Is AI Involved?

I am starting to wonder if there is some AI behind the scenes that selects puzzles most likely to cause you to fail the longer your session is. In extra hard, I've noticed that if I make it, say, to 1650, suddenly, the puzzles will change a lot to trickier puzzles. This makes sense because, as you get higher, the puzzles get harder as well.

All that is normal until I lose a hundred points, fall back to where I started, and the puzzles stay harder. It almost seems as if it learned what puzzles trip me up. On extra hard, I often log in and breeze through a bunch of puzzles with motifs I recognize, hit a wall, and never recover until I log in the next day and start the cycle over again.

This is not a complaint because the point of doing these puzzles is to improve, and if there is an AI there to help me focus on motifs that I struggle with, they are even more valuable. I just thought I might start what might prove to be an interesting discussion on whether AI is involved with these puzzles now and, if not, whether they should involve AI.

Puzzles don't use "AI" in the way that this AI trend has been taking over in some spaces. Yes, it's a computer and yes there's some algorithm or system to how puzzles are randomly chosen. I don't work for chess.com coding with how puzzles are selected, but the way I understand it is something like this:

You get a puzzle somewhere in a rating range near your rating (maybe plus or minus a few hundred rating points). There are hundreds of thousands of puzzles in this rating range (if not more) and the puzzles are randomly selected. There's some nuances I am skipping over and also many more I am probably unaware of, but I don't believe the puzzles are generated maliciously in some "AI" attempt to keep you on the site longer.

I wasn't thinking it was malicious. Instead, I was thinking it was selecting puzzles to help focus on motifs where I am weak. The idea being that the more I'm exposed to them, the more likely I will be able to recognize the motif and choose the right moves. This makes me a better chess player faster. Also, it could be a simple algorithm that weights puzzle selection by the motifs I am failing at. This could be an opportunity, if it isn't already in place. I'm thinking it might already be in place based on my daily experience.

Hi. I'm not convinced that the motifs attached to individual puzzles are very reliable. Not unless chess.com has massively improved the way they label them. Used to be that members could suggest themes and there was some members who were authorised to give the green light. Later there was a phase where puzzles suddenly had scores of "themes", most of which were nonsense. I suspect there was a computer algorithm involved with that. It felt like almost every puzzle had the "Passed Pawns" theme attached to it. Most of the time the solution to the puzzle had nothing whatever to do with PPs, and sometimes there weren't even any PPs on the board. There were other themes like that too. e.g. The theme would be Opposite coloured bishops, and the puzzle was a mate in one in which no bishops (or any other clergy) were involved.

Given the vast number of TT puzzles I rather doubt that any human has gone through classifying them all, and if the classification has been done by computer then I very much hope the algorithm is a lot better than the old one.

AI in its current state being incredibly stupid I would think people would be happy it was composing puzzles.

I would say, completely stupid, in the real sense of the word "stupid". Because, despite the name, AI is not even vaguely intelligent. The way computers process data has not really changed since computers were invented. The data has certainly changed (no longer just numbers, letters, symbols), the speed has changed massively, and therefore the amount of data that can be processed quickly has multiplied massively, but there is nothing involved that most of us would call intelligence. (MHO).

You can doubt it, but that's what happened. Not by one human, obviously, but by the entire community. For many years, users could assign "themes" to the puzzles, and up- or downvote existing ones. Also, players could report faulty puzzles. Also, the rating of puzzles corrected itself, going down or up as people soled or failed them. Then the developers deleted these options somewhere around 2018 or 19 for some reason.

The system has more than enough training data to predict what puzzles a player would be weak at and use a neural network to select the puzzles a player would most likely fail at. It feels like there is something in place for that already and that makes it even more valuable. Classical neural networks are simple to train and apply with this large of a dataset. Heck, give me pytorch and I could do it in a weekend, and I know chess.com is interested in AI, so at this point I'm convinced they already did that... and good. That is exactly where AI best shines.

Let's hope that no one from the staff reads this. The last thing chess.com (or the world) needs is AI screwing up something hopelessly. Again.

I hear the frustration. ChatGPT and other LLMs are something different. It's that we used to call "machine learning" for decades until someone invented the transformer model and a man started dreaming of a 1.3 trillion dollar IPO.

Which is going to collapse soon. As soon as people realize that LLMs go nowhere and they won't ever get close to intelligence.

The worst part will be when an LLM winds up training on output created by other LLMs. Hallucinations of hallucinations will lead to information collapse. The most useful LLMs are the ones we have right now. Machine learning will stay though. This is why I have avoided working on AIs that generative content, preferring to focus on solving problems that require classical neural networks that do not create an input feedback loop like chess and signal analysis. Those will always be useful.

neural networks and llms are two separate concepts

llms utilize the former but are more complicated and messy than the former

neural networks have been around for decades