🧠 Meta Just Entered the Coding Agent Wars

Mark Zuckerberg announced Muse Code in beta, a terminal-based coding agent that takes on full software engineering tasks across large codebases, planning the changes, writing the code, and validating the results. It runs on Muse Spark 1.2, a coding-focused update to Meta's model family, and it's Meta's first dedicated entry into the agentic coding space that Claude Code and Codex have spent the past year building out. You can follow the updates on Meta’s blog website.

On benchmarks, Muse Spark 1.2 lands just behind Anthropic's Opus 5 in Claude Code and OpenAI's GPT-5.6 Terra in Codex, but undercuts both on price through a steeply discounted contributor tier. That discount comes with a catch: you get it in exchange for letting Meta train on your prompts and completions.

Muse Code is live in beta for macOS and Linux now.

Demis Hassabis is stepping down as CEO of Google DeepMind to become the unit's chairman and take on the title of chief scientist for Alphabet, a move meant to free him up for AGI strategy and Isomorphic Labs, Alphabet's AI drug discovery spinoff. Koray Kavukcuoglu, previously DeepMind's CTO, takes over daily operations as senior VP, reporting directly to Sundar Pichai.

The bigger surprise is who's leaving. Jeff Dean, a 27-year Google veteran, is exiting to start a new company called Discovery Loop. This lands as Google faces real pressure to keep pace with OpenAI and Anthropic, and the market noticed immediately, with Alphabet's stock dropping more than 4% on the news.

Here’s the direct note from CEO Sundar Pichai and the man himself.

Anthropic confirmed it's assembling an in-house "custom silicon team" to design its own chips for Claude, aiming to co-design hardware and models together so Claude runs faster and more efficiently at scale.

Anthropic isn't first here. OpenAI unveiled its own Broadcom-built inference chip back in June, Google has run on its own TPUs for years, and Meta has been developing its MTIA accelerators. Anthropic says this sits alongside its existing multi-chip setup with AWS, Google, Nvidia, and AMD, not instead of it.

Designing an advanced AI chip from scratch typically costs close to half a billion dollars once you factor in engineering talent and manufacturing, so this is a long game. No timeline has been announced.

If you're paying for Airtable and want your data somewhere you actually own, pull your entire workspace into Obsidian as local Markdown files, views, filters, and formulas included, using the official Importer plugin. Each Airtable table turns into a folder of notes plus a Base file that recreates your original views, so nothing about how you had things organized gets lost.

It's also a good entry point into using Obsidian as a second brain: once your data lives as Markdown, you can paste research, meeting notes, or outputs from Claude or ChatGPT straight into the same vault, so everything stays searchable and linked in one place instead of scattered across apps.

As AI video models keep getting better, one of the most underrated use cases isn't short clips, it's building genuinely creative websites around them. Pear.no is a live example worth studying: an agency site that regenerates its own visuals with AI on every refresh, using GPT Image 2 and Seedance 2.0 to create new imagery and motion each time someone loads the page.

Worth a look if you want a sense of where AI-generated web design is actually headed, and a nudge to stop thinking of these tools as "video generators" the next time you're building something.

A developer created a super cool AI video making the classic two-hand "finger frame" gesture and turns whatever's inside that frame into a fully AI-animated scene, while the rest of the video, your hands, your face, your movement, stays completely real.

Here’s the code or you can try it out here!

A few weeks ago we started asking one question in the RSVP form: what's the task you'd love to hand to AI but haven't figured out yet? Twenty-five of you wrote back.

This Friday, we’re building the first one live, from scratch, submitted by one of you:

"Take a video I like on Instagram, transcribe it, and have an agent rewrite it in my language patterns so I can recreate it as my own video or article."

We’ll build it live, start to finish — how we think through it, where we get stuck, and what it looks like running. How to use skills, connectors and Claude Code. At the end, one number on screen: the hours a month this task used to cost, and what it costs now.

Your submission could be the next build. If we pick yours, we build it whether you're there live or not, and you get the finished system either way. Everyone can submit.

Stuck mid-build on something already? The last 15–20 minutes of the session are for exactly that: you started automating something, hit a wall, bring the wall. We’ll get you unstuck live. No question too basic; being stuck at step two is the normal part nobody posts about.

FYI: Practicaly AI Labs members jump the queue →

🗓️ When: Friday, August 7 2026 @ 10AM PT (1PM ET)
 📹️ Where: Virtual. (Can't make it live? RSVP anyway, every registrant gets the recording.)