So I've been focusing a lot on local AI and local inference, as they like to call it in the industry. Sam Altman was recently on an interview where he was chastising people for wanting to have their own local inference and how dare we sit on memory that's not being used 100% of the time. Because apparently that's the goal. What a tool. Anyway, I'm going to start posting more back in this community. Take a look at the classroom if you're interested in figuring out how to vibe code an app from all the way from the very beginning in Basics into launch and production and to do it securely. And then keep an eye out for more posts on local inference, software development, and where AI is headed. Okay.

Using open design and a local model you can do a lot

Andrej Karpathy spoke at AI Ascent this past week. He described something he called agentic engineering. The human stays in the loop. The agent writes detailed documentation. You oversee the architecture. The agent handles the code specifics. I've been teaching exactly that for almost a year. Here's what I built it around: Before any code gets written, you run four phases. Brainstorming. Product requirements. Technical architecture. Then implementation — story by story, with the AI reading your planning documents before it builds anything. Why? Because without a plan, every feature is built in isolation. The AI has no memory of decisions made in previous sessions. Each prompt starts from scratch. The documentation is the memory. The human is the director. Karpathy's framing yesterday was exact: vibecoding raises the floor. Agentic engineering maintains the ceiling — the professional standards, the security model, the architectural coherence that vibe coding can't talk about. He said you can outsource your thinking. You cannot outsource your understanding. That's the line I've been trying to write for a year. I built BMAD and the Vibe Code Blueprint around that principle before it had a name. Now it does. Agentic engineering. The methodology was always the missing piece. And now you can learn it too.

It started with a few failed prompts and ended with a complete lockout. If you’ve been hitting Internal Server Error 500 or getting bounced from the login screen this morning, you aren't alone. As of April 15, 2026, Anthropic is officially grappling with a major outage affecting Claude.ai, the API, and the Claude Code CLI. For those of us deep in the world of "vibe coding," where the flow depends on a tight feedback loop between our natural language and the machine, these service interruptions are more than just a nuisance: they are a complete work stoppage. What’s Happening? - Widespread Login Failures: Users are being logged out and unable to return to their sessions. - The "500" Wall: Claude Code and API requests are dropping mid-stream, returning "Internal Server Error" instead of that sweet, functional code. - Systemic Instability: This follows a week of intermittent degraded performance, leading many to wonder if the infrastructure is struggling to keep up with the latest Sonnet and Opus 4.6 deployments. The Home Lab Advantage If there was ever a day to celebrate data sovereignty, today is it. While the cloud-reliant masses are stuck staring at status pages, this is where a robust home lab pays for itself. 1. Failover to Ollama: By pointing your development agents to local Ollama endpoints, you keep your logic in-house and your throughput steady. 2. Modular Resilience: The best "vibe coding" workflows aren't tied to a single model. Use this downtime to test your current PRDs against local LLMs like Llama 3 or DeepSeek. If your prompts are truly modular, they should perform regardless of the backend. 3. Triple-Pass Validation (TPV): Even when the API returns, use the TPV protocol to ensure the "post-outage" code hasn't suffered from the "lazy output" issues that often plague models when servers are under extreme load. Staying Operational Check the official status page for updates, but don't wait for a green light to stay productive. Shift your builds to your local hardware, keep your Docker containers humming, and remember: the best AI infrastructure is the one you control.

Has anyone been having issues with Claude lately? I have wanted to support Dario through their whole GOV issue but the way they have been lashing out at the community has got me thinking about trying Codex. Whats your thoughts? Also now with the Claude Code leak so may features would be going into the other CLI tools. Whats your thoughts on this?