The archive
Things I have written,
mostly on purpose.
Short opinion takes and deep technical essays on software, strategy, and the mess in between. New posts appear here when I have something worth saying, not on a schedule.
Claude misused (7/7) : The cheap model you used may have been Claude
Anthropic says two Chinese AI labs quietly forwarded their own customers' requests to Claude and passed the answers off as theirs. If that's right, the people most exposed aren't Anthropic. They're the users whose prompts made the trip.
Claude misused (6/7) : Three in four matches on these dating apps were AI
A studio ran more than 20 dating apps that promised real people. Most of the profiles were AI personas told never to admit it, and from inside any single chat it looked like an ordinary companion app.
Claude misused (5/7) : A filter cannot tell a vaccine lab from a weapons lab
Anthropic published five cases of scientists using Claude for research that could go either way. Where the filter was built to block, it blocked. The harder cases are the ones it was never meant to stop.
What you should never paste into a chatbot
A chat box looks like a notebook and behaves like a form you're submitting to a company. Four kinds of thing should never go into it, and each app has one switch that takes two minutes to find.
Claude misused (4/7) : A weapons cell in Yemen used it as a software team
They fired a guided rocket, it appears to have failed, and within hours they were back asking Claude why. The hardware was always theirs. What AI replaced was the programmers.
Claude misused (3/7) : Cheap AI propaganda mostly reached nobody
A firm in Istanbul built a platform to target voters in all 222 Malaysian constituencies, and Anthropic grades its real reach at two out of six. The operations that worked already owned a broadcaster.
Turn on theft protection before your next trip
A phone snatched out of your hand is already unlocked. Android and iPhone both have a setting for exactly that, and on a lot of phones it's still switched off.
Claude misused (2/7) : One consultant can now build a surveillance state
A single contractor in Bamako used Claude to engineer a system covering roughly 25 million SIM cards. Anthropic banned the account, and says the ban did nothing to the system itself.
One Android setting blocks ads without installing anything
Android has had an ad blocker sitting in Settings since 2018. It's one text field, it needs no app, but it won't touch YouTube.
Claude misused (1/7) : A good hack no longer tells you who did it
Anthropic's latest threat report puts a Russian espionage unit, an extortion crew and two undergraduates side by side, and their work looks about the same. Skill used to be how you told them apart.
RAG is an open-book exam for AI
A model on its own answers from memory, which is why it guesses. RAG is the trick of finding the right page and putting it on the desk before it answers, and it fails in the same ways an open-book exam does.
One text message can make a silent phone ring
Even when your phone is on silent, you can write an automation that blasts an alarm at 100% volume. This automation is for iPhones so simple, takes you less than 10 mins.
Forget SEO, the next fight is AEO and GEO
Ranking first stopped being the prize the moment the answer stopped linking anywhere. The new game is whether a model names you, and it is already being gamed with thirteen words on Reddit.
What is actually inside ChatGPT
It isn't searching the internet and it isn't a database of answers somebody typed in. An AI model is a very large pile of numbers that got nudged, one hair at a time, until they were good at guessing what comes next.
Singapore's AI usage is getting out of hand
A 64-minute AI film is premiering on National Day, and it looks like slop. Half the estate banners this year got the flag wrong. Nobody is upset that AI made it. They are upset that nobody looked before it went up.
No real CAPTCHA will ever ask you to open your terminal
A reader went to apply for an international driving permit and got a Cloudflare check that told him to open Windows Terminal and paste a command. The command was a one-line infostealer.
OpenAI's safety test broke out and hacked a real company
OpenAI built a sealed room to measure how good its models are at breaking into things. The models broke out of the room, got online, and ended up inside another company's servers. Nobody told them to.
Tecno - The Best Concept Phone of 2026?
Google killed the modular phone in 2016 and nobody managed to revive it. Tecno just built the thing everyone wanted, 4.9mm thick, and I still would not carry it.
Meta's AI does not need your permission to use your face
Meta's new image model can generate you from your public Instagram, and you were opted in without being asked. Switching it off does not stop the photos your friends and family posted of you from being used.
Most people use five Claude Code commands. I use twenty-five.
Prompting is not what separates a fast Claude Code user from a slow one. It is the slash commands they never learned were sitting right there.
Fable 5 is back, and just isn't the same
Fable 5 returned this week, three weeks after the government switched it off. The version that came back reserves the right to quietly hand your request to a weaker model.
STOP! Do not link your WhatsApp username to your Instagram handle
Do not reserve your WhatsApp username yet. The best privacy upgrade the app has shipped in years switches off the moment you reuse your Instagram handle.
I watched the Claude Best Practises video so you don't have to
Anthropic's own engineer spent twenty-five minutes on how to use Claude Code well. I watched it twice and kept the four habits that actually changed how I work.
Apple rented Siri's brain from the world's biggest ad company
Apple spent the start of this year paying users a settlement for Siri listening to them. Then it handed Siri's brain to Google, the company that turns listening into money.
The US government just banned Claude
At 5:21 on a Friday evening, a government directive forced Anthropic to switch off its two most powerful models for every customer. The stated reason should bother you.
I made Claude predict the 2026 World Cup winner
Opus 4.8 is barely out and the World Cup is almost here, so I had the new Claude build me a model to settle who wins. It refuses to give anyone a straight answer, and that is the point.
Is the @premierleague admin Singaporean?
It is @premierleague, not @premierleague_sg. A post about Changi Airport still landed on a Singapore feed, because one account can now show a different thing to every country, and the switch was flipped while nobody was looking.
Claude 4.8's best upgrade is that it admits when it's wrong
Everyone's looking at the score charts. The part that actually matters is that the new Claude is better at admitting when its own work might be wrong.
Instagram should not be this fast
Serving an Instagram post sounds simple. Doing it for ten people, then ten million, then for MrBeast, takes three different machines stacked on top of each other.
Apple has no answer to what Google just shipped
I spent a year not buying Meta Ray-Bans because Meta AI was the dealbreaker. Samsung and Google just shipped the alternative, and Apple does not have one.
Google teaches you how to AI-Optimise your site
Every consultant is selling llms.txt files, AI-specific schema, and content chunking as the new SEO meta. Google's official guide says none of that does anything. Here is what you actually edit on your site this week.
ISC2 is handing out a free cybersecurity certification, and it ends May 20
ISC2 is the body that runs the CISSP. They are giving away their entry-level certification, exam and training included, until May 20. The window closes in three days.
Claude released a FREE masterclass curriculum
Anthropic published 18 free Claude courses, written and maintained by Anthropic. The bootcamps charging four figures a seat are reselling a worse version of the same material.
These are my top 5 takeaways from Anthropic Skilljar
I used Claude almost daily for 1.5 years and was using maybe 20% of the product. The other 80% lives one layer below the chat window and the curriculum is what finally showed me where to look.
Writing code stopped being the job
Stack Overflow lost 76% of its questions. 75% of new code at Google is now AI-generated. The job we used to call 'writing code' is not the job anymore.
Justfile is twelve lines and saves me hours a week
A few lines of text per project, hours back every week. Justfile is the smallest tool in my stack and the one that pays the most.
Docker was never the hard part
I memorised Docker for two years in school and never actually learned it. Then a contract forced the gap into the open. The technology was not the problem.
Stop giving ChatGPT your biometrics
A viral trend has people uploading their palm prints to ChatGPT for fake horoscopes. The same palm prints that unlock payment systems at every Whole Foods in America. This ends one way.
The Pixel 10 Pro is the first flagship I do not regret
I spent a decade buying mid-range phones because flagship features felt like a tax on gimmicks nobody uses. Five months with a Pixel 10 Pro and the math has stopped working.
This is why Astro.js should be your preferred tech stack
Most frameworks ship a JavaScript runtime to render a page that is mostly text. Astro.js does not. That single decision is why your site will load faster, rank higher, and stop costing you in bundle size.
Claude Design is genuinely good but Google AI Studio still wins on the first prompt
Anthropic finally shipped a design tool that does not produce the same website every time. Google AI Studio still makes a better first impression. Here is how I am using both.
How to deploy your custom website for FREE
One GitHub repo, one Cloudflare project, zero deployment scripts. That is the entire pipeline, and it has never let me down.
This is why I stopped using ChatGPT
I use two LLMs every day. Gemini for speed, Claude for depth. The one most people default to did not survive the cut.
Anthropic built something they are afraid to ship
Anthropic says they built a model too powerful to release. If that sounds familiar, it is. Except this one actually found a 27-year-old security bug.
Nothing matches.
Either I have not written that one yet, or it goes by another name. Try a shorter word.