Deep Roots

Curiosity Plug #02

The next magician of 64 squares

Chess is an amazing example of the depths to which inference can run when pushed past the surface. If you analyze the games of chess players, you'll be able to classify them into roughly the following fuzzy categories:

  1. Positional : Tigran Petrosian, Anatoly Karpov, etc. Brilliant defense, strategic, solid.
  2. Attacking/Tactical : Mikhail Tal, Garry Kasparov, etc. Fearless sacrifices, artistic tactics, magic.
  3. Universal : Magnus Carlsen, Bobby Fischer, etc. Exceptional at all three levels of the game. Squeezed water from stone.
  4. Creative : Richard Réti, Alexander Morozevich, etc. Highly creative and unorthodox openings, risk efficient, strategic play.
  5. Pragmatic : Vladimir Kramnik, Vishwanathan Anand, etc. Deep strategic understanding, positional and tactical brilliance, effective at developing counter plays on the move.
  6. Modern theoreticians : Fabiano Caruana, Wesley So, etc. Solid preparation, Strong calculational capabilities, technical precision.
  7. Defensive : Peter Leko, Viktor Korchnoi, etc. Strong openings, solid defensive capabilities.
  8. Dynamic/Modern : Hikaru Nakamura, Alireza Firouzja, etc. Aggressive, tactical style. Crafty.

It can be argued, rightly, that players cannot be categorized into these rigid groups since styles flow, mix and blend. Tal displayed exceptional strategic depth in certain games, while Caruana displayed fluidic tactics in certain positions. So it boils down to the nature of the game as it unfolds. Game theory could have some answers, but what tickles my brain particularly is how the players infer their plays as the game goes on. How do they tweak strategies and juggle outcomes on the fly? Game engines like Stockfish and Leela are excellent at going to incredible depths, and it appears that AI is actually introducing fresh possibilities into the game hitherto unseen.

That begs the question,

Is inference king?

Any intelligent system operates on roughly two fronts :

  1. Training : Here data is paramount. The system is fed loads of data and allowed to build connections among them.
  2. Inference : Here the model does the actual job - matching the output to reality.

Now imagine you're an AI and brawling it out against Garry Kasparov. You contain the knowledge of the old and new masters, and even more positions than Garry could keep in his head, and he still can keep insane amount of information in his head. And yet, you find yourself struggling to mate him, and sometimes to even find a winning advantage. Why?

No matter how good your data is, inference is where AI really shines. Models like Deepseek, Claude, Gemini, GPT, etc. are excellent at inference, but are only limited to queries as per their training. Hence, they cannot reason effectively. While [recently](tab: Google’s DeepMind AI Can Solve Math Problems on Par with Top Human Solvers | Scientific American) Alpha Geometry2 AI from Google managed to reach Gold medal levels at IMO, we're still far from some insane prodigious talent.

And I believe, as Sequoia has articulated it well, AI's true frontier is Inference, and that is where we shall find AGI, if we do find it. So we might find the next Tal or Fischer somewhere hidden deep inside the zoo of barely imagined AI models. And there might be a new class of players altogether in the list that surpasses the levels of the game altogether. We're using Fischer random and various other forms of Chess now to explore the uncharted territories hidden in the 64 squares, but the real magic just might be waiting on the board, silent and hidden in plain sight.

So, do we have the eyes to really see beyond the wall?

Let's hope we do, and till then, lock in!