Research, stories, and insights on AI and Data Engineering.
Deep dives into neural networks, NLP, and AI systems
Pipelines, warehouses, GPU-accelerated query engines, and big-data systems
Illustrated novels that teach the modern AI stack as story, with audio narration in English and Hindi
With gratitude to every unsung hero across AI and Data — the engineers, researchers, and builders whose quiet work makes all of this real.
— P.S.
What is an LLM, intuitively?
Tokens & embeddings — text becomes math.
Vector spaces, similarity, and the moment they named the model.
Self-attention, and a stranger at the library.
Their toy model says a sentence neither of them was ready for.
Layers, residual streams — and the secret hours.
In May 2026 I shipped gpudb v0.1 — the first SQL execution engine targeting Apple Silicon GPUs, built as a DuckDB extension with a CUDA backend on Linux. Three releases later, the project crossed two lines at once. v0.3.0's streaming-aggregate rewrite reached parity with native DuckDB on end-to-end TPC-H queries — the worst cell improved roughly 100×, from 11.05 s to 0.109 s. And gpudb became an official DuckDB Community Extension: INSTALL gpudb FROM community now works in any DuckDB ≥ 1.5.5, signed, no flags. This is the full arc — what v0.1 proved, what v0.2 honestly lost, what v0.3 fixed, and why the next GPU frontier is joins.
Read more
Your vector index and your query encoder drifted apart months ago. Retrieval quality is quietly collapsing, and nothing in your observability stack noticed.
In May 2026 I shipped gpudb v0.1 — the first SQL execution engine targeting Apple Silicon GPUs, built as a DuckDB extension with a CUDA backend on Linux. Three releases later, the project crossed two lines at once. v0.3.0's streaming-aggregate rewrite reached parity with native DuckDB on end-to-end TPC-H queries — the worst cell improved roughly 100×, from 11.05 s to 0.109 s. And gpudb became an official DuckDB Community Extension: INSTALL gpudb FROM community now works in any DuckDB ≥ 1.5.5, signed, no flags. This is the full arc — what v0.1 proved, what v0.2 honestly lost, what v0.3 fixed, and why the next GPU frontier is joins.
Kimi K3 proved open-weight LLMs can reach the frontier — and proved they're far too big to run where users actually are. The next race isn't a bigger model; it's the architecture review that puts frontier AI on an ordinary laptop.
Speculative decoding promises faster LLM inference without touching model weights. The math holds up — but the production story is messier than the papers admit.
A four-bit model file tells you how many elements it has — and almost nothing else that matters. Not the scale-derivation rule, not the zero-point convention, not the packing order. In 2026 alone, six documented incidents across vLLM and SGLang turned those silent agreements into silently wrong model output. GRIT is my answer: a 64-byte descriptor and an O(1) boundary check for block-scaled reduced-precision arrays, with five zero-dependency implementations that agree bit-for-bit on 96/96 cross-language fingerprints — and a read-only scanner that found real convention ambiguity in checkpoints you can download today.
Mechanistic interpretability has produced real, reproducible findings. Here is an honest accounting of what those findings actually mean for engineers.
Exploring the frontiers of AI, data, and technology. New research and insights published regularly.
About the Author