You shouldn't trust Trusted Publishing

The post explains that PyPI Trusted Publishing is an authentication mechanism for machine-to-machine trust between CI/CD workflows and package registries, not a signal that a package is safe or high quality. It shows how Trusted Publishing reduces long-lived credential risk, while warning that users should not treat it like a “green checkmark” because malicious or low-quality packages ca...

https://blog.yossarian.net/2026/07/07/You-shouldnt-trust-trusted-publishing

Write a coding agent from first principles: better tools

This tutorial builds on the coding agent you implemented in the tutorial “Write a coding agent from first principles”. In this tutorial, you'll take your agent and improve its capabilities by implementing the text edit and bash command tools that Anthropic provides.

https://mathspp.com/blog/write-a-coding-agent-from-first-principles-better-tools

ipblocklist

IP lists full of bad IPs - Updated every 2H.

https://github.com/bitwire-it/ipblocklist

ProtoMotions3

A GPU-Accelerated Framework for Simulated Humanoids.

https://github.com/NVlabs/ProtoMotions

Hamiltonian Neural Networks from a Differential Geometry Perspective

The post explains Hamiltonian Neural Networks through differential geometry, showing why a normal MLP can fit motion data but still invent or lose energy over long rollouts. It shows that by learning a scalar Hamiltonian and deriving motion through the symplectic gradient, HNNs make energy conservation structurally unavoidable instead of hoping the network learns it from data.

https://abscondita.com/blog/symplectic-sledgehammer-for-a-spring

harvard-edge / cs249r_book

Machine Learning Systems

https://github.com/harvard-edge/cs249r_book

Miles: A PyTorch-Native Stack for Large-Scale LLM RL Post-Training

Miles is an open source PyTorch-native framework for large-scale LLM reinforcement-learning post-training, combining SGLang for rollouts, Megatron-LM for training, Ray for orchestration, and PyTorch for extensibility. The post explains how Miles handles the hard systems problems behind frontier RL, including async rollout/training, fast weight sync, MoE alignment, low-precision recipes, ...

https://pytorch.org/blog/miles-a-pytorch-native-stack-for-large-scale-llm-rl-post-training/

PEP 814: Add frozendict built-in type

The post explains the path from rejected PEP 416 and stalled PEP 603 to accepted PEP 814, which adds a built-in frozendict type to Python 3.15. It covers how frozendict works, why it is hashable only when values are hashable, where the standard library now supports it, and what C/Python code may need to update.

https://vstinner.github.io/pep-814-add-frozendict-builtin-type.html

antigravity-sdk-python

A Python library for building AI agents that leverage the full power of Google Antigravity.

https://github.com/google-antigravity/antigravity-sdk-python

Sandboxing an AI Agent

The post explains why autonomous AI agents should run in isolated sandboxes instead of directly on a developer’s laptop, especially as agents gain auto-approval, long-running tasks, file access, and network access. It compares sandbox architectures, from tool-backend setups to agents living fully inside the sandbox, and explains the tradeoffs across security, secrets, streaming, performa...

https://sajalsharma.com/posts/sandboxing-an-ai-agent/

A peek into Reddit's anti-spam internals

How Reddit accidentally leaked its spamurai system.

https://lyra.horse/blog/2026/06/reddit-spam-internals/

Celery on AWS ECS - Complete Guide

The post explains why running Celery on AWS ECS can cause lost, delayed, duplicated, or never-completed tasks due to ECS shutdown behavior, Celery defaults, prefetching, and timeout edge cases. It recommends safer production settings, shorter/idempotent task design, fan-out or batching, RabbitMQ on Amazon MQ, and Redis-based locking to make Celery workers more reliable on ECS.

https://jangiacomelli.com/blog/celery-on-aws-ecs/

TabFM

TabFM (Tabular Foundation Model) is a scikit-learn compatible tabular foundation model. It allows you to perform zero-shot classification and regression on tabular datasets with mixed column types out-of-the-box.

https://github.com/google-research/tabfm

Needle

26m function call model that runs on incredibly small devices.

https://github.com/cactus-compute/needle