Hi there,

Three weeks again. Last time I said shorter gaps and shorter emails were probably the sustainable shape of this thing, and then I went and published five guides, so the "shorter email" part is going to need some work.

One of them matters more than the others, and there's a decent chance it's about you.

🔒 Your company blocked ChatGPT. Now what?

Almost every piece of AI advice out there quietly assumes you get to choose your tool. Open Claude, make an account, start using it on real work. That's fine if you're a freelancer or a founder. For most employed people it's useless, because most employed people don't choose.

You know the situation if you're in it. ChatGPT and Claude are blocked on the work network. You have a Copilot license nobody explained, or an internal chatbot the company bought a platform for and half-built. You sat through an intranet course on ethical AI use, maybe a second one on basic prompting. And somewhere in the building there's a person whose job title contains "AI" and whose actual mandate nobody could describe to you, including possibly them.

I finally wrote the guide for that: Your Company Blocked ChatGPT. Now What? The short version, in case you never click:

The block is usually the right call. I know that's not what you want to hear. But a free consumer account isn't built to hold a client contract or a patient record, your inputs can feed model training depending on plan and settings, and there's no data processing agreement behind any of it. Your IT department isn't being paranoid. Where it goes wrong is what happens next: block the risky tools, buy a sanctioned platform, run a mandatory training, quietly check the box marked "AI: handled." Nobody ever comes back to ask whether anyone actually got good at it.

Look at what you actually got before writing it off. Copilot inside Microsoft 365 runs a frontier-class model. My complaints about it are real, but they're about the packaging: it's less configurable than ChatGPT, it doesn't learn about you as cleverly, and it interacts with other software surprisingly badly, including Microsoft's own. "The model is dumb" isn't on the list. It's catching up, by the way - used to be 2 years behind cutting edge, now probably less than 1 year. And the tool you're allowed to use on real client work beats the better tool you're not allowed to touch. That's the entire point of it existing.

Run two tracks. Build the actual skill privately, on your own material: the trip, the cover letter, the rental agreement, the hobby project. Then take the same habits into whatever your company sanctioned. Briefing properly, iterating instead of expecting a perfect first answer, checking the output instead of trusting it. None of that is tool-specific. The skill transfers completely even though the data never crosses.

The one hard line: never push real work data through your private account to get around a limitation. If the sanctioned tool genuinely can't do something you need, that's feedback your company needs to hear, in specific terms. Quietly working around the policy just hides the gap and adds risk nobody agreed to carry.

One last thing from the guide, because it's the piece people skip: go find your AI champion and bring them one specific, boring task from your own work. A report you write every month. An email you send too often. Most of these people have a mandate and almost no support, and a real example from someone in the building is worth more to them than another workshop slide.

🧠 The one that made me uncomfortable

I've been in meetings where someone shares a document, I ask why they framed something a certain way, and they don't know. Because they didn't write it. The AI wrote it, they skimmed it, they forwarded it, and the moment you ask questions past the first paragraph there's nobody home.

I recognize it because I used to do it. Read the opening, read the ending, decide both are solid, ship the whole thing. Then find something buried in the middle weeks later that I'd never have signed off on. The AI hadn't lied to me. I just hadn't done my job.

There's research on this now, and it's the reason I wrote AI Makes You More of Whatever You Already Are. A team at the MIT Media Lab put EEG caps on 54 people writing essays: one group with ChatGPT, one with a search engine, one with nothing. The ChatGPT group showed the weakest brain connectivity, felt the least ownership of their own work, and many of them couldn't quote a single line from an essay they'd finished minutes earlier.

Then they ran a fourth session and swapped everyone. The people who'd been thinking unaided and then got AI? Brain activity jumped, recall improved, AI made them better. The people who'd leaned on AI first and then had it removed stayed checked out. The habit of not thinking had already set in.

Fair warning, and I put this in the guide too: it's a preprint, the sample was small, it was students writing essays under a timer, and the lead researcher herself pushed back on the headlines with "we didn't find any brain rot." Don't quote it as gospel. Treat it as an honest hint that lines up with something a lot of us have felt.

What I actually believe is simpler than the study. AI doesn't change your character, it multiplies it. If you've always looked for ways to hand in something technically finished, AI is the best shortcut ever invented. If you genuinely want your work to be good, it's the biggest lever you've ever had. Same tool, opposite outcomes, and the variable isn't the model - it's all you!

🎨 Three guides on making images that survive contact with real work

Every image on Good AI Guide is generated. The hero art on the guides, the category illustrations, the two-company race picture from last edition's game. So the series I wrote on AI images is just what I actually do, including the parts that waste an hour.

From an idea to an image worth using — start from the job the image has to do, not a pile of style adjectives. Let an LLM interview you into a proper visual brief, then generate. Treat the first result as a draft, diagnose the single biggest problem, change one thing. That last habit is what stops prompt roulette.

Edit the image, keep what works — the hard part of editing is telling the model what it's not allowed to reinterpret. Describe the change and the invariants separately, make one meaningful edit per turn, and branch from a clean source, because drift accumulates even when every instruction says to preserve the original.

Put generated images to work — the production version: approved facts and copy first so the model renders decisions instead of quietly inventing business claims, when to leave ChatGPT for the API, what to automate, and a named human between generation and publication. Not optional, that last one.

They read as a series and they're long. If you only want one, take the first.

🇫🇷 The French edition is actually finished

Last edition I told you French had "just launched with the first guides translated and the rest following." That sentence was doing a lot of work. What existed was a French homepage, three guides, and a polite note on everything else saying the translation was coming.

It's done now. French and German both carry all 143 entries: 35 guides, 42 tool reviews, 25 model entries, 26 recommended reads, 15 people worth following, plus the quiz, the checklists, the certificates and every page's chrome. A script tells me when an English edit has left a translation stale, and it currently reports zero in both languages. This would have been nearly impossible for me to achieve just 1-2 years ago.

How it got done is quite interesting. Every single file went to its own AI agent, and each one started cold: no memory of the file before it, no accumulated taste, nothing but the English source and a rulebook. The rulebook is where all the actual thinking is. It tells the agent that French keeps fewer English words than German does, because a German professional will happily say "Tool" and "Feature" in a sentence while a French one says «outil» and «fonctionnalité». An agent that doesn't know that writes French a native speaker can smell in two sentences.

They also run on a mid-tier model (Sonnet) on purpose. I ran a blind comparison against the expensive one (Opus) during the German phase and genuinely couldn't tell which was which, so the cheap fast one won and everything since has gone through it. The judgment lives in the rulebook, not the model. And the rulebook compounds: every correction that could plausibly happen again becomes a glossary row, so a given mistake gets made exactly once. My actual job was "just" three things. Spot-check each batch for consistency, settle the calls the agents flagged instead of quietly inventing an answer to (should «subagent» stay English in French the way it does in German? no, «sous-agent», because French borrows less), and put a native speaker in front of the result. That last review came back with nothing systemic to fix, which is the only reason I'm comfortable telling you it's finished.

📇 Three smaller things

Loops, graphs, and who decides what the AI does next — people keep arguing about these two words when they talk about agents. A loop is an agent that picks its own next step until it decides it's done. A graph breaks the work into fixed steps with checkpoints you control. You're trading autonomy for control, and most people firing off one-off chats never need to care. I'll say plainly what I said in the guide: I haven't built a production graph yet, and Anthropic itself has watched teams spend months on elaborate multi-agent setups that a single better-prompted agent then beat. Start simple.

Your next reader is a machine — I've been meaning to send you this one for two editions. Bots overtook humans as the majority of internet traffic while most of us kept building exclusively for people. Getting a machine's attention works nothing like the SEO you know, and it's a large part of why this site is structured the way it is.

And a small change with a lot of thinking behind it: the homepage used to say this site was for builders and leaders. It now says it's for the curious. Curiosity is the entry requirement here. Prior AI knowledge isn't, and pretending otherwise was possibly quietly turning away exactly the people I most want to reach.

🎮 One more game, and this one's hiding

I replaced the little easter egg in the site footer. It got even nerdier! I vibe-coded it with Claude Fable 5 with just 4 prompts, 3 of which to tweak the game's difficulty. It's now Gradient Descent, and it teaches the one idea underneath every AI model you've ever used: measure how wrong you are, take a small step in whichever direction reduces the error, repeat a few million times.

You are the optimizer. You roll a model down a loss landscape and try to find the bottom. Six short levels, each one a real training problem: local minima that trap you, plateaus where the gradient goes flat, a learning rate that explodes if you get greedy, noisy data. Five to fifteen minutes, runs entirely in your browser, English only like the other one.

Nobody needs to know how training works to use AI well. But if you've read the phrase "the model was trained" a hundred times and never had a feel for what that means physically, this will give you one in about ten minutes.

That's it for today. If you're in the blocked-ChatGPT situation, I'd genuinely like to hear what your company actually gave you and how it's going, because that guide is going to keep growing and the best material for it comes from replies to these emails.

Best,
Fabian