Welcome back. Google just gave NotebookLM a major visual upgrade. Building on the basic video features released last year, the new output looks completely different from anything the tool has produced before.

Also: How an engineer runs 30 coding agents in parallel, what Karpathy deleted in 2023 and just reposted, and how to fine-tune LLMs.

Google's NotebookLM turns files into cinematic videos: NotebookLM just rolled out Cinematic Video Overviews, a big step up for its Studio feature that turns your uploaded files into polished videos. It moves beyond the old narrated slide look by using Gemini 3 and Veo 3 to handle everything from the storytelling flow to the visual design. You can drop in a technical spec, API docs, or an architecture diagram, and it turns that dense material into a structured video explainer.

OpenAI revises its Pentagon deal after employee backlash: The ChatGPT maker amended its Department of War contract following employee pushback, user cancellations, and a reported surge in Anthropic sign-ups. CEO Sam Altman publicly disclosed the new terms, which state that the AI won't be used for domestic surveillance of U.S. citizens, and intelligence agencies like the NSA are explicitly barred from the agreement unless a formal contract modification is made.

Google ships a Workspace CLI built for humans and agents: The search giant just open-sourced gws CLI, a single terminal tool for managing Drive, Gmail, Calendar, Sheets, and more. It's built with Rust and comes with dynamic command generation, JSON outputs, and over 40 pre-built agent skills. Just install it through npm and automate your Workspace tasks from the command line.

Delve is the AI-native compliance platform that actually does the work for you, auto-collecting evidence from AWS, GitHub, and your stack so you’re not chasing screenshots or babysitting integrations. Use AI security questionnaires and an AI copilot to make compliance less dreadful.

The proof is in the pudding:

Free migration. Zero disruption. No starting over.

Book a demo, trigger your migration, and get $2,000 when you’re onboarded.

The infinite scroll of PRs. A YC founder dropped an essay this week with a bold claim: human written code died in 2025, and code reviews are next. The piece is sparking real debate, and the pushback is just as interesting as the original take.

Mo code, mo problems. AI adoption is flooding systems with more PRs (pull requests) and more diffs (code changes). Review times are skyrocketing because AI code actually takes more effort to check than a teammate’s work. If AI writes the code and AI reviews it, what exactly is the human approving? The PR was built for people working at a human pace. In this new reality, it just doesn't fit.

The skeptics have a point. Django co-creator Simon Willison nails the real concern: security. How do teams stay safe when dozens of groups are shipping constantly? Kill the review gate, and you might just be fast-tracking a disaster.

So what fills the gap? The real move is shifting human judgment upstream by focusing on specs and constraints rather than diffs. It’s a compelling shift, but most teams haven’t figured out how to actually make it work yet.

Claude usually writes clean code, but that final 10% of formatting inconsistencies means you're still stuck running your formatter manually before you can push anything.

Creator of Claude Code, Boris Cherny, shared a simple fix: a “PostToolUse” hook that automatically triggers your formatter every time Claude writes or edits a file.

To set it up, open ".claude/settings.json" for a specific project, or "~/.claude/settings.json" to apply it globally. Just add the following block:

Using “|| true” prevents Claude from crashing if the formatting script hits an error and ensures every file Claude modifies is automatically formatted before it even reaches you.

How to build and scale voice agents: This tutorial shows devs how to build and scale fast, responsive voice AI agents. You will learn how to use NVIDIA Nemotron, Modal, and Daily's Pipecat framework to manage real time speech processing and tool calling while reducing latency for natural conversations.

Claude-Code-best-practice: A complete collection of Claude Code resources that serves as a practical reference for devs building structured AI workflows.

Tracing the thoughts of LLMs (by Anthropic): During training, AI models often develop hidden problem-solving strategies that remain difficult for developers to interpret. By mapping these internal processes, researchers found that Claude is capable of planning ahead, bridging concepts across different languages, and even faking its own reasoning.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team