Blog
Thinking
out loud
Notes on AI, agents and building products, from someone who ships both the design and the code. Written by me, every word.
AI
Fable 5.1 is here. The 20x plan is still not 20x.
Anthropic shipped Claude Fable 5.1 today at the same token price and a much cheaper cache. The subscription math underneath did not move, and the 50 percent boost ends in two weeks.
September 1, 2026
AI
Apple caught up to the language. The 512GB Studio is still a waitlist.
Apple now sells Mac Studio on token bills and frontier-class inference. I ordered the 256GB M5 Ultra and retired the laptop. The configuration that holds GLM-5.2 properly is the one they will not take an order for.
August 28, 2026
Tools
Omarchy Quattro: the Linux desktop I did not expect to like
DHH spent a year turning his Arch setup into a distro. I installed it on two machines this week. Notes on why it works and where the agent story actually lands.
August 23, 2026
AI Workflow
Grok Bot has a computer. The product is that you never operate it.
Making the machinery disappear is the interesting part. The trust boundary should not disappear with it.
August 12, 2026
AI Workflow
I stopped choosing between Fable and Sol
Claude Fable 5 in one cmux pane, GPT-5.6 Sol in another. One works, the other reviews, and I keep the final call.
July 19, 2026
Tools
When a mobile app changes the privacy contract
Cursor's iOS launch exposed a deeper product problem: enabling a new feature should not quietly weaken a privacy promise users already chose.
July 3, 2026
Self-hosting
A homelab should give time back
Self-hosting is useful when recovery, updates and monitoring are designed in. Otherwise you bought yourself another job.
July 2, 2026
AI
The frontier went behind a velvet rope. Mine runs on a Mac Studio.
GPT-5.6 opened with a limited partner preview and Fable 5 disappeared after its first release. Meanwhile GLM-5.2 handles much of my daily work on the Mac Studio in the other room.
June 28, 2026
AI Workflow
What running 12 AI agents actually changed
The models were not the hard part. Priorities, permissions, evidence and stopping rules were.
June 20, 2026
AI
ChatGPT is on your phone. Apple just put AI in it.
Almost every phone already has a frontier model on it, running in someone else's data center. WWDC was Apple betting on the opposite architecture: computation that starts in the computer you bought and reaches for GPU server halls only when it has to.
June 12, 2026
AI & Development
Three Months With Opus 4.6, Then Everything Changed
Three months in, Opus 4.7 happened and a free local model caught what I couldn't. Notes from the daily driver seat.
April 27, 2026
AI & Development
Monitor VRAM usage across your LAN
I needed to know how much VRAM my Mac Studio was using from another room. No tool showed it. So I built a widget.
April 12, 2026
AI
Anthropic built its most capable model yet. Then decided the world wasn't ready.
Claude Mythos Preview changes the economics of finding software vulnerabilities. Anthropic's decision to gate access is understandable, and still concentrates an uncomfortable amount of power.
April 8, 2026
AI
Anthropic didn’t kill OpenClaw. It killed the illusion.
Anthropic made it explicit: third-party harnesses like OpenClaw no longer count against Claude subscription usage. That changes the economics, the trust and the ground AI builders stand on.
April 6, 2026
AI
Dream belongs in the workflow, not another tab
Dream is my name for multimodal production inside an agent job. The useful part is carrying context through review, delivery and a passing build.
April 6, 2026
AI & Development
What 100K GitHub Stars Can and Cannot Tell Us About AI Coding
Claw Code crossed 100,000 GitHub stars in a day. The record claim matters less than the engineering underneath it: verification loops, small skills and disciplined context.
April 2, 2026
AI
Claude Code's source leaked. Axios was hijacked the same day.
Anthropic shipped source maps in their npm package, exposing 500k+ lines of Claude Code. The same day, two malicious Axios releases turned a routine dependency update into remote-code execution. Two npm trust failures in one day.
March 31, 2026
AI
What the Mythos leak revealed before the launch
Before Anthropic announced Project Glasswing, an exposed draft showed how the company was thinking about capability, cybersecurity and access.
March 28, 2026
AI Development
It's now faster to build the app than to find it
I built TermAway because no SSH client did what I needed. Weeks later, someone else shipped the same thing.
March 25, 2026
AI Tools
Claude Cowork Just Got Serious
Anthropic's latest Cowork update lets Claude carry work across files, apps, devices and schedules.
March 24, 2026
Design Leadership
AI-native design isn't a chatbot on top
The hard design work is deciding when the system acts, asks and stays quiet.
March 23, 2026
AI Tools
OpenClaw on VPS vs Local: Why the Current Advice is Wrong
Most AI creators are recommending VPS setups for OpenClaw. For most people, this is the wrong approach.
February 12, 2026
Startups
The 48-Hour Validation Rule
If you can't get 20 people to sign up for a free waitlist in two days, you don't build it.
February 5, 2026
AI Workflow
Managing My AI Team From a Dashboard
How the Codex app changed the way I think about AI-augmented development.
February 4, 2026
AI & Learning
Anthropic Just Studied How AI Affects Learning to Code
AI can help you ship faster. Anthropic's study asks what happens to the learning underneath.
February 3, 2026
Career
The New Junior Playbook: Trust Over Credentials
Why the traditional career ladder is broken, and what's replacing it.
February 2, 2026
AI & Teams
Who Reviews the AI's Code When the Senior Leaves?
The sustainability question nobody's asking about AI-augmented development.
February 1, 2026
Ventures
Kerber.ai: A Venture Studio Powered by AI
What began as curiosity about AI coding turned into a venture studio.
January 31, 2026
A second opinion with scars
Need a technical and creative perspective on what you're building?
I advise teams where product, design, engineering and AI meet — especially when the obvious answer is not the useful one.