"2025: where a Turing Award winner leaving Big Tech to start another AI company is just a regular Tuesday."
Yann LeCun is departing Meta after 10+ years to launch a startup focused on world models, while ElevenLabs just dropped a voice marketplace letting brands license AI celebrity voices. Plus, we're diving into whether you actually need GraphRAG or if it's just hype.
Hello Builders! 👋
🔥 Today's Top Story
Yann LeCun is leaving Meta to build "world models", the godfather of deep learning is ditching his cushy gig to chase what he's been preaching for years: AI that actually understands how the world works, not just pattern-matching text.
His "world models" approach means AI that can plan, reason about physics, and predict consequences—think agents that don't just write code but understand what it'll actually do when deployed.
The HN thread hit 413 comments because everyone knows: when a Turing Award winner leaves FAANG money to start fresh, they're seeing something the rest of us aren't yet.
🚀 Ships & Launches
ElevenLabs Drops Video Generation - The voice AI company now does video with Veo 3.1, turning images into full clips in their studio.
ElevenLabs Launches Celebrity Voice Marketplace - Brands can now license AI-cloned famous voices for ads, which definitely won't create any weird legal precedents.
Google Maps Gets AI Tools - New AI-powered features for Maps Platform developers to build location-aware apps with trusted data.
Google Brings Free Gemini to Africa - Partnership with Cassava Technologies makes Gemini accessible at zero data cost across the continent.
📺 Learn & Build
Measure RAG Retrieval Like a Pro - Final part on DCG@k and NDCG@k metrics so you actually know if your retrieval sucks or not.
GraphRAG: Skip the Hype - Real talk on when GraphRAG makes sense versus when you're just burning money on complexity you don't need.
Build a Gmail AI Agent - Create an n8n workflow that scores and routes your emails by priority so you can ignore the noise.
Your Everything Notebook Setup - How to use NotebookLM as a searchable brain dump for your entire professional life and actually find stuff later.
Perfect Code Zero Revenue - Meysam spent 2 months building FindForce with perfect code, made $0 because he feared competitors instead of seeing them as validation.
Free Claude Context Engineering - André released free .context files on Github that supposedly match paid builder results using Claude and context engineering.
💬 Builder Conversations
Composer 1 Over Claude - Bobby ditched Claude and OpenAI entirely for Cursor's Composer 1, claiming it's faster and more accurate for coding.
Vibe Coding Means Rebuilding - Luca argues you'll rebuild the same feature 5 times when vibe coding, and that's learning, not waste.
Ask Don't Command AI - Jaequery gets better results by phrasing prompts as questions rather than direct commands when vibe-coding.
YouTube Breaks Yt-dlp Again - The popular video downloader now requires external JavaScript runtime for full YouTube support, sparking 245 comments on workarounds.
📰 Industry Moves
Anthropic's $50B Infrastructure Bet - Building data centers in Texas and New York with UK partner Fluidstack, going full vertical integration mode.
World Labs Drops Marble - Fei-Fei Li's startup launches persistent, downloadable 3D world generation that actually sticks around instead of vanishing like competitor demos.
Google Clones Apple's Private Cloud - Rolling out their own version of Private Cloud Compute for on-device AI that keeps your data local-ish.
Europe Still in AI App Race - Accel's 2025 report shows Europe lagging in model funding but competitive in the application layer where most money gets made anyway.
Immortality Startup Gets Real - Eternos pivots to Uare.ai with $10.3M seed for personal AI clones, because nothing says pivot like dropping the death-cheating brand.
Figma Plants Flag in Bengaluru - New office signals push beyond design tools into broader developer territory with local engineering talent.
Microsoft Backs Superconductor Data Center Tech - VEIR pivots from power transmission to solving data center energy problems with megawatt-class superconductors.
Open-Source OCR Catches Up - DeepSeek hit 84% accuracy on messy documents, beating most commercial OCRs that cost $500/month for worse results.
⚡ Quick Links
The 5 FREE Must-Read Books for Every AI Engineer - Free resources to level up skills
How AI startups should be thinking about product-market fit - PMF strategies for AI companies
One More Thing...
Wild to see LeCun betting on world models while everyone's arguing about whether Composer 1 is fast enough to ditch Claude entirely. Feels like we're watching two different timelines of AI development play out in real-time.
What are you shipping this week? Testing any of these new tools?
Keep shipping,
P.S. If you made a cat mockumentary with ElevenLabs video today, you're legally required to share it. I don't make the rules.
