Blog - Musings by S Akash

The Best Agent Is the One That Survived Your Benchmark

Sep 15, 2026

StartupSelf-Improving AgentsEvolution

re-forge (The Adaply), the startup I am building, is after the best possible coding agent for a team: capture two to three weeks of Claude Code sessions, turn what went wrong into candidate changes, evolve the agent's configuration with an evolutionary search against a benchmark built from the team's own prompts, and keep only what survives. Plus the research behind it: our paper Adaptation Without Inheritance, joint work with Shine Gupta, submitted to NeurIPS 2026. App screenshots and the paper included.

The First Binary That Works Is Not the Best One

Aug 19, 2026

CompilersGPU ComputingConference

I took a poster to the 2026 EuroLLVM Developers' Meeting in Dublin as an LLVM Foundation travel grantee. LLVM's offloading runtime picks the first GPU binary that works; the poster measures why it should pick the best one instead (selection costs 3–6 ns against a 36 μs module load) and proposes metadata-ranked, profile-guided dispatch. Poster included.

Survival of the Fittest Agent

Aug 17, 2026

Paper PresentationSelf-Improving AgentsEvolution

Session 02 of the ML Reading Club NCR (venue partner: VirtualVakil). I presented the Darwin Gödel Machine, a coding agent that rewrites its own code and evolves from 20% to 50% on SWE-bench with zero human edits, plus its descendants: the Huxley GM (select by clade, not score) and the Red Queen GM (evolve the judge too). Slides included.

RL Didn't Teach Your Model to Think

Aug 2, 2026

Paper PresentationReinforcement LearningReasoning

I started the ML Reading Club NCR in Noida (venue partner: Lyric), and gave the first session's talk on the NeurIPS 2025 paper asking whether RL with verifiable rewards teaches models anything new. The pass@k crossover says no: RL sharpens the base model's distribution, it does not expand it. Slides included.

The Best KV Cache Eviction Is the One You Never Have to Do

Jun 30, 2026

ML SystemsAttentionInference

I do KV-cache eviction for a living. An OCR report out of Baidu, R-SWA, bounds the decoder to the last 128 generated tokens, keeps the cache constant, and gets higher accuracy for it. Here is why that quietly reframes the whole eviction problem for reference-grounded tasks.

Can C++26 Reflection Pick Your GPU at Runtime?

Jun 18, 2026

C++GPU ComputingCompilers

C++26 reflection collapses every program entity into one scalar handle, std::meta::info. I ask whether you can steal that trick to select a GPU backend at runtime, and stay honest about why the host/device gap does not vanish.

Attention Mechanisms That Actually Matter: From Multi-Head to PagedAttention

Mar 25, 2026

ML SystemsGPU Computing

A deep dive into four attention mechanisms that power modern LLMs, from the original transformer to the serving tricks that make inference feasible at scale.

How Shazam Identifies a Song in 10 Seconds (The Math Is Beautiful)

Mar 24, 2026

Signal ProcessingEssay

A 20 year old paper turned an intractable search problem into a hash table lookup. Fourier transforms, constellation maps, and the most elegant algorithm I have ever encountered.

Dyson Spheres and the Kardashev Scale: Engineering at the Scale of Stars

Mar 23, 2026

ScienceEssay

What it means to harness a star's full output, why Dyson Spheres are the defining structure of a Type II civilization, and where humanity sits on the Kardashev Scale right now.

NVIDIA GPU Architectures: Ampere to Blackwell, Through a Kernel Developer's Eyes

Mar 22, 2026

GPU ComputingCUDA

Four generations of NVIDIA GPUs, from the perspective of someone who actually writes the kernels that run on them. What the specs mean when you are profiling at 2am.

To Climb the Kardashev Scale, We Need AI That Thinks Like Terence Tao

Mar 21, 2026

EssayAITechnology

Why civilizational scale engineering challenges require not one brilliant mind but thousands working simultaneously, and why that means superhuman AI is not optional.

I Built gpucheck Because GPU Kernel Testing Shouldn't Feel Like Guesswork

Mar 20, 2026

Open SourceGPU Computing

The story behind gpucheck, a pytest plugin for GPU kernels. 511 test configs, 8 real bugs found in Triton and PyTorch, and why the GPU community deserves better testing tools.