Kabir Murjani — Writing https://kabir.codes Technical essays on machine learning, optimization, combinatorial search, and intelligent systems by Kabir Murjani. en-us kabir.murjani@iimb.ac.in (Kabir Murjani) kabir.murjani@iimb.ac.in (Kabir Murjani) Fri, 05 Jun 2026 03:50:04 GMT https://kabir.codes/Of_H.jpg Kabir Murjani — Writing https://kabir.codes Playing Chess in 57 Megabytes https://kabir.codes/writing/chessnano https://kabir.codes/writing/chessnano Wed, 01 Apr 2026 00:00:00 GMT A transformer that sees only notation, compressed via polar-angle quantization into a file smaller than a photograph. Abstract: There is something philosophically uncomfortable about treating chess as a language modeling problem. Chess has geometry. It has material balance, pawn structure, king safety, tactical patterns that span a dozen moves. A language model has none of that built in. It operates on a flat sequence of tokens with no explicit notion of a board, a square, or a piece. And yet, if you train a transformer on the move sequences of strong players, it has no choice but to learn *something* about chess in order to predict those sequences well. The only interesting questions are how much, and whether that something is enough to make legal moves reliably.

Read the full article →

Playing Chess in 57 Megabytes ]]> kabir.murjani@iimb.ac.in (Kabir Murjani) Machine Learning Dreaming in Proofs https://kabir.codes/writing/dreaming https://kabir.codes/writing/dreaming Sun, 01 Mar 2026 00:00:00 GMT Novelty is not a better optimizer. It is a wild dreamer chained to a merciless judge. Abstract: Here is something I have come to believe after staring at search problems for too long: optimization cannot be creative. Not "is bad at being creative." *Cannot be.* It is the wrong kind of process.

Read the full article →

Dreaming in Proofs ]]> kabir.murjani@iimb.ac.in (Kabir Murjani) Machine Learning Walls, Not Whispers https://kabir.codes/writing/whispers https://kabir.codes/writing/whispers Sun, 01 Feb 2026 00:00:00 GMT A small case for communicating *geometry* instead of *rewards* in distributed search. Abstract: There is a quiet dishonesty in how most distributed reinforcement learning systems talk to each other about failure.

Read the full article →

Walls, Not Whispers ]]> kabir.murjani@iimb.ac.in (Kabir Murjani) Machine Learning Rational Amnesia https://kabir.codes/writing/amnesia https://kabir.codes/writing/amnesia Thu, 01 Jan 2026 00:00:00 GMT Stop decaying your gradients. Start amputating your graphs. Abstract: Here is a failure mode I keep seeing, and I think it is more fundamental than people treat it as.

Read the full article →

Rational Amnesia ]]> kabir.murjani@iimb.ac.in (Kabir Murjani) Machine Learning The Calculus of Laziness https://kabir.codes/writing/lazy https://kabir.codes/writing/lazy Mon, 01 Dec 2025 00:00:00 GMT Why procrastination is not a flaw, but an optimal compression algorithm. Abstract: Every intelligent computation in the universe, whether it is a biological neural network or a student at three in the morning, reduces to exactly two questions:

Read the full article →

The Calculus of Laziness ]]> kabir.murjani@iimb.ac.in (Kabir Murjani) Machine Learning Water Finds the Crack https://kabir.codes/writing/water https://kabir.codes/writing/water Fri, 05 Jun 2026 03:50:04 GMT Why your solver doesn't care about your game, and how to force it to play honest. Abstract: Last Tuesday I watched a reinforcement learning agent completely break a agent routing environment I had spent a week building.

Read the full article →

Water Finds the Crack ]]> kabir.murjani@iimb.ac.in (Kabir Murjani) Machine Learning