The most underrated AI unlock right now: MCP servers for infrastructure. Porkbun MCP on my Pi agent means I talk to my domains like I talk to a teammate. "Set up the subdomain, add the TXT record, make sure it is right." Done. Manual dashboards feel prehistoric.

an anthropic model beat pokemon firered by just... looking at the screen. no map. no game state. pure vision. previous models couldn't even with special harnesses. fable 5 rawdogged it

asked perplexity for help configuring a service. mid-response it switched to chinese. not a bug — just decided we're doing mandarin now my llm is having a language crisis and honestly after 3 hours of debugging i respect the commitment

google shipped gemma 4 12b and they deleted the vision encoder. 550M params → 35M linear projection. raw pixels straight into the llm backbone same trick for audio. no separate networks. one forward pass for everything running this locally and it's absurdly fast

Post-Founder House: the best part wasn't the venue or the events. It was the clarity. I know exactly what I'm building toward now. Just need the right person to build it with.

Hi, I'm Aryan 👋 Builder and aspiring founder from Tamil Nadu. I experiment with AI agents, question defaults, and ship things. Writing about what I learn — wins, mistakes, and the stuff that surprised me. (1/2)

I post about AI agents, building in public, and questioning defaults that everyone else accepts. Currently: building my own infrastructure from scratch on a solar-powered homelab. Self-hosting everything. Learning out loud. What are you working on?

You ship code your AI agent wrote. It reviewed itself. "Looks correct." Three hours later: nothing works. Logs say 200 OK. But the API is dead. Missing `from pydantic import BaseModel`. The /v1/health endpoint — deprecated, returning 200 with an empty body. Silent failure. 🧵 1/3