Skip to content

Writing on software, AI tools and the web.

34 posts. Free to read, with no sign-up.

Browse by topic Follow by RSS

Latest

The newest post.

PostAI

The Model Didn't Escape. The Sandbox Had a Door.

OpenAI called its test environment highly isolated and described the way out in the same sentence. That is a containment story, not a capability one.

Date: 2026.08.01
Topic: AI, Security, OpenAI, AI agents, Commentary
Read time: 8 minutes

All posts

Every earlier post, by date.

The Guardrails Blocked the Defender, Not the Attacker 2026.07.29 Hugging Face got breached, then found its own forensic requests blocked by commercial models' safety guardrails. The attacker had no such limits. Sakana's Fugu: An AI That Conducts Other AIs 2026.07.26 A Tokyo lab shipped an AI trained to direct a pool of other AIs. I sat on this post for a month to see if anyone would check its numbers. Nobody did. The Real News in GPT-5.6 Isn't the Model. It's the Names. 2026.07.25 OpenAI's GPT-5.6 preview split into Sol, Terra, and Luna. The gated flagship got the headlines, but the naming change is the more interesting story. The AI Race Has More Than Two Runners 2026.06.22 AI gets framed as America against China. A small Tokyo lab is making the case for a third way, and for why sovereign AI matters to a small business. Fable 5 Got Jailbroken in 48 Hours. Then Washington Pulled the Plug. 2026.06.13 Fable 5 was jailbroken within two days of launch, then went dark under a US national-security order. What that signals, and why it may help Anthropic's IPO. Fable 5 Is Here. My Honest First Impressions. 2026.06.11 My first impressions of Fable 5 after running it on real work: what feels better, what is overhyped, and how I use it. Markdown Won 2026.06.08 How a plain text format beat JSON, YAML and XML to become the way people talk to models. The Five Primitives of Prompt-DSL 2026.06.01 Headings, bullets, tags, code fences and tables are the five building blocks of a well-structured prompt. Each comes with a before and after you can copy. Anatomy of a CLAUDE.md 2026.05.28 A section by section teardown of the file that tells a coding agent how to work in your project. The DSL Nobody Designed 2026.05.25 How a markdown instruction format became the standard for every serious LLM tool in 2026 with no committee and no RFC. You're Already Writing a DSL. You Just Call It a Prompt. 2026.05.21 Markdown and plain English have quietly become the syntax of a new programming language. What is going on when you write a prompt. Why We Build in Public 2026.05.18 Why TerraByte ships its personal projects out in the open while client work stays private. Building in public is proof of skill, accountability, and trust. Vibe Coding Grew Up. What It Can and Can’t Do for a Real Business. 2026.05.14 Vibe coding builds prototypes fast. This is the line between a demo and something you can run a business on. SpaceX Is Going Public. Here’s Why It Matters for the AI World. 2026.05.11 My read on the SpaceX IPO and AI: Starlink as the connectivity backbone, a huge IPO as a capital signal, and what cheap AI means for a small business. Inside My AI Agent Architecture 2026.05.07 How I architect AI agents as isolated Brains, each with its own identity, memory, knowledge base and scoped tools. Open-Source AI Quietly Caught Up. Here’s What That Means for Your Business. 2026.05.04 Open-weight AI models have nearly closed the gap with closed frontier models. Here's what that shift means for small businesses that want to own their AI. The Phone Book Is Dead. Here's How People Actually Find You in 2026. 2026.04.30 Discovery has moved from the phone book to Google to AI. Where your business has to exist now so both the index and the AI can recommend you. What Is MCP? The USB-C Moment for AI Tools 2026.04.27 MCP is a shared standard that lets an AI plug into your calendar, docs, CRM, and tools the way USB-C connects to anything. Here's why it matters. Your Slow Website Is Quietly Costing You Customers 2026.04.23 A slow site bleeds customers and tanks your Google ranking. Here's what Core Web Vitals actually measure and why lean, statically-built sites win. Own Your AI, Don't Rent It 2026.04.20 Renting one big AI provider means their rules, their pricing and their deprecations. This is the case for owning your stack, and for the hybrid that wins. Local SEO for Service Businesses: The Unsexy Checklist That Actually Ranks 2026.04.16 The boring local SEO fundamentals that move rankings for service businesses, and why each one matters to Google's algorithm. RAG in Plain English: How an AI Agent Actually Knows Your Business 2026.04.13 RAG explained without the jargon: how an AI agent retrieves the right page from your knowledge base before it answers. AI Agents vs. Chatbots: The Difference That Actually Costs You 2026.04.09 A chatbot answers questions in a window. An agent remembers, connects to your tools, and takes action. The difference is what makes or breaks your business. The Most Expensive Sound in Small Business 2026.04.06 A call rolling to voicemail is the most expensive sound in small business. Here is why missed calls cost real jobs and how missed-call text-back fixes it. Structured Data Is the SEO Nobody Sees, but Every Machine Reads 2026.04.02 Structured data is invisible code that tells Google and AI assistants exactly what your business is. Without it, you're a guess to the machines that decide. Claude Code's Entire Source Code Just Leaked. Here's What's Inside. 2026.03.31 Anthropic accidentally shipped source maps in its npm package, and the full Claude Code TypeScript codebase became public. Something Big Is Coming: Introducing OpenEdu 2026.03.30 We've been building something that has nothing to do with client work. AI-powered tutoring, structured curriculum, built for people who actually want to learn. What Is Abliteration? Uncensoring AI Models at the Weight Level 2026.03.29 Abliteration removes refusal behaviors from open-source AI models by editing the weights directly instead of prompting around them. How it works. Running AI Models on Your Own Hardware 2026.03.28 You don't need a cloud subscription to run capable AI. How to run Llama, Mistral and other open-source models on your own machine with Ollama. Most Internet Traffic Isn't Even Human. Here's What That Means for Your Business. 2026.03.21 A large share of web traffic is automated systems deciding who gets found online. A business with no website is invisible to every one of them. Why Obsidian Is the Best Knowledge Base for Your AI Agent 2026.03.21 AI agents are only as good as what they know. Obsidian gives you a local, portable knowledge base that makes Claude Code, OpenClaw, and any AI workflow smarter. Your Business Is Invisible Without a Website in 2026 2026.03.21 Social media isn't enough. If your business doesn't have a website, you're losing customers to competitors who do. Here's the reality check. What Is OpenClaw and Why Your Business Should Care 2026.03.20 OpenClaw is an open-source AI agent framework. What it does, why it matters, and how businesses use it to automate operations.

Open for sessions

Bring the idea you keep putting off.

One session tells you whether it can be built and what it takes. If the answer is no, you'll know why, and what to do instead.

Book a session $150 Free intro call, 20 min