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.
Read moreI 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.
Read moreC++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.
Read moreA deep dive into four attention mechanisms that power modern LLMs, from the original transformer to the serving tricks that make inference feasible at scale.
Read moreA 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.
Read moreWhat 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.
Read moreFour 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.
Read moreWhy civilizational scale engineering challenges require not one brilliant mind but thousands working simultaneously, and why that means superhuman AI is not optional.
Read moreThe 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.
Read more