Matthias Fiedler is a colleague of mine at SPIEGEL – he covers sports from Munich. In his newsletter StoryCodes, he gives journalists tips on AI, and occasionally he interviews a guest. And this time I got to be the one! Since his newsletter is published in German, here’s my English translation:

GPT-5. When the mechanics click and the big thinking model fires up, writes some code or outlines a topic.

Claude Sonnet 4. I use it as my first reader and sparring partner. Does what it’s supposed to, doesn’t get weirdly chatty on me.

Google Colab with Gemini. Build small Python applications, run them in the browser, fix errors directly with AI.

I let headlines lead me astray. Classic. When you read breathless reports about studies showing what AI can’t do or has spectacularly failed at again, always look at the methodology: research often lags months behind current development. What held true for GPT-3.5 back in 2022 may not apply to GPT-5 anymore.

There was this recent hope that AI couldn’t be funny. During new employee onboarding the other day, I put GPT-5 to the test: “What can you say to new employees but not in a relationship?” A question from that “wrong but funny” comedy format that’s big on TikTok. The answer: “The probation period is six months.” Everyone can decide for themselves who comes off worse here: my humor, ChatGPT, or German comedy.

The media company Every doesn’t just publish good AI articles, they’re also building a text editor called Lex. Pretty interesting. What I’m still missing: AI autocomplete. When I’m typing, the machine should semantically search through my files and texts and serve up suggestions. Programmers have it better – their dev tools have had this forever. You can test it somewhat by using an AI code editor like Cursor as a word processor, creating a project and copying in your own texts or pulling them via MCP.

“Dumb” is an anthropomorphization that doesn’t really help, right? World knowledge compressed into a chatbot – that’s anything but dumb. If the question is what these probability calculators aren’t suitable for, I’d say: replacing humans. The providers train emotional simulations into the models that fake understanding disturbingly well. People are even forming parasocial relationships with LLMs. We should really knock that off. It’s like the developers watched the movie “Her” and thought: Perfect, that’s exactly how we’ll build this thing. That film is from 2013 and it’s a damn dystopia.

I still haven’t gotten into all these “second brain” tools. Fabric, Kortex, Mymind, whatever they’re all called – note-taking with AI assistance. I try them out, but I get more out of clicking through bookmarks myself, thinking, making connections. But hey, to each their own. Maybe I’ll come around eventually.

Get excited. Seriously, I mess around with this stuff, I get outputs of varying quality, but when I’m trying to get ChatGPT to be creative, unexpected, fantastical and it actually works, that’s a pleasant surprise. Doesn’t happen often. These machines produce average by default – they’re boring and predictable by design, plus artificially constrained on top of that.

Proofread texts for simple errors. I used to be a copy editor.

For me it works the other way around.

Error messages from programs. I copy them straight into a chatbot.

Facts.

Everything around AGI, the supposed superintelligence. It’s a marketing gimmick, and nobody knows what it actually means. It’s on my list of 10 facts journalists should know about AI.

Personal: I wanted a minimalist website and retrofitted a WordPress. With Claude’s help: different types of entries, custom sorting, a button for summaries.

Follow the data: If the AI lives in the cloud (now I’m talking like that too), then the providers can theoretically read the inputs. Often they’re stored for future models. Authorities or hackers could try to intercept them too. Where’s the company based, where are the servers? You can also run some AI models locally on your own computer – a three-year-old MacBook Air and a program like Ollama are enough.

Ask the model itself how the prompt could deliver better results. The newer models are pretty good at this now. For GPT-5, there’s an official prompt optimizer.

I created an ambitious marathon training plan with GPT-5 and discuss my training sessions with it. It’s a dumb idea, but I tell myself I can assess what I’m doing. Whether it works, we’ll see (hopefully) on September 21st in Berlin.

Been a while, but I saw someone sketch out their plan for coding first, then ask: Understood? Logical? Obstacles? Unclear points? Works for data analysis or document work too. First check if the data’s even readable. Ask something you already know the answer to.

It’s a cliché from the AI world, but: garbage in, garbage out. If you have good thoughts, you can extract more with AI. If you lazily type short instructions, you’ll get mediocre results at best.

Ask more questions! We’re experiencing unprecedented hype around an almost eerie technology – everyone should be asking everything all the time.

I like AI when it challenges me, when it forces me to be better, more creative, smarter. Not as a shortcut.

Math problems, which AI then at least solves cleanly with a Python script. But that’s the kind of laziness that leads straight to getting dumber.

Francine Sucgang knows her way around AI, data, language – and just like I used to be a “digital native,” she’s part of the “AI natives” generation.