Thinking out loud
Blog
Notes on AI, GPUs and engineering, written to check whether I really understood something.
If I can’texplain it,I don’tget it yet.
Latest entry · Oct 2026
Stop Giving Every Model Its Own GPU: Serving Embeddings, Rerankers, OCR and LLMs on One Card
Karpathy recently made the same bet at the level of the model itself, arguing that a “cognitive core” of roughly a billion parameters, distilled from a larger model, is…
Read on MediumOct 2026
Stop Recomputing Tokens: KV, Prefix, Prompt, and Semantic Caching Breakdown
Everything you need to understand where your input tokens are being recomputed and what to do about it. It covers the four cache layers from first principles, their…
Oct 2026
Keep the GPU Busy: A Hardware-First Guide to CUDA Kernels
The best way to understand and go about writing CUDA kernels would be to understand the hardware, look at how the compiler instructions work and connect them both.…
Oct 2026
Scaling LLM Reasoning with Amazon SageMaker RLVR
Reinforcement Learning from Verifiable Rewards (RLVR) is useful when an answer can be checked automatically, but writing an ideal response is expensive.
Oct 2026
Stop Waiting on Local LLMs: How Uzu Supercharges Inference Speed
In 2023, Karpathy explained why local LLMs are slow. When a model generates one stream of text on your own computer, the chip spends most of its time waiting for weights…
Oct 2026
Everything an AI Engineer Needs to Know About GPUs
What is inside a GPU, how training and inference use it differently, the generations, what you actually rent, NVIDIA’s competitors, and running models locally.
Oct 2026
CUDA from zero to hero #1
These are my notes as I learn CUDA; here I will try to break down things and represent them in a way that makes the most sense to me. I like to truly understand things…
Sep 2026
Jev without Jargons
We have been using LLMs like a hammer for every AI problem, even simple decisions. Jev handles those decisions in milliseconds at a fraction of the cost. Let’s…
Sep 2026
Winning the Modern AI Search Landscape
For decades, the fundamental search was relatively straightforward: crawl the web, index documents, retrieve relevant pages, rank them, and present the results to the…
Aug 2026
Maps, Keys, and Digital Fingerprints: Demystifying Linux Security and Safe App Installation…
Every time you install modern, secure software outside the default Ubuntu/Debian store, you follow the exact same 3-step pattern: