Floating point from scratch

Floating point from scratch
41 minutes by Julia Desmazes

Julia explores how floating point numbers work by rebuilding them from the ground up, from mathematical definitions to hardware design. She explains complex features like NaNs, infinities, rounding modes, and subnormals in simple terms, then shows how to implement a custom bfloat16 floating point unit in hardware. She also describes the challenges of verification, compiler behavior, and chip tapeouts, highlighting how theory and real silicon often differ.

Voice Features That Sound Human and Start Fast
sponsored by Async

Real-time voice is hard to get right when latency and quality pull in opposite directions. Async Voice API gives developers streaming TTS, voice cloning, and 15+ languages, with audio starting in just 166 ms. It ranks Top-3 on Hugging Face TTS Arena, supports REST and WebSocket APIs, and works with Twilio, LiveKit, and Pipecat. From $0.50/hour

How I built a sub-500ms latency voice agent from scratch
14 minutes by Nick Tikhonov

Nick explains how he built a fast voice agent from scratch with about 400ms response time. He shows why voice systems are harder than chat, focusing on real-time turn-taking and coordination between speech recognition, language models, and speech output. By improving latency, model choice, and system design, the custom setup outperformed existing tools and made conversations feel more natural.

The peril of laziness lost
5 minutes by Bryan Cantrill

Good software needs laziness — the drive to build clean abstractions that make future work easier. LLMs lack this virtue entirely, since effort costs them nothing, so they pile on code without any pressure to simplify. This makes human constraints more important, not less. LLMs work best as tools guided by engineers who still care about building systems that are smaller and cleaner.

Cleaning up merged git branches: A one-liner from the CIA's leaked dev docs
2 minutes by Spencer Dixon

A 2017 WikiLeaks dump of CIA internal documents included some everyday developer tips, one of which is genuinely useful. A single command lists all merged git branches and deletes them at once, skipping your current branch and main. Saving it as an alias means cleaning up a cluttered branch list takes one word instead of several minutes of manual work.

Agent harness engineering
17 minutes by Addy Osmani

Harness engineering focuses on building the system around an AI model so it works reliably. An agent is the model plus this “harness” of prompts, tools, memory, and rules. Instead of blaming the model for mistakes, engineers fix the harness so errors don’t repeat. Strong harness design—like feedback loops and safeguards—often matters more than the model itself for real-world performance.

And the most popular article from the last issue was:

• IAM: Everything you need to know