ChatGPT accords

[mnitabach]

What's sarcastic? I actually live in 1992 and I've never heard of it so it must have been invented later.

Yes computers beat human champs, but people said they never would years before it happened, which parallels saying they can't do other things like create good fragrances.

We will obviously have to wait & see what transpires with regard to other types of "AI", but IMO the particular form of non-symbolic AI represented by LLMs & image-creation algorithms will not succeed at expert-level perfume composition (just as they can't even follow the rules of chess). The reason chess is so amenable to algorithmic expertise is that the causal structure of the universe is completely trivial & easily represented by a compact symbolic model. This is pretty much the opposite of perfumery. But we shall see!

 

[xii]

If Magnus Carlsen did perfumery in the very manner he does chess, he would evaluate daily hundreds if not thousands of blends without need of ever making notes. Chess position evaluation is very similar to blend evaluation; the size of the space sampled, or its finiteness for that matter, is utterly irrelevant. For rather obvious reasons. Now, a classical chess engine does essentially the same thing with somewhat better accuracy: it won’t blunder, ever. Such an engine doesn’t rely on machine learning and will beat every chess player in the world, Magnus Carlsen being arguably the strongest human chess player of all time. Leela Chess Zero, a deep learning chess engine utilises only a tiny fraction of the node space size of classical engines but stands ground against them.

 

[tyriekm]

Whoa! Sea salt with 20% Ethyl Maltol that's a pretty sweet salt

 

[David Ruskin]

. Chess position evaluation is very similar to blend evaluation

Except that one is spacial, and the other is to do with the sense of smell. Evaluating a fragrance is, or should be, quite complex, involving many criteria.

 

[mnitabach]

Except that one is spacial, and the other is to do with the sense of smell. Evaluating a fragrance is, or should be, quite complex, involving many criteria.

Exactly. A chess computer has a complete closed-form representation of the state of the universe & possible moves AND the definition of success or failure is purely objective & unambiguous & trivially hard-coded in the algorithm. Designing perfume compositions is about as opposite to this as is conceivable.

 

[jsweet]

Exactly. A chess computer has a complete closed-form representation of the state of the universe & possible moves AND the definition of success or failure is purely objective & unambiguous & trivially hard-coded in the algorithm. Designing perfume compositions is about as opposite to this as is conceivable.

My thoughts precisely. A game is a closed system that can entirely be described through language. Nothing in real life works like that.

 

[jfrater]

All I know is a computer will never beat a human chess champion. I wasn't impressed with "AI" until I was, with midjourney making really cool movie screenshots. It won't surprise me in the least when a program invents a great fragrance, which is only a matter of time. (Though I wonder how people will decide which ones to actually make and test.) All the philosophy about what it takes to be creative doesn't matter much if it smells good and unique.

I like your last sentence except the "unique" bit. Looking for the "unique" has driven creative people away from beauty to novelty. "Uniqueness" is not, in my opinon, relevant at all to something being beautiful.

 

[jfrater]

One of the problems with texts on the internet is Google. I worked for a search engine optimisation office for a summer once during my uni years. That was really disillisioning. When writing a text on a certain topic for a webpage, algorithms tell you which keywords to use and how many times to use them in order for Google to give the page a better rank. (You don't think a keyword makes much sense in that specific context, you're struggling to integrate it in the text? Doesn't matter. Put it in.) Then there's the problem that using copied text hurts your ranking. So if you're offering a product sold on lots of other platforms and it comes with a ready-made advert text, you need to reformulate that so that it looks more "original"/"new", or risk a bad ranking. I'd expect that AI is much better at this, especially since it doesn't get frustrated.

This is part of what has destroyed much of the usable internet in my opinion. I owned and ran listverse.com which is a top 10 lists site and I did most of the writing initially. It was reliant on ad revenue. Every time I met with my account reps at Google they would give me an ever diminishing list of "requirements" that were "not mandatory" but were I not to comply it would reduce considerably my ability to make money from my content. I told Google they were tactily censoring me and they immediately (and excessivly) went into defensive mode stating that they were absolutely not censoring content.

I had staff to pay so I had no choice. I allowed Google to begin dictating what content was to be allowed. Eventually I found relacement advertisers and was able to resume writing the content I wanted and loved. Google consequently began "shadow banning" the site. I sold the business.

 

[xii]

I doubt I can reasonably contribute to this thread. At a beverage of choice and lots of paper, perhaps... Instead, some food for thought, feel free to consider irrelevant: Bowerbird males build stunningly elaborate altars to attract females. The rate of success of a randomly chosen altat is very low because the winner pretty much takes all. So it seems very hard to observe them mate. Fortunately, we, humans can tour these altars, pick the prettiest and it happens to be very likely bowerbird females will pick the very one.

 

[ErroneousFact]

I like your last sentence except the "unique" bit. Looking for the "unique" has driven creative people away from beauty to novelty. "Uniqueness" is not, in my opinon, relevant at all to something being beautiful.

I just meant that if a computer makes something like existing fragrances it won't impress many people, since we didn't need a computer to come up with it.

 

[G&W]

To all people comparing chess to god knows what... I suggest you to code a simple chess program (little tip: Beta and Alpha algo), then code or at least compare it to a deep learning chess AI to see the difference.
Not wanting nor having time to pick up a thread fight here, but you can't talk effectively about AI without coding at least a pseudo AI let me tell you that.
And to whom affirming that there is no AI capable of beating a chess champion.. you need a history udpate my friends !
Cheers

 

[ErroneousFact]

And to whom affirming that there is no AI capable of beating a chess champion.. you need a history udpate my friends !
Cheers

I've already said I live in 1992 and my knowledge is all up to date, thank you very much.

 

[Darren Alan]

No Darren, you really, really don't. You can be much more creative and original than that.

You're right. It is kind of crazy what it generates though. But if you look at all the details that Grant loaded into the prompt, he practically wrote it himself anyway. The AI just filled in the fluff...

*Edited for typos

 

[G&W]

A chess computer has a complete closed-form representation of the state of the universe & possible moves AND the definition of success or failure is purely objective & unambiguous & trivially hard-coded in the algorithm

For anyone reading this enormity, please look otherwise for IT knowledge.
@mnitabach Why are you talking about what you clearly don't know a single bit about?!!

 

[steventeddy]

I agree with jsweet's position on whether or not chess position evaluation serves as a good analogy to perfume evaluation. Namely, it's a poor analogy because all legal chess positions must by definition conform with the rules of chess. I.e. the set of chess rules in one's mind always corresponds with the chess position, be it laid out on a board, or in your mind's eye. On the other hand, perfume has no such guarantee, i.e. there is no necessary connection between what one judges as a "good formula" in their mind's eye compared to whether or not they will judge it as a good formula when they smell it. (Indeed, chess was used by my Kant professor as the example to illustrate the so-called "problem of representation"--loosely speaking, correspondence between ideas in ones mind and objects in the real world).

I believe the amenability of chess to at least one non-symbolic machine learning approach has been demonstrated. In an arXiv paper about "AlphaZero" from 2017, a deep neural network was trained (by playing against itself) starting with random play and knowledge of the game rules and got to a point where it routed Stockfish, which in my understanding is a symbolic AI approach. The authors say that the neural network takes in the board position as input, and outputs probabilities for candidate moves. One can search for "AlphaZero vs Stockfish" on youtube to see some analyses of some games that were released by DeepMind. The way that Stockfish was beaten was very interesting to say the least.

I would argue that chess position evaluation can involve many criteria, at least from a human perspective. That and chess having a finite number of positions/a well defined notion of "good" doesn't buy you very much in evaluation. For the second/third points, in short, the search space is still huge. If at a first pass we define a good chess position as one that leads to a forced mate (for white lets say), then even with only 7 pieces on the board, that is very difficult (with respect to the amount of computation required). Not only are there many possible legal positions with 7 pieces (423,836,835,667,331 to be exact), the forced mate could be 100+ moves. Indeed, there is a tablebase that contains the outcomes (and the optimal move order to achieve such outcome) of all legal positions with 7 pieces. Conservatively, it's 18TB in size. Here is the wiki page on endgame tablebases.

Of course, if we retreat from that position and just want to know if a position is 'winning', whatever that might mean, a human would consider a multitude of factors. For example: space on the board, good/bad bishops (is your bishop blocked behind your own pawns), doubled pawns, knight outposts (knights defended by a pawn, but cannot be attacked by an enemy pawn), rank/file control, king safety, piece harmony (are your pieces defending each other). To my knowledge, Stockfish evaluates a position by considering factors like this, weighting them to compute a score for whether a position is 'winning' (where this weighting is hand tuned). I agree that one would just be looking at how the pieces are placed on the board, but as a human, I ask what the strengths and weaknesses of each player's positions are, and use those to predict whether a position is good or not.

 

[mnitabach]

I agree with jsweet's position on whether or not chess position evaluation serves as a good analogy to perfume evaluation. Namely, it's a poor analogy because all legal chess positions must by definition conform with the rules of chess. I.e. the set of chess rules in one's mind always corresponds with the chess position, be it laid out on a board, or in your mind's eye. On the other hand, perfume has no such guarantee, i.e. there is no necessary connection between what one judges as a "good formula" in their mind's eye compared to whether or not they will judge it as a good formula when they smell it. (Indeed, chess was used by my Kant professor as the example to illustrate the so-called "problem of representation"--loosely speaking, correspondence between ideas in ones mind and objects in the real world).

I believe the amenability of chess to at least one non-symbolic machine learning approach has been demonstrated. In an arXiv paper about "AlphaZero" from 2017, a deep neural network was trained (by playing against itself) starting with random play and knowledge of the game rules and got to a point where it routed Stockfish, which in my understanding is a symbolic AI approach. The authors say that the neural network takes in the board position as input, and outputs probabilities for candidate moves. One can search for "AlphaZero vs Stockfish" on youtube to see some analyses of some games that were released by DeepMind. The way that Stockfish was beaten was very interesting to say the least.

I would argue that chess position evaluation can involve many criteria, at least from a human perspective. That and chess having a finite number of positions/a well defined notion of "good" doesn't buy you very much in evaluation. For the second/third points, in short, the search space is still huge. If at a first pass we define a good chess position as one that leads to a forced mate (for white lets say), then even with only 7 pieces on the board, that is very difficult (with respect to the amount of computation required). Not only are there many possible legal positions with 7 pieces (423,836,835,667,331 to be exact), the forced mate could be 100+ moves. Indeed, there is a tablebase that contains the outcomes (and the optimal move order to achieve such outcome) of all legal positions with 7 pieces. Conservatively, it's 18TB in size. Here is the wiki page on endgame tablebases.

Of course, if we retreat from that position and just want to know if a position is 'winning', whatever that might mean, a human would consider a multitude of factors. For example: space on the board, good/bad bishops (is your bishop blocked behind your own pawns), doubled pawns, knight outposts (knights defended by a pawn, but cannot be attacked by an enemy pawn), rank/file control, king safety, piece harmony (are your pieces defending each other). To my knowledge, Stockfish evaluates a position by considering factors like this, weighting them to compute a score for whether a position is 'winning' (where this weighting is hand tuned). I agree that one would just be looking at how the pieces are placed on the board, but as a human, I ask what the strengths and weaknesses of each player's positions are, and use those to predict whether a position is good or not.

Exactly my point: the chess search space is of course beyond enormous & cannot be traversed even remotely densely, but the causal structure of the space is trivially & unambiguously describable AND the success or failure of ultimate outcome (checkmate) is also trivially & unambiguously describable. While chess & perfumery share exponential enormity of the search space, the causal structure of the two universes, their amenability to compact unambiguous representation, and the unambiguous detection & representation of the success or failure of ultimate outcomes could not be more different.

 

[xii]

OK, I can work with that. I'll be more than glad to ditch chess analogy.

Well, once we're clear causality, spatiality, compactness and other somewhat embarrassing notions are of no relevance whatsoever.

I have a long, history of implementing tree pruning for game strategy searching. It's also a rather distant past but not much has really changed: we navigate through a complete game tree by evaluating its branches. For instance, in solved games the evaluation function returns win, loose ar draw for every branch. Such a function can be hard to compute, but there are surprises. For instance in Renju, a glorified version of five-in-a-row, I actually found a fancy proof of existence of an essentially linearly computable function solving it. There obviously exists an exponentially computable function running through the entire game tree but computing it is a hopeless task. The full game tree of chess likely contains more information humans have ever verbalised.

On the other hand, we know skilled players can just look at the position and have a pretty good idea about the outcome. Which pleased proponents of the rather amusing idea that human brain is so very powerful to no end. Somewhat less optimistic view on the matter is, if we can train ourselves to evaluate a position in a game of go, then perhaps it is because there exists a quickly computable and effective evaluation function that simply takes so many factors into account that we just couldn't see. This doesn't apply only to board games, but, in fact, to any activity where the role of language is emphasised. And that leads to perfumery, BN DYI perfumery in particular.

Students in French perfume schools are often told they will learn all the tricks of the trade but creativity. I keep meeting these students and can absolutely confirm their skill is impressive but they revere creativity way more. BN is dominated by western people mostly from Anglosphere so no wonder there is so much bias towards analytic approach here. We often read about truth, common sense, empiricism and whatnot, as opposed to speculation, intuition, anything broad really. However, analytic approach inadvertently reduces all matters to study of very confined spaces ruled by doctrines with very little nuance. I find it fascinating and, to be honest, feel rather comfortable in this sort of a habitat. But I can't shake the feeling ChatGPT is literally thriving in it.

 

[Leshutch]

People should have no doubt that the first 'AI' designed fragrance will be available within the next 5 years and it will be an absolute commercial success.

 

[mnitabach]

People should have no doubt that the first 'AI' designed fragrance will be available within the next 5 years and it will be an absolute commercial success.

I agree, albeit with the proviso that all the "AI" will do is generate test formulas, which will then be evaluated by human beings & iterated by hand. There is IMO zero chance that the first version of a formula emitted by an "AI" will be a quality perfume by any definition: commercial success, aesthetic merit, or any other.

 

[mnitabach]

I agree, albeit with the proviso that all the "AI" will do is generate test formulas, which will then be evaluated by human beings & iterated by hand. There is IMO zero chance that the first version of a formula emitted by an "AI" will be a quality perfume by any definition: commercial success, aesthetic merit, or any other.

Note this particular extreme difference btwn chess & perfumery: we know how human chess players evaluate positions & we can program computers to do so, but we have effectively zero idea how human beings perceive even the simplest accords, let alone complete perfumes.