Local AI Models Just Killed the Cloud Era

"The year we realized running a 70B model on our laptop was easier than explaining our AWS bill to the CFO."

Hello Builders 👋

This is our last one for the year. Hoping this daily newsletter has been as valuable to you, it has been to me. I’ll be working though. My AI Resume Builder needs marketing support to be ready for the new years.

Here is Christmas card for you. I created using Wishbot, my first vibe coded app.

Anthropic quietly dropped LSP integration into Claude Code, and the Hacker News thread exploded to 268 comments. This is huge because it means Claude can now tap directly into your language server, the same thing that powers your IDE's autocomplete, go-to-definition, and error checking.

Translation: Claude stops hallucinating about your codebase. It knows your actual types, function signatures, and project structure without you copy-pasting context like a caveman. This is the difference between an AI that "kinda gets it" and one that actually understands your code.

Replit Moves Into ChatGPT - Connect your Replit account and build actual working apps directly from ChatGPT conversations, no tab switching required.

v0 Meets Enterprise Knowledge - Vercel's v0 now connects to Glean workspaces, letting you generate UIs based on your company's internal docs and data.

AI Engineering Trends for 2025 - The New Stack covers agents, MCP, and vibe coding trends actually worth paying attention to this year.

Vibe Coding for Marketing Workflows - How search marketers are using AI coding tools to automate tedious SEO and analytics work.

Code Mode vs MCP Explained - Rizèl Scarlett breaks down what Code Mode actually does and why it's not killing MCP anytime soon.

Symfony's AI Development Assistant - MCP server specifically built for Symfony projects, because PHP devs deserve AI tooling too.

Gemini 3 Flash A/B Tests Code - Google demos real-time code generation A/B testing with Gemini 3 Flash evolving loading spinners based on user feedback, which is either the future or just overengineered spinners.

Vibe Coding Is Promotion - Federico argues spending 90% of time reviewing AI code versus writing it isn't less work—it's an architect role.

15-Year-Old Ships 50k Users - WSJ profiled a high schooler who built an AI financial research platform where Claude wrote almost all the code.

AI Fooled Her Own Readers - Journalist fine-tuned a model on her writing and seven regular readers couldn't consistently tell what she wrote versus the AI.

Vibe Coding's 2026 Ceiling - Builder predicts platforms like Lovable and Bolt will hit 500M users before year-end, questioning if we've peaked or just started.

ChatGPT Does Spotify Wrapped - OpenAI's year-end review shows your chat stats with awards, poems, and pictures because apparently we needed another thing to screenshot and share.

Prompt Injection: Forever Problem - OpenAI admits AI browsers like Atlas will always be vulnerable to prompt injection attacks, so they're fighting fire with fire using an "LLM-based automated attacker" for defense.

Small Business Gets AI Help - Gemini 3 Flash can analyze customer feedback and draft launch emails for small businesses, turning hours of work into minutes if you trust it enough.

Apple Wallet Pass Generator - Turn any barcode into an Apple Wallet pass because your local coffee shop still lives in 2015.

HN Wrapped 2025 - LLM roasts your Hacker News activity from the past year with generated stats and commentary.

Wild that we're getting year-end wrapped experiences for our AI chats now. Between Claude's LSP support and that stat about spending 90% of time reviewing AI-generated code instead of writing it, feels like we've crossed some kind of Rubicon this year.

Keep shipping,

P.S. If your ChatGPT Wrapped shows more debugging conversations than actual feature builds, you're doing it right. That's just called Tuesday.