Your network blocks the Lichess assets!

lichess.org
Donate

AI Slop is Invading the Chess World

I don't know if we will ever achieve AGI (I have a very strong doubt on that; I don't think we humans are that good at anything, especially programming, or encapsulating the essence of human reasoning in particular.) Till then, this kind of slop will happen.

In the end, LLMs are but a [probability calculator] in a nutshell. It doesn't reason. Or think. It parrots out what the probability function says is the most likely optimal output. So, all the LLM users need to be reminded that, one just cannot outsource thinking to LLMs. (which is exactly what these slops attempt to do - and fail.)

I don't know if we will ever achieve AGI (I have a very strong doubt on that; I don't think we humans are that good at anything, especially programming, or encapsulating the essence of human reasoning in particular.) Till then, this kind of slop will happen. In the end, LLMs are but a [probability calculator] in a nutshell. It doesn't reason. Or think. It parrots out what the probability function says is the most likely optimal output. So, all the LLM users need to be reminded that, one just cannot outsource thinking to LLMs. (which is exactly what these slops attempt to do - and fail.)

@RuyLopez1000

hey, I think you did a mistake.. i don't think chatgpt has only 1.8 billion parameters, that's very low, it's around 175 billion for gpt 3.5 and 1.8 trillion for gpt 4...

@RuyLopez1000 hey, I think you did a mistake.. i don't think chatgpt has only 1.8 billion parameters, that's very low, it's around 175 billion for gpt 3.5 and 1.8 trillion for gpt 4...

For your first question Google AI gives the correct answer: https://share.google/aimode/vx7ZL2O51kKf5vrVl

That said, indeed they made errors: https://share.google/aimode/A10pzhYMnEQFqRUBB

And they still do: https://share.google/aimode/NZAD0sbdkkVulzavw

Edit: I updated that last link, but it was the wrong and now I lost the old. Hopefully that one works:

https://www.google.com/search?smstk=ChhscVRBbzV0NTgzQlZuSWsrQlo3TUc1bz0QAQ%3D%3D&smstidx=0&q=What+is+the+evaluation+of+the+following+chessposition,+given+as+FEN?%0A%0A4k3/8/8/8/8/8/7P/4KB2+w+-+-+0+1&udm=50&csuir=1&aep=34&shndl=37&shmd=H4sIAAAAAAAA_3WOsU4DMRBECeV9AtXWiJxzEAGKIkVCCg0SoqO0NueNbcX2Wl4n5hf4Nb6KCwUNQtPNaOZN93nZ0bvDCl6gOgI6YThi9ZyA9z_OnkPg5pOF0ZFIZvHn-AasP1ECFHjevm66bnm4U4-_enhTy5enW2gwn7SA4Wrtas2yUqq11lupE2TsR45KCMvo5rlwZIU-anFYSFd3jLuEPvQ52euLr5n-b8BHtCRqVzCZ6aeyzDaQ1bag8ZSqGj7-eHrqJ4PF6OF-YfKZ8Q0mIzKIDAEAAA&shmds=v1_AdeF8Kj8mes8_8kacn41mQoP3WGdDHEC99WybM6eqjDgsIsHog&source=sh/x/aim/m1/1&kgs=14ee2eb1d147a89a

For your first question Google AI gives the correct answer: https://share.google/aimode/vx7ZL2O51kKf5vrVl That said, indeed they made errors: https://share.google/aimode/A10pzhYMnEQFqRUBB And they still do: https://share.google/aimode/NZAD0sbdkkVulzavw Edit: I updated that last link, but it was the wrong and now I lost the old. Hopefully that one works: https://www.google.com/search?smstk=ChhscVRBbzV0NTgzQlZuSWsrQlo3TUc1bz0QAQ%3D%3D&smstidx=0&q=What+is+the+evaluation+of+the+following+chessposition,+given+as+FEN?%0A%0A4k3/8/8/8/8/8/7P/4KB2+w+-+-+0+1&udm=50&csuir=1&aep=34&shndl=37&shmd=H4sIAAAAAAAA_3WOsU4DMRBECeV9AtXWiJxzEAGKIkVCCg0SoqO0NueNbcX2Wl4n5hf4Nb6KCwUNQtPNaOZN93nZ0bvDCl6gOgI6YThi9ZyA9z_OnkPg5pOF0ZFIZvHn-AasP1ECFHjevm66bnm4U4-_enhTy5enW2gwn7SA4Wrtas2yUqq11lupE2TsR45KCMvo5rlwZIU-anFYSFd3jLuEPvQ52euLr5n-b8BHtCRqVzCZ6aeyzDaQ1bag8ZSqGj7-eHrqJ4PF6OF-YfKZ8Q0mIzKIDAEAAA&shmds=v1_AdeF8Kj8mes8_8kacn41mQoP3WGdDHEC99WybM6eqjDgsIsHog&source=sh/x/aim/m1/1&kgs=14ee2eb1d147a89a

Of course a dumb techbro who knows nothing about AI would use AI generated images to make his AI slop article. Fuck off. Nobody's obligated to pay you.

Of course a dumb techbro who knows nothing about AI would use AI generated images to make his AI slop article. Fuck off. Nobody's obligated to pay you.

I use AI to have some sort of a mindset, and to glaze my wins. You can win because you took your opponent's queen when before you were losing, and the AI will say "You dominated positionally this game. Your opponent resigned because of your space on the board and your control and attack." Never use AI to analyze your games. Actually, you can, but fix it, because it will miss tactics.

I use AI to have some sort of a mindset, and to glaze my wins. You can win because you took your opponent's queen when before you were losing, and the AI will say "You dominated positionally this game. Your opponent resigned because of your space on the board and your control and attack." Never use AI to analyze your games. Actually, you can, but fix it, because it will miss tactics.

@miksuu15 said in #39:

  • 1.8B parameters would be considered an SLM, not an LLM. In early 2023, GPT-4 was rumored to have ~1.76T parameters, so I’m guessing the author rounded that up alongside with the wrong amount. Later iterations don’t have any widely accepted leaks about parameter count.

Good spotting! Someone else also pointed this out to me. Should be trillion, not billion, I edited the blog to fix it.

  • A token isn’t a word; it’s closer to ~0.75 words on average. Many common words are one token, but it depends on capitalization and context. Either way, the article’s claim is false.

But that isn't where I got the 300 billion from, I got it from this link: 'OpenAI, the company behind ChatGPT, fed the tool some 300 billion words systematically scraped from the internet: books, articles, websites and posts – including personal information obtained without consent.' https://www.sydney.edu.au/news-opinion/news/2023/02/08/chatgpt-is-a-data-privacy-nightmare.html

(They made the conversion without taking into account what you said below it seems). Anyway, I edited the blog to reflect that ChatGPT-4 was trained on 1PT of data. https://seifeur.com/chat-gpt-4-data-size/

  • It predicts the next token, not the next word.
  • The author tries to familiarize the reader with the term “token” by defining it as 1:1 with “word,” which is misleading.

Thanks for the tip! I edited the blog to clarify that tokens are not necessarily a whole word.

  • Correlating “parameters” with adjusting the output doesn’t make any sense.

Parameters are the backbone of producing the output.

'LLM parameters are the settings that control and optimize a large language model’s (LLM) output and behavior. '
https://www.ibm.com/think/topics/llm-parameters

  • Using the internet as training data is not clearly “theft”, the case is just not legally settled. I think that only content behind a paywall should be illegal to scrape, but again, this is just my opinion.

Saying that something couldn't be theft unless someone in a wig and a gavel says so doesn't feel right. Morality doesn't equal law. You absolutely can call something theft without some institution first saying so.

  • It’s true that LLMs can make logical errors (especially older or non-reasoning models). However, for example, this is even with reasoning mode turned off in ChatGPT’s default mode (I don’t know which model the author used, and the site might utilizing smaller/older ones):

I said it in the article: perchance.org. I put it on reasoning mode. https://perchance.org/ai-text-generator

It’s true that some companies are building bad products with AI while capitalizing on the hype. However, there’s no doubt the technology can be utilized, especially when the product is built well.

How? I'd like to hear. No one has succeeded.

@miksuu15 said in #39: > - 1.8B parameters would be considered an SLM, not an LLM. In early 2023, GPT-4 was rumored to have ~1.76T parameters, so I’m guessing the author rounded that up alongside with the wrong amount. Later iterations don’t have any widely accepted leaks about parameter count. Good spotting! Someone else also pointed this out to me. Should be trillion, not billion, I edited the blog to fix it. > - A token isn’t a word; it’s closer to ~0.75 words on average. Many common words are one token, but it depends on capitalization and context. Either way, the article’s claim is false. > - GPT-3’s training data was described as ~300B tokens, not words (July 2020; source: https://arxiv.org/pdf/2005.14165.pdf). But that isn't where I got the 300 billion from, I got it from this link: 'OpenAI, the company behind ChatGPT, fed the tool some 300 billion words systematically scraped from the internet: books, articles, websites and posts – including personal information obtained without consent.' https://www.sydney.edu.au/news-opinion/news/2023/02/08/chatgpt-is-a-data-privacy-nightmare.html (They made the conversion without taking into account what you said below it seems). Anyway, I edited the blog to reflect that ChatGPT-4 was trained on 1PT of data. https://seifeur.com/chat-gpt-4-data-size/ > - It predicts the next token, not the next word. > - The author tries to familiarize the reader with the term “token” by defining it as 1:1 with “word,” which is misleading. Thanks for the tip! I edited the blog to clarify that tokens are not necessarily a whole word. > - Correlating “parameters” with adjusting the output doesn’t make any sense. Parameters are the backbone of producing the output. 'LLM parameters are the settings that control and optimize a large language model’s (LLM) output and behavior. ' https://www.ibm.com/think/topics/llm-parameters > - Using the internet as training data is not clearly “theft”, the case is just not legally settled. I think that only content behind a paywall should be illegal to scrape, but again, this is just my opinion. Saying that something couldn't be theft unless someone in a wig and a gavel says so doesn't feel right. Morality doesn't equal law. You absolutely can call something theft without some institution first saying so. > - It’s true that LLMs can make logical errors (especially older or non-reasoning models). However, for example, this is even with reasoning mode turned off in ChatGPT’s default mode (I don’t know which model the author used, and the site might utilizing smaller/older ones): I said it in the article: perchance.org. I put it on reasoning mode. https://perchance.org/ai-text-generator > It’s true that some companies are building bad products with AI while capitalizing on the hype. However, there’s no doubt the technology can be utilized, especially when the product is built well. How? I'd like to hear. No one has succeeded.

@NHL_1024 said in #43:

For your first question Google AI gives the correct answer: https://share.google/aimode/vx7ZL2O51kKf5vrVl

I used perchance.org AI model https://perchance.org/ai-text-generator. Probably it's less powerful. So the top AI are able to answer the question it seems.

That said, indeed they made errors: https://share.google/aimode/A10pzhYMnEQFqRUBB

And they still do: https://share.google/aimode/NZAD0sbdkkVulzavw

Edit: I updated that last link, but it was the wrong and now I lost the old. Hopefully that one works:

https://www.google.com/search?smstk=ChhscVRBbzV0NTgzQlZuSWsrQlo3TUc1bz0QAQ%3D%3D&smstidx=0&q=What+is+the+evaluation+of+the+following+chessposition,+given+as+FEN?%0A%0A4k3/8/8/8/8/8/7P/4KB2+w+-+-+0+1&udm=50&csuir=1&aep=34&shndl=37&shmd=H4sIAAAAAAAA_3WOsU4DMRBECeV9AtXWiJxzEAGKIkVCCg0SoqO0NueNbcX2Wl4n5hf4Nb6KCwUNQtPNaOZN93nZ0bvDCl6gOgI6YThi9ZyA9z_OnkPg5pOF0ZFIZvHn-AasP1ECFHjevm66bnm4U4-_enhTy5enW2gwn7SA4Wrtas2yUqq11lupE2TsR45KCMvo5rlwZIU-anFYSFd3jLuEPvQ52euLr5n-b8BHtCRqVzCZ6aeyzDaQ1bag8ZSqGj7-eHrqJ4PF6OF-YfKZ8Q0mIzKIDAEAAA&shmds=v1_AdeF8Kj8mes8_8kacn41mQoP3WGdDHEC99WybM6eqjDgsIsHog&source=sh/x/aim/m1/1&kgs=14ee2eb1d147a89a

Amusing example. AI thinks h8 is a light square.

Also they recommend Kd8 after h4, a trolling move XD.

@NHL_1024 said in #43: > For your first question Google AI gives the correct answer: https://share.google/aimode/vx7ZL2O51kKf5vrVl I used perchance.org AI model https://perchance.org/ai-text-generator. Probably it's less powerful. So the top AI are able to answer the question it seems. > That said, indeed they made errors: https://share.google/aimode/A10pzhYMnEQFqRUBB > > And they still do: https://share.google/aimode/NZAD0sbdkkVulzavw > > Edit: I updated that last link, but it was the wrong and now I lost the old. Hopefully that one works: > > https://www.google.com/search?smstk=ChhscVRBbzV0NTgzQlZuSWsrQlo3TUc1bz0QAQ%3D%3D&smstidx=0&q=What+is+the+evaluation+of+the+following+chessposition,+given+as+FEN?%0A%0A4k3/8/8/8/8/8/7P/4KB2+w+-+-+0+1&udm=50&csuir=1&aep=34&shndl=37&shmd=H4sIAAAAAAAA_3WOsU4DMRBECeV9AtXWiJxzEAGKIkVCCg0SoqO0NueNbcX2Wl4n5hf4Nb6KCwUNQtPNaOZN93nZ0bvDCl6gOgI6YThi9ZyA9z_OnkPg5pOF0ZFIZvHn-AasP1ECFHjevm66bnm4U4-_enhTy5enW2gwn7SA4Wrtas2yUqq11lupE2TsR45KCMvo5rlwZIU-anFYSFd3jLuEPvQ52euLr5n-b8BHtCRqVzCZ6aeyzDaQ1bag8ZSqGj7-eHrqJ4PF6OF-YfKZ8Q0mIzKIDAEAAA&shmds=v1_AdeF8Kj8mes8_8kacn41mQoP3WGdDHEC99WybM6eqjDgsIsHog&source=sh/x/aim/m1/1&kgs=14ee2eb1d147a89a Amusing example. AI thinks h8 is a light square. Also they recommend Kd8 after h4, a trolling move XD.

The statement claims "never less than 32," which is incorrect—the correct phrasing would be "never more than 32" or "at least 32."
More or least at least
or immoralist which may be a list of immorals at least or morralist where i m moral which would never be immoralist, never the least its false unless its never less or never more than 32.

  • I think I'll just sit by, have a dirty gin martini and enjoy the slop, I mean show. Thank you Ruy Lopez!! ;)
The statement claims "never less than 32," which is incorrect—the correct phrasing would be "never more than 32" or "at least 32." More or least at least or immoralist which may be a list of immorals at least or morralist where i m moral which would never be immoralist, never the least its false unless its never less or never more than 32. * I think I'll just sit by, have a dirty gin martini and enjoy the slop, I mean show. Thank you Ruy Lopez!! ;)

I tested a few more LLMs, and turns out, most of them have issues with that relatively simple question:

  • ChatGPT, not logged in: wrong answer (no link sharing supported, so I give a quote):

The statement "There are never less than 32 empty squares on a chess board" is false.

Here's why:

  • A chess board has 64 squares in total (8x8).
  • At the start of the game, there are 32 pieces (16 for each player), which means 32 squares are occupied by pieces.
  • The remaining 32 squares are empty at the start of the game.

However, during the course of a game, pieces can be captured and removed from the board, which can leave more than 32 squares empty. In some positions, it's possible for fewer than 32 squares to be empty, especially in the endgame when fewer pieces remain on the board.

For example, if only a few pieces are left (like one or two kings), there could be fewer than 32 empty squares remaining. So, it is not guaranteed that there are always at least 32 empty squares.

That shows that there is still a long way to go until AGI, if we ever get there.

I tested a few more LLMs, and turns out, most of them have issues with that relatively simple question: * ChatGPT, _not_ logged in: wrong answer (no link sharing supported, so I give a quote): > The statement **"There are never less than 32 empty squares on a chess board"** is **false**. > > Here's why: > > * A chess board has 64 squares in total (8x8). > * At the start of the game, there are 32 pieces (16 for each player), which means **32 squares are occupied** by pieces. > * The remaining **32 squares are empty** at the start of the game. > > However, during the course of a game, pieces can be captured and removed from the board, which can leave more than 32 squares empty. In some positions, it's possible for fewer than 32 squares to be empty, especially in the endgame when fewer pieces remain on the board. > > For example, if only a few pieces are left (like one or two kings), there could be fewer than 32 empty squares remaining. So, it is not guaranteed that there are always **at least** 32 empty squares. * ChatGPT, logged in: correct answer (but I already had asked the question when not logged in, so it may have learned from that, or it is a stronger model, or both): https://chatgpt.com/share/694b4186-f9c4-8006-86de-cf99b9bcb8ea * Grok, not logged in: wrong answer: https://grok.com/share/c2hhcmQtMg_eaacd901-f006-4817-bb9e-7c367279260f * Claude, logged in: wrong answer: https://claude.ai/share/4aaee5f5-7bc6-4022-b195-5fe90a93adea That shows that there is still a long way to go until AGI, if we ever get there.