Taste-Bench

Taste-Bench measures the taste of an LLM agent: its ability to choose the better direction at a real decision fork in a long-horizon task.

https://github.com/wbopan/tastebench

type_enforced

Fast runtime type enforcement for Python 3.11+ type annotations. Zero dependencies and uncompromising performance.

https://github.com/connor-makowski/type_enforced

Lemma

The open-source workspace where humans and AI agents work as one team.

https://github.com/lemma-work/lemma-platform

Making np.searchsorted up to 25× Faster in NumPy 2.5

The article speeds up binary search by batching independent searches with NumPy’s vectorized operations, then reformulates the algorithm to use only O(1) extra memory. After porting it to C++, the approach achieves up to a 25× speedup over NumPy 2.4’s implementation.

https://blog.scientific-python.org/numpy/searchsorted/

Building a Faster (Rust-Based) Python Driver for ScyllaDB

ScyllaDB is building a new Python driver on top of its Rust driver using PyO3, with most serialization, deserialization, request execution, and concurrency handled in Rust for better performance and maintainability. Benchmarks show especially large gains under concurrent workloads, while the project still needs production hardening, broader feature coverage, and a compatibility layer for...

https://www.scylladb.com/2026/09/28/building-a-faster-rust-based-python-driver/

AI Development on Windows: from PyTorch and llama.cpp to Windows ML

Microsoft is expanding local AI development on Windows with experimental GGUF support in Windows ML via llama.cpp, new native inference APIs, and an OpenAI-compatible endpoint for running models locally. The broader stack also adds native PyTorch and Triton support on Windows Arm64, giving developers a path to train and optimize models with familiar tools and deploy them through Windows ...

https://devblogs.microsoft.com/foundry-on-windows/build-on-winml-oct-7-26/

ldraw-nova

Agent tooling for generative LEGO models building, built with Astra and Opus 5.5, powered by Jev

https://github.com/anteloc/ldraw-nova

Orchestrating Disposable Browser

Glazer replaced race-prone parallel startup with readiness signals between Python and Docker containers, allowing components to start concurrently without assuming that running means ready. That design cut cold starts from 8–10 seconds to about 4.5 seconds, and caching the Tor data directory brought startup down to a consistent 1.9–2.1 seconds.

https://blog.glazer.ee/posts/building-glazer-disposable-browser/

The things you own end up owning you

NumPy is being ported toward Python’s Limited API so it could eventually ship abi3 wheels that work across multiple Python versions, reducing the large per-version wheel matrix and easing support for new Python releases. The same work is also pushing NumPy toward subinterpreter support by replacing static types and global state with heap types and per-module state, though multiarray rema...

https://labs.quansight.org/blog/limited_api_numpy

Build & Train a GLM-5.3-Flash Model From Scratch with Python

This tutorial demonstrates how to build and train a compact 25-million parameter multimodal language model from scratch using modern techniques like mixture-of-experts and sparse attention. It further explores the AI researcher's role by guiding viewers through pre-training, reinforcement learning experiments, and the critical importance of designing effective data-driven environments.

https://www.youtube.com/watch?v=-gfgQfw2g_E

rGPU

Keep Python on your laptop. Run PyTorch operations and hold tensors on a remote GPU, including from a Mac with no CUDA installation.

https://github.com/ymcrcat/rgpu

Pyronaut

Pyronaut is a high-performance polyglot application platform that brings Micronaut features like HTTP routing, dependency injection, configuration, validation, and Java library access to Python via GraalPy.

https://github.com/micronaut-projects/pyronaut

Can Jev save my inbox from the Democrats?

The author outlines an end-to-end ML project evaluating whether TypeSafe’s Jev, accessed through OpenRouter, and smaller local models can effectively filter political fundraising emails from a personal inbox. He ultimately builds a 23.4 MB quantized ONNX SetFit model using mdbr-leaf-mt and confidence thresholding, creating a fast, CPU-friendly local classifier while exposing the remainin...

https://huggingface.co/blog/stephen-solka/use-jev-to-delete-fundraising-emails

How we built an async-aware Python profiler

Learn how Datadog made Python profiling async-aware while reducing profiler overhead and preserving context across asyncio tasks.

https://www.datadoghq.com/blog/engineering/async-python-profiler/

8 Python 3.15 Features You Need to Know

Python 3.15 introduces practical enhancements such as lazy imports, a built-in immutable dictionary, and streamlined unpacking in comprehensions. These updates, alongside a new sampling profiler and default UTF-8 encoding, aim to improve developer productivity and application performance.

https://www.youtube.com/watch?v=bgeqL4Btou0

Jev for Python engineers

Vercel’s Python AI SDK adds support for Jev, a fast universal classifier that answers structured choice, scoring, and probability questions without task-specific training. The post tests Jev on classifying Python vs. English and generating Python ASTs, showing that it works well for narrow decisions but still struggles when pushed into general code generation.

https://vercel.com/blog/jev-for-python-engineers

openTPU

An open-source AI accelerator, developed by AI: RTL, ISA, simulator, compiler and profiler in one repo. Runs Qwen3, LFM2.5 and Qwen3.5 on a Kintex-7 PCIe card.

https://github.com/FeSens/openTPU