Tiny Models That Overdeliver, Post-Training Tooling Goes Stable, and Edge AI Gets Crowded

Tools

Liquid AI's LFM2.5-350M Outperforms Models Twice Its Size on Tool Use and Instruction Following

Trained on 28T tokens with large-scale RL, this 350M model beats Gemma 3 1B and Qwen3.5-0.8B on tool use and data extraction benchmarks. If you're building edge pipelines that need fast, cheap function calling, this is worth dropping into your stack today.

TRL v1.0 Ships With 75+ Post-Training Methods and a Stability Guarantee

TRL is now a production-grade library, not just a research repo, covering PPO, DPO, GRPO, and everything in between. The v1.0 commitment to API stability means you can actually build on it without chasing breaking changes every sprint.

Analysis

Claude Code Source Leak Exposes Anti-Distillation Traps, Frustration Detectors, and Stealth Mode

Anthropic accidentally shipped readable source maps in the Claude Code npm package, revealing fake tool injection to poison training data scrapers and server-side summarization that strips reasoning chains from API logs. Whether you build agents or compete with Anthropic, these design decisions are directly relevant to how you think about protecting and structuring your own pipelines.

Claude Code Architecture Deep Dive: Five Layers, AsyncGenerator Loop, and a 100+ Tool Runtime

This is the most thorough technical breakdown of the leaked Claude Code internals, covering the QueryEngine singleton, streaming via async generators, and a permission model that gates every tool call through human approval. If you're designing an agent framework, this is a free look at how a production system at scale actually handles state, interruption, and multi-agent coordination.