling together a description based on past writings or a human essentially doing the same" like I am genuinely baffled that you think humans just amass averaged opinions from other sources and don't have their own. Just using our blog for example, we are only writing tasting notes on bottles we've tasted, we aren't scraping the internet for other people's opinions who, by your logic, did the same. Where's the source material in that loop? At any rate, appreciate you taking the time to comment, looking forward to seeing your deeper analysis in the days to come!
AI art will always be problematic, whether ironic in its use or not. You may be able to argue that it isn't ironic (though I still don't buy that, either - this is a lot of pushback on two line items for a "not an anti-AI post") but that still doesn't argue against the issue of AI art on the whole.
"The idea that humans don't have honest personal opinions and subjective sensory experiences is insane."
That's not what I wrote, speaking of bad faith interpretations - or perhaps simply faulty interpretations, let's not conflate those two thing, I think we are both doing this in good faith. I literally wrote "Humans can tell you what they taste..." Humans of course have a subjective experience, but transforming that into language is rarely subjective or particularly personal. If it was, wine tasting notes wouldn't be as homogenous as they are. Have you ever tasted "garrigue"? How about a raw gooseberry? Or fresh churned earth? Certainly some human writers have written more personal tasting notes than others, but not the majority. And then consider even if they did: how would this be helpful? Sometimes I see tasting notes about very specific species of cherry or orange or mushroom that I've never even heard of. Is that helpful? And would I peronally taste the same thing even if I was familiar?
We're not talking about if a person can slowly cobble together their own palate translations to another's - we're talking about whether an AI in a somm position is more or less useful than a human in that same role. When it comes to giving tasting notes in a moment of service, I have to say I'm not seeing the difference between the two, in this specific context.
"I am genuinely baffled that you think humans just amass averaged opinions from other sources and don't have their own."
Humans are social creatures and very, *very* vulnerable to group think, mob mentality, and social modes of thinking at any given time in history. Sociology is pretty clear on this topic. There will always be outliers, unlike with AI, but outliers alone don't create a major enough distinction when considering somms-in-service as a whole. And when you couple this with the fact that the majority of humanity lacks any significant understanding of wine to begin with (something that is key), thn whether a somm states something like the a god-given truth or an AI does, the guest isn't likely to distinguish the difference between the two. And, again, we're talking about whether AI can replace somms. And I think somms as they currently exist - there's a LOT that AI can replace, given where we currently are.
Anywho, I was trying to keep it simple for the comment section, but there is much to unpack here, as you said.
Maybe it's not what you meant by it, but saying that somms are cobbling together a description based on past writings implies that they aren't using subjective experience to provide honest personal opinions on a case by case basis. There's undoubtedly descriptors that "trained palates" are prone to using, but as you noted, there's a clear justification for the use of familiar descriptors so that people do understand them, and I strongly disagree that translating that subjective experience into language isn't personal.
As you sort of touch on here, homogeneity is a byproduct of familiarity, but it's true that getting more specific has its own pitfalls (though, I'm not sure what kinds of oranges are throwing you off 😂). Still, I think the differences are often nuanced with things like that – I don't think you lose someone when you get more specific within a familiar category for them; even if they can't tell you from memory the difference between a satsuma and a blood orange, it doesn't render the distinction meaningless and someone who has tasted a "standard issue orange" will still get a sense of what that means. Personally I think if you are being more specific in a way that makes sense to you and is honest, then as long as it isn't that rare flower you and 2 other people in existence smelled that one time you were on the top of Mt Fuji, specificity is usually a net positive.
As for the descriptors you called out, I totally get the point you're making - garrigue and gooseberry (which I had to order a can of online to familiarize myself with) are much more obscure to people in the US (though I think "fresh churned earth" is a pretty familiar scent to most people who have used a shovel). I will say that I refuse to use sensory descriptors that I am not familiar with and I go out of my way to familiarize myself with the ones that are commonly used in wine notes. It doesn't make those descriptors more useful though, and I tend to shy away from using them because I am aware of that level of unfamiliarity. This is something that Zach and I disagree on sometimes, but one thing we are aligned on is never providing notes that don't reflect what we are actually getting from a bottle. Something that is only possible if you've actually tasted the bottle in question.
Another key aspect I think you're overlooking here is that descriptors come up repeatedly for certain wines because there are patterns that exist. It's the reason people are able to guess wines blind in the first place. Undoubtedly that contributes to similar notes coming up again and again for different bottles from the same varietal or region, and it doesn't suggest some conspiratorial movement in the wine community to pretend those aromatics or flavors are there. It comes full circle back to using notes that people understand when you notice that the variety is lacking, but again, you have to balance familiarity with specificity as artfully as you can.
Ultimately what we were trying to hammer home in our post is that sensory experience is still uniquely human. The very important distinction between a human and an AI when describing a specific bottle is that the AI has never tasted it. It disregards the importance of individual specificity and trust of an individuals honest conveyance of their personal tastes. And I think saying that we're not talking about cobbling together our "own palate translations to another's - we're talking about whether an AI in a somm position is more or less useful than a human in that same role" misses the core point we were making, which is that an AI simply cannot describe a specific bottle of wine if humans haven't already done so. We aren't saying that getting a collated group of tasting notes from human sources provided by an AI won't be useful, we're saying that without people doing it first, it's literally impossible.
I'm *not* saying - nor have I said - that humans are cobbling together a description based on previous writings - I'm saying/asking: "what's the difference?" in the context of a somm service position. Questioning the difference of the result, especially as all AI writing is the result of something another human wrote. Whether it's that hiuman or another human standing in front of you - what's the difference in getting tasting notes culled from a stranger? And especially especially as our tasting notes, when shaped into words, are funneled through a cultural group think that is rarely personal or unique. People not into wine writing/reading or education will always surprise with notes that veer far away from the norm, and yet giving them these norms are somehow more uniquely human than an AI compliation of other humans? How? In what way?
I also never used a word or idea like "conspiratorial", and quite consciously so - if you need to reframe an argument as an illogical extreme in order to counter it, you're not actually countering it. Human being are gonna human, and sociological effects of group think are well documented, and "industries" are littered with them moreso than other groups. I'm equally as baffled as to how you think this is not a basic reality of human thinking and interactions in social situations, of which somms and the guests they cater to very much are.
Will we still need humans tasting and writing up wines in some way, somewhere, to some extenet? At least insofar as AI currently exists, that would be necessary, but that in no way = AI cannot replace somms, at least to a significant degree. And if the AI tech that allows them to breakdown aroma and taste like a human, that could very easily replace humans needing to do that in the first place.
Which, again, is not the same consideration as how much we do or shoiuld value the difference of a human vs. an AI. It's just looking very soberly at whether it CAN be done, and how much generations that AI will be native to and/or people who know so little about wine in the first place that they themsleves could never appreciate the difference, would ever care. And I'm thinking they won't, unless we give them a hell of a lot more reason to. If we don't, the replaceement isn't going to be hard, at least in many contexts.