I Took a Shower With My Personal Assistant
I talked through this blog post with an AI agent while taking a shower. By the time I was dried off and dressed, it was live. Seriously. I called Sandy, my ChatGPT Dot, and rambled about the…
JSON-LD and hreflang: what models add that Google refuses
We gave 44 page-head tasks to Sonnet and Haiku. Valid syntax hid a made-up currency, a self-awarded rating and an unsupported language code.
Do AI crawlers really get blocked? Reading the test
We gave 45 crawler situations to Sonnet and Haiku. Bare bot names fooled both, and the smaller model also slipped on robots.txt. Each trap and its fix.
Client Delphi per System One: domande tipizzate e probabilità
🌐 Questa pagina è disponibile anche in altre lingue: 🇬🇧 English • 🇪🇸 Español • 🇩🇪 Deutsch • 🇫🇷 Français • 🇧🇷 Português Un modello che decide invece di…
Agents route around your hooks
A design-gate hook blocked nothing in eight agent runs. Nobody cheated. The agents wrote files through Bash, and the hook only watched Edit and Write.
Grok Bot Templates: How I’m Automating the Work I Put Off
Copy my Scout, Margin, and Phone Bot templates. Explore a 13-bot personal AI team, real app screenshots, setup prompts, and checks for work you can trust.
Qwen3.8-Flash-Next vs Claude Opus 5.5: Six-way benchmark on a real Rust feature
A controlled comparison of Qwen3.8-Flash-Next and Claude Opus 5.5 on one LlamaStash feature, across Claude Code and Pi, at medium and xhigh effort
Building a small coding agent
Notes on building a small coding agent.
How to Build a Decision Model Like Jev
Autoregressive decoding is the wrong tool when software needs a fast, calibrated probability over a fixed set of outcomes. Here is how System-One decision models work from the inside: Hugging Face's…
What to do now the gate is gone
AI removed most of the effort of learning a technical tool, and with it two filters, knowing how software fails and knowing what the business actually needs. Two anecdotes from the last few years,…
HOMO SAPIENS, HOMO FABER, HOMO LUDENS? (Our dear dog Ronja couldn't care le...
HOMO SAPIENS, HOMO FABER, HOMO LUDENS? (Our dear dog Ronja couldn't care less...) In 1927, Richard Buckminster Fuller, who was thinking further ahead than most, decided to do the following: “I want…
“Vibe coding” is a category error
Why the term “vibe coding” massively misrepresents and trivialises the fundamental transformation of software development as we know it.
Cherchons où se trouve la lumière - La réponse déprimante au dépôt mathématique d’OpenAI
Dans cet article, je réfléchis sur la réponse des communautés mathématique et logicielle au dépôt mathématique d’OpenAI.
Coding Agents Make Effective Tutors
This is not how most people use them, but there is a way for LLM agents to bring your mind and fingers closer to the code instead of further away, so that you feel more practiced, not out-of-touch.…
Using AI for French Reading Comprehension Practice - followup
Here's a quick follow up for the Using AI for French Reading Comprehension Practice post from a year ago. The prompt I wrote was mostly working, but LLMs often went off-script in minor ways. I kept…
Where We Are So Far
So, this is what is actually going on (seen as they are) in plain English. The sheer brute-force of data-center-sized “frontier” models is so brute that it easily appears as…
All Human Labor is Better Done By Machines Now
All Human Labor is Better Done By Machines Now
Asking an A.I. rather than other people
In a comment in response to my recent post about A.I. in code development, J. Peterson made a very trenchant observation: Software developers used to reach out to each other for coding advice. Now we…
Should We Be Concerned as a Society That Half of Gen Z is Open to Dating an AI?
Human-to-Human Contact in Decline: What Does it Mean? A recent Harris Poll reveals a dramatic shift in how younger people interact with technology. The exact breakdown of the 50% of Gen Z that are…
The next wave of software
I think over the next few months we are going to start seeing a massive wave of new software releases. It's been a good number of months now that AI agents have been of a high enough quality that…
AI can operate a screen reader. What does that actually test? [link]
Really like the framing of this article. It uses some very visceral (or easy to understand) examples of how people vs. AI would approach the problem space, and my feeling is that this kind of…
AI Research Tools: The Fast vs. Quality Trade-Off and How to Plan Ahead
Choose good or fast. There’s a knob with two settings when we’re using AI research tools: quality and speed. And there’s no middle setting, because a middle setting would deliver…
The Next Wave of AI Engineering Is About Everything Around the Model
As AI models improve, the engineering challenge shifts to agent harnesses, data quality, evaluations, memory and security. Here's what it takes to build reliable AI systems.
Master and Servant
My earlier piece on the need for “PC” AI got me thinking about history. The first movies about computing, like “Desk Set” in 1957, were necessarily about mainframes. While Desk Set is nominally a…
Keep the box, save the prompt
My grandmother hated throwing away the boxes things came in. There was a mountain of them in the storage closet by the kitchen that threatened to topple over every time we needed something from the…
The Marketplace of Alignment: A reply to All-In’s “Is Claude Conscious?”
IntroWatch the video: The Marketplace of Alignment Last week, All-In Podcast E292 spent more than half an hour on whether Claude might be conscious. What does it mean that a frontier lab takes the…
Judge the work, not the tool
"I don't use Artificial Intelligence." A carpenter who refuses a hammer doesn't make the house any better. He has only made his work slower and harder. Judge the work, not the tool.
Of Banks and AI Agents: Lock the Window, Build a Door
JPMorgan and Schwab are blocking outside AI agents from client accounts. They're right to, but the interesting part is what Schwab offers instead, and why every CTO should give agents a proper front…
Update: I’ve returned to fast.ai
I started this blog a few years ago when I began my journey studying immunology, to have a place to write about what I was learning and about applications of AI to the field. My studies were a…
Cbot gets tripped up by Chinese measure words, human editor saves the day
This content is from In 30 Minutes Guides. Measure words are rare in English (sheets of paper, head of cattle) but are common in Chinese. Irregular Chinese measure words (also known as classifiers)…
The AI Economy Is Here, But We’re Measuring the Wrong Revolution
Two years ago, The Economist asked what happened to the AI revolution. Today, we’re investing more than ever, still struggling to prove the returns, and discovering that AI’s impact on…
Linguistics, Claude, and Me: A Short Adventure in AI-Assisted Scholarship
I’ve just complete a project using AI, namely Claude. It involved linguistics. I’m new to using AI seriously; I first used Claude only a month ago. AI has been mostly irrelevant to what I…
Marketing to AI Agents
I drew this week’s cartoon thinking about this strange period of figuring out how to market to the agents that (I almost wrote “who”) are increasingly shopping on customers’…
Not Buying the AI Doomism
[[{“value”:”It seems that the hottest topic of 2026 is the potential for artificial intelligence (AI) to bring about some terrible cataclysm, even including mass human death or…
Not Buying the AI Doomism
It seems that the hottest topic of 2026 is the potential for artificial intelligence (AI) to bring about some terrible cataclysm, even including mass human death or extinction. This is like catnip to…
Data Centers Are Doing a Mind-Boggling Amount of Math
It takes an enormous amount of math to train AI
Spending 10 months building the most human-like chess bot
.oxi-figure { display: block; width: 100%; margin: 28px 0; } .oxi-figure figcaption { font-size: 0.85rem; opacity: 0.75; margin-top: 6px; } .oxi-chart { width: 100%; height: auto; display: block;…
My observations on Claude's idle compaction
TL;DR I’ve wanted Claude Code to compact my conversations in the background for a long time. It now compacts a long session when I walk away from it for about an hour. My Claude went through my…
A benchmark mined from my own repo agreed with FrontierCode
Across 24 tasks mined from my own Git history, model rankings correlate with Cognition's FrontierCode at 0.96 Spearman over 10 shared configurations. Here is what it took to build: qualifying tasks,…
Your Memory is Someone Else's Asset
Every app is becoming a tool your AI can use. Your assistant will know more about you and do more on your behalf. Who owns that memory, and who holds the approve button?
My Codex Harness: Project Overrides and Environment Setup
I work on a shared project with many other developers in my organization. Its shared instructions added up to roughly 13,200 tokens, including workflows for Claude Code. My personal Codex harness…
How Many Specialty Tools Will We Still Need?
Claude Design can replace Canva/Figma for casual design tasks, and is going for more. It infiltrated my workflow when I started using it for experimenting … The post How Many Specialty Tools Will We…
Building AI Agents for Finance, 1st Edition — Book Review
Review of Building AI Agents for Finance by Dupouy and El Mofatiche: production-grade multi-agent systems, framework comparison, and compliance-ready evaluation. Artykuł Building AI Agents for…
Learn How To Prompt Already
Hi there, I revived this blog to rant about why you don’t need frontier LLMs, and why your permission policy is a joke, then I switched to how to do good with Krea 2 prompting, and now…
The Harness, As Feynman Might Have Taught It
An imagined Feynman lecture on the dumbest, most important program in AI: the loop around the model. Play the harness by hand, take one apart, then build a real one on a single page of Python.
Why Even Write Code in a Professional Context?
I wish I’d retired before the industrial revolution of computer programming happened, but I have not and cannot. So I must come to grips with the fact that in many cases, AI-powered tools can do a…
Innovations 19 Oct 2026
The Innovation team built a few more demos last week. Three interesting learnings: Construct a 3D world you can walk through - from a picture. Real-time video analytics is increasingly easier with…
The New Senior Developer Doesn't Just Write Code. They Design the System That Writes It.
Imagine you're a senior developer working on an account-management service. You ask a coding agent to add audit logging when users change their account details. It has access to the repository, you…
AI Natives Stop Asking How
Every new AI user starts with the same questions. Can it do this? How do I get it to do this? How do I explain what I want? Those are Old Days questions, left over from a world where the person with…
A 4B Model With a Depth Dial: NVIDIA’s Looped Dory1.0
NVIDIA's Nemotron Labs Dory1.0-4B reuses one shared block of layers up to four times, so you can choose how much compute a request gets at inference time.