DeepSeek Tried Chess... HUGE Mistake.

⭐️ Follow Me If You Are Amazing:

[GothamChess]

See you all in May 2025 for the Gotham European tour - Brussels, London, Vienna, Berlin, Munich, and Warsaw:

[arshiaprogaming]

Deepseek barely missed the bot championship

[michaelamra]

Time for the ultimate DeepSeek vs. ChatGPT AI match

[raul-ydob]

Brother got on the DeepSeek train immediately

[MrAzizul]

In Chinese chess, there is this piece named CANNON which can jump a piece. I think that explains Re1 and Rxe4.

[medusagorgonchsnnel]

this is genuinely one of the funniest videos ive seen and i actually laughed out loud like 6 times watching it. i know people have said this before but you really have a way of storytelling and engaging the audience

[dgamer]

Refusing to make a move? Sounds like a resignation to me. I'd say you win it.

[headlesshorseman]

Rematch with a "DeepThink (R1)" toggle, which uses different version of DeepSeek that is specifically designed for tasks that require comprehension.

[XiagraBalls]

In another universe, another version of Levy is playing an AI bot called SeekDeep, he's playing white and his Queens keep disappearing off the board.

[spizzlo]

Levy, if you happen to see this, you can give chatbots pretty in-depth instructions for how you want them to respond. You could start with “you are a chess grandmaster and we are going to play a game of chess. Each response will be one legal move. Do not elaborate on chess strategy or any other explanations. It’s extremely important to remember which pieces are left and which squares they are on”. Something along those lines.

An added benefit is I don’t think it would make as many mistakes because it will have more memory of the hands when it leaves out all the extra commentary.

Chatbots hallucinate, meaning they make things up and pass them off as absolute truth. But you can prompt them in a certain way to make them less likely to hallucinate.

[TheZombiesAreComing]

8:48
"The computer is a cheating bastard" is a trope for a reason.

[ThyVile]

I decided to try to play with Deepseek myself. My general takeaways are:

1. It's better not to use R1 reasoning because then it takes too long to make a turn and starts hallucinating too early. The reasoning part is a fairly fun read though.
2. If it makes an illegal move and you point it out the next move is either illegal or of blunder quality.
3. I suspect that the best way to "play" is to provide the whole game move set in PGN format in every message so it's easier for the model to work with. Might be more accurate but less meme-y.

Nice commentary as always, Levy. Удачи с гроссмейстерством!

[outpizzaddahut]

12:23 The atoms alighted perfectly TWICE for that rook.

[skumparn]

can't wait for the botclasico ChatGPT vs DeepSeek

[Pizzashahark]

19:56 we need stockfish vs deepseek. I’m already placing bets stockfish will win

[Alchemistic]

When your opponent refuses to play a move, that's resigning. However, rematch anyway.

[adriansticoid]

05:12 When the move is castle and Levy manually moves the rook after the king, you know the bot will summon pieces outta nowhere or make illegal moves later on.

[Kraken-wn]

0:00 Levy jumpscare
1:26 Chessly jumpscare
2:02 Chess board jumpscare

[Pihsrow]

I only want to see the 5 dimensional chess that DeepSeek and ChatGPT will play, probably being able to create new peaces with unique skills.

[SKOLinoYT]

I never really understood the concept of deusexmachina, but this is probably the strongest definition I've ever seen.