Jay Singh

Convexity buys guarantees, but every deep loss landscape lives without them, and the saddle points, overparameterized valleys, and implicit bias of the optimizer quietly take over.

Probability supplies the semantics and information theory supplies the ledger, and every loss function in machine learning is a cross-entropy wearing a disguise.

The singular value decomposition hands matrices a clean theory of approximation for free, and that theory turns out to be an accident of two dimensions.

A model too large for one chip gets split along three axes, each technique answering a different bottleneck, until the binding constraint stops being memory and becomes failure itself.

A loss curve looks like smooth decay, but the real story of training lives in the plateaus, the sudden jumps, and the circuits quietly assembling out of sight.

The transformer replaced recurrence with direct attention, and every major variant since is just a different masking rule applied to the same skeleton.

Deep learning breaks the rules classical statistics says it should obey, and the explanation hides in the geometry of a loss surface too vast to picture.

Quantum mechanics is linear algebra dressed up in physics clothing; the physical content lives in a handful of postulates the math itself never delivers.