Building a GPU Compiler from Scratch

Progressive multi-level IR lowering systematically transforms high-level PyTorch graphs into optimized CUDA code through six explicit compiler stages. By separating graph fusion, hardware scheduling, and autotuning, the architecture makes GPU optimization easier to reason about, measure, and improve.

https://kernelspace.substack.com/p/building-a-gpu-compiler-from-scratch

google / artemis

ARTEMIS turns natural-language instructions into reliable Android automation. It automates end-to-end workflows, captures logs, and integrates seamlessly with AI coding assistants such as Antigravity, Codex, and Claude Code. It also achieves 99%+ success rate on AndroidWorld Benchmark.

https://github.com/google/artemis

Polars Query Optimization: Lazy vs. Eager Mode

Learn how Polars' lazy mode improves memory efficiency and performance compared to eager mode by building an optimized query plan before execution. By utilizing predicate and projection pushdown, lazy mode significantly reduces resource usage when working with large datasets, making it the preferred approach for most data analysis tasks.

https://www.youtube.com/watch?v=ux7MlXxFDQY

Automating Django deployments with Deployer

The article demonstrates how to implement a multi-lift scheduling algorithm in Python using LOOK to serve requests in the current direction before reversing. It then explores multi-lift optimization by separating responsibilities between a Dispatcher and a LiftController state machine, with requests assigned using weighted cost functions.

https://codewithstein.com/automating-django-deployments-with-deployer/

Stop one person claiming your Python free tier twice

Fold Gmail dots, plus tags and googlemail.com into one key, then grant the trial in one transaction. Django and SQLAlchemy.

https://elijahbrown.hashnode.dev/stop-one-person-claiming-your-python-free-tier-twice-gmail-dots-plus-tags-and-googlemail-com

whatsnewt

Python core developer Barry Warsaw text adventure game to test your knowledge of Python 3.15 by solving puzzles.

https://pypi.org/project/whatsnewt/

Jumper

Jumper is an open-source Python toolkit for a 22-DoF crab robot that lets AI coding agents design appearances and scenes, train motions with reinforcement learning, and package them for deployment to the real robot.

https://github.com/KingKongRobotics/jumper

Python 3.14 vs Python 3.15 – performance testing

Performance benchmarks on AMD and Intel systems reveal that Python 3.15 is overall ~1.45× faster than Python 3.14. While massive speedups in Base32, Ascii85, and Base85 C-rewrite encoding operations heavily skew the average, most other workloads show a more modest performance improvement of around 1.20×.

https://en.lewoniewski.info/2026/python-3-14-vs-python-3-15-performance-testing/

AirLLM

AirLLM is a Python library for running very large language models on consumer GPUs with limited memory by loading model layers sequentially instead of keeping the entire model in VRAM.

https://github.com/lyogavin/airllm

A Clean Way to Implement Undo in Python

This video explains how to implement the Memento design pattern in Python to handle state restoration and undo functionality without breaking encapsulation. It demonstrates a clean, modern approach using frozen data classes, protocols, and generics to manage state snapshots independently of application logic.

https://www.youtube.com/watch?v=ZXAt8AP3DnY

Parsing JSON at tens of gigabytes per second… in Python

Free-threaded Python 3.14 can push JSON parsing into the tens of gigabytes per second when fastsimdjson spreads work across many cores and avoids eagerly materializing Python objects. On a 64-core machine, Lemire reports throughput around 66 GB/s, showing how dramatically multithreaded parsing can scale when Python’s GIL is no longer the limiting factor.

https://lemire.me/blog/2026/10/12/parsing-json-at-tens-of-gigabytes-per-second-in-python/

why .parquet dominates modern data engineering

This video explains why the Parquet file format has become a dominant standard in modern data engineering by using columnar storage to improve speed and reduce file size compared to traditional CSV formats. It details key technical features like compression, encoding, and predicate pushdown that allow for significantly more efficient data retrieval and storage for large-scale analysis.

https://www.youtube.com/watch?v=6Qkzfwk0W58

fastapi-stream-lease

Distributed stream and SSE concurrency lease manager for FastAPI and Starlette backed by atomic Redis Lua scripts.

https://github.com/agustin18/fastapi-stream-lease

LongLive

Long video generation and world-action modeling research from NVIDIA. Each project has its own directory, code, documentation and model weights.

https://github.com/NVlabs/LongLive

CacheCanary

CacheCanary is a Python CLI and GitHub Action that detects why Claude prompt caching silently stops working on Amazon Bedrock, checks saved requests in CI, and analyzes Bedrock logs to show cache hit rates and the cost of misses.

https://github.com/Haarris/cachecanary

Lifeguard

Lifeguard is a static analyzer to detect Lazy Imports incompatibilities and ease the adoption overhead for Lazy Imports in Python.

https://github.com/Facebook/lifeguard

python-date-parser

A Rust-backed Python library for fast fuzzy parsing of heterogeneous date strings, with bulk-list and Polars APIs that avoid Python-level iteration and can outperform pure-Python alternatives by orders of magnitude.

https://github.com/SEIU-Tech/python-date-parser

Building a small Django app with help from AI

This tutorial demonstrates how to build a basic Django idea-voting application using AI as an interactive assistant rather than an autonomous developer. It emphasizes using targeted prompts to generate boilerplate, explain error messages, write tests, and catch edge cases like database race conditions while keeping architectural decisions in the developer’s hands.

https://codewithstein.com/building-a-small-django-app-with-help-from-ai/

Speculative Decoding: The Art of Guessing Well

Speculative decoding lets smaller models guess several tokens ahead while the larger model verifies them in parallel, reducing LLM inference latency without changing the final sampling distribution. From the acceptance math to Gumbel coupling, EAGLE-3, and multi-token prediction, the piece shows how increasingly sophisticated drafters make those guesses accurate enough to deliver real sp...

https://astralord.github.io/posts/speculative-decoding-the-art-of-guessing-well/

loopinfo: let your for loop know if it is on the first or last item

Python loops do not know where they are. loopinfo adds loop.first, loop.last, loop.index, loop.count, previous, next and more.

https://pypi.org/project/loopinfo/