Granda | Ideas & Code

Patching LLM Weights by Hand

Wed, 15 Apr 2026 00:00:00 +0000

I wrote a from-scratch implementation of ROME — rank-one model editing — using nothing but torch and transformers. The goal: rewrite a single fact inside GPT-2 Medium with one matrix addition, and see what happens to everything else the model knows.

I ran four edits. Three were surgical. The fourth exposed that ROME’s “surgicalness” is extremely sensitive to implementation details you don’t necessarily think about up front.

The Four Edits

I picked facts GPT-2 Medium actually knows with confidence (P > 0.5 on the correct answer) across four different domains:

Analytics

Fri, 03 Apr 2026 00:00:00 +0000

Building a Spacecraft Computing Simulator with Claude Code

Mon, 30 Mar 2026 00:00:00 +0000

I wanted to understand how spacecraft computers work. The real-time scheduling, the radiation hardening, the delay-tolerant networking that keeps Mars rovers talking to Earth. With Artemis II on the horizon (the first crewed lunar mission in over 50 years) it felt like the right time to dig in. So I built a simulator. From scratch. With Claude Code as my pair programmer.

No physical hardware. Everything runs in QEMU and Docker on a laptop. Code on GitHub.

Your Parked Tesla Is a Data Center

Fri, 13 Mar 2026 00:00:00 +0000

Your car is parked 95% of the time. Inside it sits a chip capable of 300-500 trillion operations per second, connected to cooling, power conversion, and a cellular radio. It does nothing.

Tesla and xAI want to change that. On March 11, Elon Musk unveiled “Macrohard” — internally called Digital Optimus — a joint project that turns parked Teslas into personal AI agents. Not chatbots. Agents that watch your screen, control your mouse and keyboard, and do actual work.

Visual QA as a CI Pipeline Stage

Fri, 06 Feb 2026 00:00:00 +0000

I merged a PR last month. The code review looked good. Tests passed. Then I opened the site on my phone and the sidebar was completely broken.

The fix was trivial—a missing media query. The bug was obvious once you actually looked at the mobile view. Nobody did.

So I added a pipeline stage that looks.

I open a GitHub issue that says:

Implement the Manual Entry option for adding clients. When clicked, opens a wide slide-over drawer with a form to create a new client.

Who Is Agent Trace For?

Fri, 30 Jan 2026 00:00:00 +0000

Cursor released Agent Trace, an open spec for tracking which code in a repo was written by an LLM. It records the model, the tool, the conversation, and the exact line ranges — all appended to a JSONL file in your project.

The pitch: “As agents write more code, it’s important to understand what came from AI versus humans.”

I spent time reading the spec and the reference implementation. The engineering is solid — clean schema, thoughtful extensibility, good partner list (Amp, Amplitude, Cloudflare, Cognition, Google, Vercel). But I kept circling back to one question: what do you do with this data?

Rich Links for Lazy Developers

Sat, 10 Jan 2026 00:00:00 +0000

When I share a blog link on Twitter or Slack, it shows up as plain text. No preview image. Just a URL like granda.org/en/2026/01/02/claude-code-on-the-go/.

I needed Open Graph images. The standard approach: manually create a 1200x630 image for each post. That’s tedious. I asked Claude to automate it.

flowchart LR
 Push[git push] --> GHA[GitHub Actions]
 GHA --> Hugo[Hugo Server]
 GHA --> PW[Playwright]
 PW -->|screenshot| Hugo
 PW --> IMG[OG Image]
 IMG --> Commit[git commit]
 Commit --> Deploy[Deploy]

flowchart TB
 Push[git push] --> GHA[GitHub Actions]
 GHA --> Hugo[Hugo Server]
 GHA --> PW[Playwright]
 PW -->|screenshot| Hugo
 PW --> IMG[OG Image]
 IMG --> Commit[git commit]
 Commit --> Deploy[Deploy]

The Setup

I told Claude the problem: posts need social preview images, but I don’t want to create them manually. Screenshot the article content, save it as an OG image, update the frontmatter automatically.

Unsubscribe

Sat, 10 Jan 2026 00:00:00 +0000

Processing...

My QA Engineer is an LLM

Fri, 09 Jan 2026 00:00:00 +0000

Claude can click buttons.

That sounds trivial, but it changes how I build UIs. With Playwright MCP, Claude doesn’t just write code—it opens a browser, navigates to localhost, and verifies that things actually work. It catches bugs I’d miss in code review.

The Setup

Playwright MCP gives Claude browser automation. I run it with headless Chromium:

{
 "mcpServers": {
 "playwright": {
 "command": "npx",
 "args": ["@anthropic-ai/mcp-server-playwright", "--headless"]
 }
 }
}

Now Claude can navigate, click, type, and screenshot. It sees what users see.

Rolling My Own Newsletter with Claude

Wed, 07 Jan 2026 00:00:00 +0000

Day two of the Hacker News traffic spike. Forty thousand visitors, no way to reach them again. I needed a newsletter signup.

I looked at Buttondown, Beehiiv, Substack, ConvertKit. All overkill. I just needed to collect emails. I didn’t need campaigns, analytics, or subscriber management. And I wanted to own my data.

So I asked Claude to build it.

flowchart LR
 User([User]) --> Form[Newsletter Form]
 Form -->|POST /api/subscribe| Worker[Cloudflare Worker]
 Worker --> KV[(Cloudflare KV)]
 KV -.->|Daily sync| GHA[GitHub Actions]
 GHA -.-> Repo[(subscribers.jsonl)]

flowchart TB
 User([User]) --> Form[Newsletter Form]
 Form -->|POST /api/subscribe| Worker[Cloudflare Worker]
 Worker --> KV[(Cloudflare KV)]
 KV -.->|Daily sync| GHA[GitHub Actions]
 GHA -.-> Repo[(subscribers.jsonl)]

The Setup

I gave Claude a Cloudflare API token via environment variable and described what I wanted: a form that collects emails and stores them somewhere I control.

Claude Code On-The-Go

Fri, 02 Jan 2026 00:00:00 +0000

I run six Claude Code agents in parallel from my phone. No laptop, no desktop—just Termius on iOS and a cloud VM.

The Setup

flowchart LR
 A[Phone] -->|Termius + mosh| B[Tailscale VPN]
 B --> C[Vultr VM]
 C --> D[Claude Code]
 D -->|PreToolUse hook| E[Poke webhook]
 E -->|Push notification| A

The loop is: kick off a task, pocket the phone, get notified when Claude needs input. Async development from anywhere.

Infrastructure

A Vultr VM in Silicon Valley:

Letting the AI Do It

Sun, 28 Dec 2025 00:00:00 +0000

I needed a dev server: VM with SSH access and dev tools installed. I described what I wanted to Claude Code and let it handle the setup.

The Prompt

I asked for a Vultr VM with 8 cores and 32GB of RAM, accessible only through Tailscale, with persistent sessions that survive network drops.

What Claude Code Built

1. Provisioned the VM via Vultr API

It created an 8-core/32GB Ubuntu instance, waited for the IP, then SSH’d in as root.

Automatic Blog Translations With Claude and GitHub Actions

Tue, 23 Dec 2025 00:00:00 +0000

Every post I write gets translated automatically. Claude handles the translation, commits the results to main, and the translated versions deploy alongside the original.

The Flow

The translation workflow runs in GitHub Actions when English content is merged to main:

on:
 push:
 branches: [main]
 paths:
 - 'content/**/*.en.md'

flowchart LR
 B[Merge to main] --> C[Translations generated]
 C --> D[Committed to main]
 D --> E[Production deploy with all languages]

The Prompt

The workflow uses claude-code-action:

About

Mon, 22 Dec 2025 00:00:00 +0000

Abstract

This document provides background information about the author of this blog.

1. Background

Software engineer with 15 years of experience, 10 spent working in the field. Areas of focus include:

Distributed systems
Backend development and infrastructure
Full-stack and end-to-end product development
User interface design
Building secure and reliable systems

2. Contact

Always happy to connect with others in the industry: @mtt

Authorship

Mon, 22 Dec 2025 00:00:00 +0000

Abstract

This document describes the authorship methodology employed by this blog.

1. Introduction

This blog is largely generated with the assistance of Large Language Models (LLMs). The author provides direction, editing, and quality control; the machine provides initial drafts and translations.

2. Rationale

The author believes in transparency regarding AI-assisted content creation. Rather than obscuring the tools used, this disclosure aims to:

Set appropriate expectations for readers
Contribute to honest norms around AI content
Acknowledge the collaborative nature of modern writing

3. What This Means

Ideas: Human-originated
Structure: Collaborative
Prose: Primarily AI-generated, human-reviewed
Translations: Entirely AI-generated (not human-reviewed)

4. Quality Assurance

English content undergoes human review before publication. Translations are generated automatically and may contain errors or awkward phrasing. The author takes responsibility for the final published English content.

Hello World

Sat, 20 Dec 2025 00:00:00 +0000

Welcome to my blog! This is my first post.

I’ll be sharing ideas, notes, and things I find interesting here.