The AI Friday Brief
Good Monday, NOLA. Today's vibe: the Claude Code story keeps reshaping what's possible—a developer just shipped a project in three months that took eight years of dreaming, and the industry keeps proving that agentic engineering is the real deal. Meanwhile, building with AI is getting genuinely faster, running big models locally is getting easier, and there's a whole ecosystem of tools making AI more practical by the day.
The Big Story: Agentic Engineering Works
Eight Years of Wanting, Three Months of Building with AI
Lalit Maganti's deep-dive is one of the best accounts we've seen of what AI-assisted development actually looks like at scale. They spent eight years thinking about building a SQL playground, then Claude Code shipped it in three months. The post walks through real prompts, failures, iteration cycles—not theory, but the actual workflow of working with an AI agent on a complex project. Discussion on HN.
Simon Willison / Hacker News
Syntaqlite: A SQL Playground Built in Three Months
The actual product from that story above. It's a web-based SQL environment where you can query live data, save queries, and share them. The fact that this exists at all—as a polished, usable product—is the point. This is what happens when you have a good prompt, a patient developer, and Claude Code doing the heavy lifting.
Hacker News
Running AI Locally: It's Getting Real
Google Dropped Gemma 4 Locally—With LM Studio and Claude Code
Google's Gemma 4 models (the E2B and E4B versions) are now small enough to run on your own machine. This post walks through setting them up with LM Studio (a GUI for local inference) and using Claude Code to prototype with them. The barrier to running serious models locally just dropped again. HN discussion.
Hacker News
Google AI Edge Gallery: Run Gemma on Your Mac (No API Keys)
Google's official app for running Gemma models directly on your machine. Terrible name, legitimately useful app. If you want to experiment with on-device AI without messing with the CLI, this is the path of least resistance.
Simon Willison
Real-Time Audio/Video AI on an M3 Pro (Parlor)
A working demo of real-time voice I/O running on a MacBook. You talk to it, it talks back, no cloud calls. Uses Gemma and E2B (a model format optimized for edge inference). This is the "AI on your laptop" future people have been talking about—it's actually here. HN thread.
Hacker News
Tools & Releases You Can Use Today
scan-for-secrets 0.3: Catch Leaked API Keys Before They Leak
A CLI tool that scans your codebase for accidentally committed API keys and secrets. New in 0.3: a
Simon Willison
--redact flag that finds matches, asks for confirmation, and auto-replaces them. Run this before pushing to GitHub if you've been sloppy with `.env` files.Gemma Gem: Run AI Models in Your Browser (No Backend)
Embed Gemma directly in a web browser. No API keys, no cloud calls, no backend. Everything runs client-side using WebAssembly. Wild for prototyping lightweight AI experiences or building demo apps that don't require a server.
Hacker News
datasette-ports: Run Multiple Local Services from One CLI
A small utility for spinning up multiple local services on different ports. Useful if you're prototyping something complex locally and need Redis, a database, and a web server all talking to each other. No longer requires Datasette—works standalone.
Simon Willison
Learning & Understanding AI
Building a Tiny LLM to Understand How Language Models Actually Work
A minimal, readable implementation of a language model from scratch. If you've ever wondered what's actually happening inside Claude or GPT, this is a great way to build intuition without wading through 10,000 lines of production code. HN discussion.
Hacker News
How Coding Agents Actually Work (And What It Means for Your Team)
A breakdown of what goes into a system like Claude Code: prompt engineering, function calling, context management, error handling. This is genuinely useful if you're trying to understand where the magic ends and the engineering begins.
Sebastian Raschka's Magazine
Anthropic Research: Emotions as Function Vectors in Claude
Anthropic published research showing that Claude has identifiable internal representations of emotional concepts (like 'sadness' or 'joy'). It's not that Claude has feelings—it's that these abstract concepts have consistent mathematical structure inside the model. Interesting for understanding how LLMs actually work under the hood.
Anthropic Research
Industry & Culture
Microsoft's Copilot Is "For Entertainment Purposes Only" (Actually)
Microsoft's terms of service literally say Copilot is for entertainment. The reminder: AI companies themselves are telling you not to blindly trust their outputs. This isn't AI skepticism—it's coming straight from the vendor.
TechCrunch
The Musician vs. The AI Company: Copyright Trolling in Real Time
A folk musician's voice got cloned by an AI company. Then someone filed a copyright claim against *her* using the AI clone. This is the chaos frontier of AI-generated content: it's not just 'will AI replace creators,' it's 'who gets blamed when AI fakes your work?'
The Verge
Also
- Show HN: Nanocode—Training an LLM on TPUs with JAX — For folks interested in model training infrastructure—a minimal, efficient setup.
- Qwen-3.6-Plus: First Model to Hit 1 Trillion Tokens/Day — Interesting milestone in model scale and demand—shows which models are actually getting used at scale.
- OpenAI's Stargate Datacenter: Real or Hype? — Iranian IRGC released satellite imagery of what they claim is OpenAI's $30B datacenter. Worth the skeptical read.
- Why You Can't Trust AI Outputs (And Why That's OK) — Research on 'cognitive surrender'—why people trust AI outputs even when they're obviously wrong. Self-awareness matters.
- Andrej Karpathy's 'Idea File': How to Think About LLM Design — A thoughtful take on what matters when designing with LLMs. Even if you don't build models, the thinking is useful.
- GitHub Commits Surging: 14 Billion Projected for 2026 — 1 billion commits happened in 2025. We're on pace for 14 billion in 2026. AI-assisted coding is reshaping what 'productivity' means.
- Writing Lisp Is AI Resistant (And That's Weird) — An interesting observation: AI struggles with Lisp-style code. Why? The terseness and syntactic flexibility confuse models trained on Python/JavaScript.
- OneUptime Committed 12k AI-Generated Blog Posts at Once — A wild example of AI-generated content at scale—12,000 posts in a single commit. Questions about signal vs. noise in AI content remain.
- Proof of Human Authorship: Why We Need It Now — As AI content becomes ubiquitous, 'made by humans' becomes a competitive advantage. Weird timeline.
- OpenAI Cap Table Leak: Microsoft's 18x Return, SoftBank's $50B Gain — Financial numbers on OpenAI's cap table leaked. Microsoft's doing very well. Sam Altman owns almost nothing despite running the company.
Pass it on