Swayam’s Scripts

My take on Anthropic’s formal verification post

Where should the proof live: in the source or in the binary?

C++

Formal Verification

Anthropic, working with Theorem, just showed that real C and C++ code can come with a proof instead of just a test report. Using Claude, they proved that 239 OSS-Fuzz entry…

Floating Point Lab: the map of all 256 codes of an 8-bit float, coloured by class (zero, subnormal, normal, infinity, NaN), with 0x38 = 1 highlighted

Floating Point Lab

Interactive companions to Floating Point, Without the Lies

Floating-Point

Numerics

Interactive

Every value in these widgets is computed exactly, with rational arithmetic or from the checked format tables, and only rounded for display. Where a widget is about binary64…

Cpp-Verify: Formal deductive verification for C++, built into Clang

Formal verification for C++. Prove correctness instead of testing examples.

C++

Formal Verification

Published on my Substack, The Solution.

[WIP] Discrete Mathematics

Mathematics

ongoing

C++ Tips

This article summarizes some items I often keep forgetting

C++

Ongoing

constexpr to if statements must need to evaluated at compile time, else error while constexpr functions are allowed to be compile-time evaluation if function logic allows…

[WIP] Effective C++

This article summarizes all the items mentioned in Effective C++ by Scott Meyers

C++

Ongoing

C++ encapsulates multiple programming paradigms Object Oriented, Template Metaprogramming, Functional, Generic. It is more of a federation of languages.

CUDA Notes: Cooperative Groups

Beyond Block Barriers, Flexible Thread Synchronization from Warps to Grids

CUDA

A flexible model for synchronization and communication within groups of threads

[WIP] A Hitchhiker’s Guide to LLVM

Just Another Day in the Life of a SSA Variable

Compilers

LLVM

ongoing

This guide is written during the release of LLVM version 21.0.0

Gradient Flow and Variance Propogation Analysis of Dynamic Tanh Layer

A Mathematical Investigation into DyT’s Potential to Mitigate the Curse of Depth in Pre-LN Transformers

ML-Theory

Transformers

The “Curse of Depth” in Pre-LN transformers, identified by Sun et al. (2025), reveals that deeper layers often function as near-identity mappings, contributing minimally to…

[WIP] The Transparent Algorithm

A bridge between formal mathematical proofs and intuitive understanding, guiding readers through the theoretical foundations of machine learning with clarity and precision.

ML

Theory

Book

ongoing

Concurrent C++ A Practical Guide

A concise and practical guide to mastering C++ concurrency, covering threads, synchronization, parallelism, and debugging—all in one place.

C++

Book

Information Theory in Machine Learning: A Fun Approach

Teaching Neural Networks to Handle Their Trust Issues

ML

Information-Theory

Have you ever noticed how the best stories start with uncertainty? Like that moment when you’re watching a thriller and think you know who the culprit is, but you’re not…

Understanding Perplexity

A New Perspective on Model Uncertainty

LLM

Recently, I was reading the Chapter 5 (Pretraining) of the book “Build a Large Language Model (From Scratch)” by Sebastian Raschka. I stumbled upon an intriguing…

Numpy QuadDType: Quadruple Precision for Everyone

Quad Precision for All: Simplifying High-Accuracy Computing with numpy_quaddtype

NumPy

Open-Source

Self-Attention Mimicking Gradient Descent

NLP

Transformers

This section of paper Uncovering mesa-optimization algorithms in Transformers presents a theoretical construction where a linear self-attention layer in a Transformer…