December Awakening

Everything accelerated, again, in Nov-Dec-Jan. Lots of people are noticing.

MCP Apps: Custom UIs Inside Claude

MCP Apps bring your custom UIs inside Claude. I built 4 and share what I learned: how they relate to MCPs and skills, rules of thumb, and more.

Helping Claude See ... Diagrams

I built an MCP that lets Claude see its own Excalidraw diagrams, iterate on them, and save them to Obsidian.

The Short, Sweet Life of Clawdbot

Peter Steinberger's move to OpenAI marks the end of the saga; or Part 1 anyway.

I Can Now Improve, Automate, Fix and Document Everything

Thanks to The Workflow and my AI pair, Claude Opus 4.6.

Leveling Up Claude Code with Boris Cherny's Help

Cherny's thread on how he personally uses Claude Code gave me a roadmap.

Claude Cowork Now Supports Plugins

Anthropic brings Claude Code's plugin system to Cowork, along with a collection of knowledge work plugins and a slick customization system.

Skillport: Claude Skill Sharing at Scale

Or: How I spent my Christmas vacation.

OpenAI Quietly Adopts Skills

Which makes me happy. Via Simon Willison.

Antigravity Gets an A+ on My Browser QA Task

Google DeepMind's Antigravity nails a simple autonomous browser QA test of this blog where Claude Code stumbled

Kevin Hou on Google DeepMind's Antigravity

My thoughts on Kevin Hou's talk about Antigravity, DeepMind's new agent-first development platform.

Really Sam? Code Red? We Have Questions.

Starting with, are you familiar with A Few Good Men?

AI Vision vs Computer Vision: Are the Curves Crossing?

I was wrong about Computer Vision OCR being dead in 2024, but it appears that's about to change as AI Vision rises.

Talk Like Ethan Mollick, Matey

A Claude Skill that drafts LinkedIn posts from my longer essays, using Ethan Mollick's plainspoken style as a model.

Anthropic: Stop Building Agents, Build Skills Instead

A great AIECS session from Barry Zhang & Mahesh Murag, the creators of Anthropic Agent Skills.

What If You Don't Need MCP At All?

Perhaps the best way to lower MCP token overhead is to use Bash CLI tools instead.

Claude Skills in Claude Code: A Compleat Guide

I migrated my Cursor Rules to Skills in Claude Code. Simple, elegant, powerful.

These Comments from Vercel CTO Malte Ubl Struck Home

Vercel CTO Malte Ubl shares wisdom on building AI products: stay humble about user needs, dogfood your abstractions, and pick the right problems for agents.

Autonomous Coding Agents: Fire and Forget

Autonomous coding agents let you fire off coding tasks and walk away. Safe YOLO mode changes everything.

When Claude Code Skates on Thin Ice

I love Claude Code and Sonnet 4.5, but when training data is thin, the reasoning failures are hilarious

Pydantic AI Reaches v1

Which, for me, begged the question, what exactly is Pydantic AI?

ChatGPT 5 Thinking Is So Damn Smart

I needed to round out my knowledge of Pydantic AI. ChatGPT 5 Thinking absolutely nailed it.

Chroma: RAG is Dead; Long Live Context Engineering

Jeff Huber explores how Chroma is changing the conversation away from RAG to context engineering

Great Rules In, Great Results Out

An investment in writing great rules for your AI tools will be repaid many times over.

Claude Code Deep Dive: Hugo to Astro+Beehiiv in 9 Days

Migrating from Hugo to Astro/Beehiiv with Claude Code doing ~87% of the work

Andrew Ng on Leaders Still Doing Things the Way They Were in 2022

Ng put into words something that I've been thinking for a while now.

Two Interviews with OpenAI President Greg Brockman

Highlights from two recent interviews by the Latent Space team with Greg Brockman

Use GPT-5 with Playwright MCP to Auto-link Posts

Quick Cursor workflow using GPT-5 plus Playwright MCP to add named-entity links in blog posts.

Early GPT-5 Impressions

My vibe assessment: different, powerful, bumpy launch

Running OpenAI's gpt-oss on Mac Mini with MLX

Thanks to LM Studio and the community, it's insanely easy. And gpt-oss is insanely good.

Orta Therox: 6 Weeks of Claude Code

Orta Therox reflects on six weeks using Claude Code for programming, describing it as transformative for maintenance tasks and side projects

AI Engineer World's Fair 2025: My Day 2 Highlights

My observations on Day 2 of the AI Engineer World's Fair 2025

A Better Way to Give Claude Code Context

I like Scott Werner's more structured approach

ChatGPT Images Pictorial

Cthulhu chimes in on the environment

AI Engineer World's Fair 2025: My Day 1 Highlights

Key takeaways and observations from Day 1 of the AI Engineer World's Fair 2025

AI Writing Index, Feb '23 to Jun '24

My early AI work, in index form, with reflections

Apple Intelligence Initial Thoughts

Mostly from other people, I'm still digesting

Observable Framework Delivers Blazing-fast Data Dashboards

An exciting new offering, launched in February, from the team that brought us Observable Notebooks

A Quick Look at GPT-4o

I did a quick benchmark of the new GPT-4o model versus GPT-4-turbo. Roughly twice the speed, with improved Vision accuracy.

Using Tailscale to Access Amazon VPCs, EC2 Instances, and RDS Clusters

Tailscale has been simple to set up and manage, but also amazingly flexible.

Russ Cox's XZ Timeline

Nice writeup from Russ Cox on the (incredibly long) timeline of the XZ backdoor

Simon Willison's LLM Tool: Now I Have 50 LLMs

I updated my installation of Willison's LLM tool to add plugins, and now I have 50 LLMs at my fingertips, including 15 local models, which get installed on demand.

Two Worthwhile Reads From Simon Willison

Claude 3 Opus and GPT-4 to Tackle a GIS 'Sidequest'; and Getting GPT-4 to Write, Compile, and Run C Code

Moxie on Murder

I did a double-take when I saw Moxie Marlinspike in the credits as co-starring in A Murder at the End of the World episode 6 (SLIGHT SPOILERS WARNING)

Willison: The killer app of Gemini Pro 1.5 is video

Simon Willison tries out Gemini Pro 1.5 on video, and suggests its 1M token context size opens up powerful new opportunities using video prompts

Using Multimodal AI to Capture and Enrich Heirloom Recipes

I applied OpenAI's GPT-4 Vision model and Chat Completions API to preserve a treasure-trove of legacy family recipes.

Gruber Responds to Gurman's Report of AI Anxiety at Apple

John Gruber: Apple anxiety about AI/ML team's ability to deliver, constrained by privacy requirements.

The Ballad of Bull Cook, A Wacky Book

George Leonard Herter's classic guide

Cloudflare CAPTCHA Hell

Cloudflare's CAPTCHA nonsense: a sign they're getting too dominant?

Gruber Translates Linda Yaccarino's Company-wide Memo on the X Rebrand

He calls it 'translation from hostage code'—so funny, one of his best in this genre.

Twitter and Its Successor States

Medieval historian Eleanor Janega hilariously draws the parallels between Elmo's Twitter (aka X) and the 'fall' of Rome

Ethan Mollick: "How to Use AI to Do Stuff"

Great roundup, and I agree with most recommendations. Bing, maybe not.

OpenAI Concedes: AI Can't Detect AI

OpenAI quietly shuts down its AI detection tool due to poor accuracy

An Appllama Week in AI

Meta makes waves with Llama 2, while Bloomberg pumps itself with Apple LLM non-news.

Did GPT-4 Code Interpreter Escape From Its Sandbox?

I gave Code Interpreter a workout this morning, and it appeared to exit the building.

Just-released GPT-4 Code Interpreter is a Big Deal, Part 2

Latent Space had an "emergency pod" about Code Interpreter and 17,000 people joined.

Just-released GPT-4 Code Interpreter is a Big Deal, Part 1

Wharton Associate Professor Ethan Mollick has an excellent introduction.

'Commoditizing the Petaflop' with George Hotz of the tiny corp

This Latent Space podcast was a mind-blowing conversation that puts George Hotz and the tiny corp on my 'Follow Closely' list.

Really, OpenAI?

ChatGPT+ subscribers can prevent OpenAI from using their inputs as training data. That is, so long as they forego the service's second-best feature.

How to Ask ChatGPT a Technical Question (BoorishBears on Hacker News)

BoorishBears: Three-part ChatGPT prompt technique for technical questions - considerations, implementation, review.

The Lone Banana Problem

Some images are tough to write prompts for in your favorite Art AI

Hugo Responsive Images, Thanks to Bryce and ChatGPT

TIL how the html `picture` element actually works ...

Of Moats and Moat Busters

Open Source, Commercial-friendly AI Challenges the Major Closed AI Players

Midjourney V5.2 (Ars Technica)

This Midjourney update looks fantastic. I'm working on a Crafty's Illustrated essay, and plan to give V5.2 a thorough workout when it's time for imagery.

Apt Description of React on Hacker News

jankiel: React's complexity grows gradually, with many hidden footguns requiring idiomatic patterns.

MosaicML's Open Source MPT-7B Model Writes an Epilogue to The Great Gatsby

MPT-7B AI model: Creative epilogue to The Great Gatsby featuring narrator's awakening from dream.

Comparing Adobe Firefly, Dalle-2, OpenJourney, Stable Diffusion, and Midjourney

Muhammad Usman does a nice comparison of art AIs. To me, the richness of the outputs is amazing. I'm a heavy Midjourney user, and still prefer its output, though the alternatives are impressive as well.

Apple Vision Pro (Part 2) – Hardware Issues

Karl Guttag on the devil-in-the-details of Apple Vision Pro hardware.

I-JEPA: The first AI model based on Yann LeCun’s vision for more human-like AI

As pointed out by Jesus Rodriguez on TheSequence, 'With all the hype surrounding generative AI, we sometimes overlook the thrilling advancements in other areas of the deep learning ecosystem.' I-JEPA from Meta research is one such.

A Gnarly Hugo-Cloudflare Build Problem, Resolved

I had Hugo builds running on Cloudflare for a day. Then, mysteriously, they stopped working.

Apple Really Did Ignore the AI Emergence at WWDC

As I watched the WWDC Keynote and Platforms State of the Union from this year's WWDC, I was amazed that Apple appeared to be ignoring the current massive emergence in generative AI. Later, I confirmed it.

Unlimited Stupid Questions

Possibly the most important benefit of ChatGPT ...

Ownership in Small and Medium Technology Companies

An important topic, addressed thoroughly by Eric Brooke, former SpotHero CTO.

Le Mans, Unabridged

This being Le Mans weekend, here's a Le Mans deep-dive about watching the entire 24 hours, unabridged.

Vivek Haldar on GitHub Copilot

Vivek Haldar: GitHub Copilot reduces project activation energy, enabling weekend-sized scripting projects.

First Impressions of Vision Pro and visionOS

Gruber conveys the essence of the experience better than other writeups I've read so far.

Is Runway Gen-2 Midjourney for Video?

We Tested Gen-2 and Share the Resulting Videos

Faster sorting algorithms discovered using deep reinforcement learning

From nature.com, an AI story that isn't *generative* AI.

Orion Browser by Kagi

I'm late to the party, but the Orion browser from Kagi is really interesting.

Apple Vision

Stratechery's Ben Thompson with an excellent take on Vision Pro.

It’s infuriatingly hard to understand how closed models train on their input

Simon Willison digs into the question of whether the big closed LLMs are training their models based on users' input.

ChatGPT Code Review

Another great use: coming up to speed on an unfamiliar codebase.

A Day Late

What happens if I miss a day ??!!

OpenAI’s plans according to Sam Altman

Summary of Raza Habib's interview last week with Sam Altman.

Lawyer cites fake cases invented by ChatGPT, judge is not amused

Simon Willison does a deep review of the ChatGPT aspects of the recent case where a lawyer submitted ChatGPT-hallucinated case law.

How to Write Every Day—and Why

With help from others, I made the transition from painful-episodic to enjoyable-daily writing.

Camus on True Knowledge (From Myth of Sisyphus, 1955)

Albert Camus: Modern despair of true knowledge except among professional rationalists.

Why I Use Mimestream for Gmail

Mimestream has me reconsidering Apple Mail as my daily driver.

CLI tools for working with ChatGPT and other LLMs

I installed Simon's llm, ttok, and strip-tags CLI tools and got them working, great stuff.

Working With My ChatGPT Intern

Yesterday was a new record.

The Emergence

Season 1, Episode 4

March Madness, AI Style

Observations on AI's March acceleration

Venture Funds Arrive in the Mastodon Space

How to think about the arrival of commercial offerings in the Fediverse

Alarmed About AI?

John Seely Brown, former director of Xerox PARC, has helpful advice—from April 2000!

Introducing Crafty CTO

Turning the page: the story of Crafty CTO

Accelerating Crafty's Launch Using AI Models

Thanks to AI-based tools like Midjourney, Crafty shaved at least a month off its time to launch.

Blog/Website Relaunch: Under the Hood of craftycto.com

Why Fractional?

Building with Shape Up

Shape Up Your Agile