Blog
22 posts
All posts
I Gave GPT-6 Astra and Claude Fable 5.1 the Same Building Brief and Let Them Choose Everything Else. Both Held the Dimensions to the Metre and Shipped Two Different Tools.
Six fixed dimensions, four things I wanted to do in a browser, and no other constraints. Astra took 6 minutes 50 seconds and 21,385 characters. Fable 5.1 took about 21 minutes and 51,165. Both put every dimension exactly where I asked. Then one of them added buttons that teleport you to each floor at eye height, and the other built a dog-leg stair you have to climb. The first draft of this test was mine, and it was worthless: I banned libraries and materials, both scored 6 out of 6, and the test measured nothing.
I Drew My Own Floor Plan So I Would Know the Right Answers, Then Erased the Dimensions and Asked AI to Measure It
Floor plan tools promise the AI 'works within your dimensions, not a generic template.' I drew an 8.0 by 6.0 metre flat myself, so I know where every wall sits to the millimetre. With the dimensions printed on the drawing, ChatGPT got all of it right and named the 1:100 note and the scale bar as its source. With the numbers erased and only that bar left, it measured the rooms to within about four per cent. Feed it the identical file twice and the wall thickness comes back 150 mm, then 160 mm. And a dedicated floor plan tool put a window in a bathroom that does not have one.
I Redesigned the Same Room Seven Times on a Free ChatGPT Account. It Never Hit a Limit, and It Only Broke Once.
Seven edits in one chat, on the free plan, no upgrade prompt and no cap. The room survived a style change, a floor change, a furniture swap and a night scene. It fell apart on the one request people actually want most: show me the same room from another angle. Every prompt is printed so you can repeat it.
AI Interior Design Prompts: Exactly What to Type to Change the Floor, the Ceiling, the Walls or the Sofa
Eight copy-and-edit prompts for testing an interior change before you spend money on it. One for swapping a sofa, one for the flooring, one for the ceiling, one for wall colour, one for lighting, and three more. Each one was run on the same room so you can see what it actually did, including the three that quietly changed things I did not ask them to.
I Gave Free Gemini Three Public-Domain Black-and-White Photographs to Colorize. Two Came Back in Colour, One Was Refused Twice, and Both Successes Arrived With a Watermark and a Quarter of the Pixels.
Same prompt, same free account, three old photographs. Einstein at his blackboard and the Hindenburg burning came back colorized and convincing. Dorothea Lange's Migrant Mother came back with 'I don't seem to have access to that content', twice, with the chat auto-titled Colorization Request Denied. Both successful files carry a Gemini sparkle in the bottom right corner, and a 2523 by 3313 original returned as 896 by 1177.
I Gave ChatGPT the Same Grey Massing Model Five Times. It Never Wrecked My Geometry, and It Broke Something Different on Almost Every Run.
One 74-character prompt gave me a photograph of a study model. It had no windows, nothing around it, and nothing in frame to tell you whether the building was five storeys or fifty. Four longer prompts turned it into a building. Across all five runs the massing survived, but a cantilever flattened into a roof on one run and the timber jumped to the wrong volume on another. Every prompt is printed in full so you can repeat it.
Three AI Stories From This Week, Read Against the Primary Documents: Grok 4.6 Doubles Its Price at 200k Tokens, Qwen3.8's License Bans No Country, and Claude's New Watermark Says It Is Not Conclusive.
I took three stories from the past 48 hours and opened the document each one is built on. xAI's launch post quotes $2 per million tokens; its pricing table doubles the rate past 200k and bills the whole request at the higher rate. A widely repeated claim says Qwen3.8's license forbids use in the US, EU, UK, and Korea; the license is sixteen lines long and contains no territory clause at all, while a different model's license does contain exactly that list. Anthropic's new watermarking is real, and Anthropic's own support page says a detected mark is not fully conclusive and a missing mark proves nothing.
ChatGPT Ads Reached Korea on August 11. I Opened a Korean Free Account the Next Day: No Ads Yet, But the Ad Settings Are Already There and Both Switches Are On.
OpenAI's own page says ChatGPT Ads launched in South Korea on August 11. I logged into a Korean free-tier account on August 12 and looked for them. Two shopping questions produced product carousels with won prices and star ratings, and not one thing on screen was labeled as an ad. The ad machinery is installed though. It sits under Data controls, it has two personalization switches, and both were already turned on before I touched anything. The documented free-tier opt-out was the one thing I could not find.
ChatGPT vs Claude on the Free Plan: I Asked Both for 3D Scenes. The Second One Only Rendered on One Side.
Two prompts, two free accounts, same day. The first task was a cube in a room and both delivered something that runs, though only one of them survived losing the network. The second task was a solar system, and here the result was not close: one produced planets, orbits, an asteroid belt and a working camera, and the other produced a loading spinner that never stops. The cause was not model quality in any vague sense. It was one URL, and I can show you the exact line.
Three AI Launches in One Week. I Tried All Three Without Paying, and 'Free' Meant Three Different Things.
OpenAI moved free users to GPT-5.6 Luna, xAI shipped Grok Imagine Image 2.0, and Alibaba released Qwen3.8-Max. All three were announced as things you can use at no cost. I opened all three the way a new user would. One gave me a working model but will not tell me which one. One made me sign up, then generated images fine, then put its headline feature behind a subscription. One answered me without an account and printed its model name in the header. Here is what each one actually does when you have not paid.
I Automated the AI Approval Prompt. It Missed the Same Three Commands in Every Session.
A developer published 409,000 approve/deny decisions from a browser game and reported that the average player misses 1 in 3 threats. The obvious next move is to stop asking the human. So I wrote a twenty-line filter, gave it the study's own threat taxonomy as an answer key, and ran it three times. It caught 29 of 46 threats — and the ones it waved through were the same ordinary-looking commands every session. Here are all three scorecards, the full decision log, and the one substitution that beat my filter outright.
I Checked My Claude Code Setup for Shai-Hulud. Then I Found Three of My Four Package Counts Were Wrong.
The reporting says the payload wakes up the next time someone opens the repo in VS Code or starts a Claude Code session inside it. So I ran the check on my own machine: 35 config and hook files hashed, 351 installed packages against a 79-entry list, zero matches. Then I reopened all four sources to quote them properly and found that three of the four counts in my own notes were misattributed. Here are the commands, the real output, and why a zero is not a clean bill of health.
Gemini Made Me a Video for Free. First It Asked for My Gmail.
Google's free video promotion came with a number attached: ten clips. I went looking for that counter on the screens I used and never found it. What I did find was two dialogs standing between the sidebar and the prompt box, the second one asking to connect Gmail, Calendar, and Photos. I said no, and the video generated anyway.
I Used Google Slides AI on a Free Account: 10 Minutes, 3 Slides
Google lists AI deck generation in Slides as a Business Standard or Google AI Pro feature. On my free personal account it built a three-slide deck anyway. Here's the path I took, the six minutes and forty seconds of build time, the questionnaire it ran me through first, and the prompt that got refused.
The EU's AI Labeling Rule Starts August 2 — and Most of It Isn't Your Job
Article 50 takes effect on August 2. The machine-readable watermark isn't your obligation, your diagrams aren't deepfakes, and AI-assisted writing you edited is exempt as written. Which half of the rule is actually yours, read from the statute text.
Why AI-Built Websites Look Fake: 9 Tells I Found on the Page Claude Built Me
In Part 1, one message to Claude produced a finished-looking SaaS landing page — with 12,000 invented users. Before fixing anything, I went through that exact page top to bottom and named every tell that gives it away. All nine, with screenshots.
Opus 5 vs Fable 5: The Shorter Answer Cost 28% More Tokens
I gave Opus 5 and Fable 5 the same one-line prompt: a self-playing neon brick game in one HTML file. The smaller answer cost 28% more. Here's every number.
Landing Page Anatomy: Why Every SaaS Site Has the Same 9 Sections — Dissected on the Page an AI Built Me
I can't read code, but I own a landing page an AI wrote. So I learned the first real web-design lesson: what the sections of a landing page are, why they're always in that order, and how knowing their names lets you command an AI with surgical precision. First structural edit included.
What Is Kimi K3? The Chinese AI That Just Out-Coded GPT and Claude — Explained Simply
A Chinese lab dropped an AI model last week that topped a coding leaderboard above Claude and GPT — then got so popular it stopped letting new people sign up. Here's what Kimi K3 actually is, why everyone's losing it, and whether you can use it right now (in plain English).
I Asked Claude to Build My SaaS Landing Page — No Code, No Design, Just One Message
I can't code and I can't design, so I typed one sentence into an AI chat and asked it to build a landing page for an app I made up. It handed me a full, dark-mode startup page — including 12,000 users and a 4.9-star rating for an app that has never existed. Here's exactly what I typed and what came back.
Claude Fable 5 and Sonnet 5: What Actually Changed (and Which One to Use)
Anthropic renamed its top tier and shipped a new mid-tier in the same month. Here's what's real, what's marketing, and the one pricing trap nobody's flagging.
Gamma Review: I Generated a Real Deck on the Free Plan — Here's the Honest Verdict
I ran a real presentation through Gamma's free tier — one prompt, ten slides — then came back 47 days later to check what had changed. It fills blank data with confident numbers of its own, and its own export screen warns that 3 of my 10 slides will shift in PowerPoint. Here's the honest review, with my own screenshots and pricing re-checked in August.