SEBASTIAN DIEMER / A PERSONAL VIEW / SEPTEMBER 2026
AI / Human.
The machines got better. I finally caught up with my own ideas.
ADHD, curiosity, impatience, and about 200 ideas ahead of myself. With AI, that energy finally has somewhere useful to go.
THE SHORT VERSION / TL;DR
The idea is human. The crew is AI.
The things people found exhausting about working with me — impatience, parallel ideas, constant microfeedback — can be useful in a well-run AI workflow. The basics still matter: clear instructions, separate tasks, one owner. The fit feels a little like 1 + 1 = 6. Or 7, on a good day.
01 / A QUICK OPERATING MANUAL
First, the human.
I’m Sebastian. This is roughly what you’re dealing with.
Curious
My default question is “what happens if we try this?”
Creative
Ideas come easily. Quite a few are worth trying. All at once, ideally.
Impatient
I can see the result in my head. Waiting for it is the difficult part.
All at once
One interesting idea has a habit of becoming six open projects.
Very particular
Quality matters to me. Yes, including those two pixels.
Microfeedback
“Almost. Change this bit.” I have been known to say it a lot.
My personality type: Entrepreneur. Curious about yours? Find your personality type on DEEPDIEMER ↗
THE PART THAT USED TO DRIVE ME MAD
I could see it. I couldn’t just build it.
My problem was never coming up with ideas. It was getting the thing in my head into the world without losing half of it on the way.
Whether it was Kreditech and better lending decisions with machine learning, or prediction markets at Wallfair: there was always a gap between what I could imagine and what I could actually try.
The feedback from people I worked with was fairly consistent: too impatient, too many things at once, too much microfeedback. I understand it. I was already changing the next detail while somebody was still working through the first one. I frustrated people. Sometimes I simply annoyed them. And the project still didn’t turn out as good as I had imagined.
Now I can say what I mean, look at what gets built, and immediately say, “Almost. But this bit.” That sounds small. To me, it is enormous.
WHY IT FITS / MY PERSONAL LEARNING
The same personality. A much better fit.
What made collaboration difficult can become useful input. Provided I organise it properly.
ME + AI / A VERY GOOD FIT
One curious human. A crew of AI agents.
My very unscientific maths. Rosi is still counting on her paws.
WHAT CARRIES OVER FROM PEOPLE
A bad brief is still a bad brief.
- Vague instructions give somebody a lot of room to build the wrong thing.
- Give two people — or two agents — the same task without coordination, and you get duplicate work or a collision.
- If nobody owns the whole result, activity can look very impressive while the project goes nowhere.
WHAT FEELS DIFFERENT FOR ME
The next correction is just input.
- An agent doesn’t take “please change this again” personally. I can be precise without another awkward conversation.
- Several clearly separated tasks can run in parallel. My curiosity has somewhere to go.
- Fast, specific feedback helps improve the result. My habit of noticing every little thing finally has a very good use.
AI still has limits. Too much context, contradictory instructions and messy coordination still cause problems. The fit is in the feedback loop, not in pretending structure no longer matters.
02 / HOW I THINK ABOUT THE SHIFT
We’ve automated things before. This feels different to use.
A few steps that help explain it. The technologies overlap; the point is what I can actually do with them.
- 01
AUTOMATION
A specific job.
Chess computers searched for good moves. Autopilots helped fly a plane. Impressive automation, inside a carefully defined job.
- 02
MACHINE LEARNING
Learn the pattern.
Machine learning made predictions from data. That was already part of my world when I wanted to improve credit decisions.
- 03
CHATBOTS
Have a conversation.
Then the interface became a chat. Ask a question, write a text, put a document together. Suddenly a lot more people could use it.
- 04
MY CURRENT SETUP
Actually build the thing.
Now my agents can write and run code, work in local consoles, use connected tools and divide a project between them. The conversation turns into working software.
AN ERA OF ASTRA / CODEX
The tools became a crew.
For me, the jump is a console on my computer with several agents working through a real project. One builds. Another checks. Tools get connected. Things run. I give feedback by voice and watch an idea become something I can actually open.
I’m taking on things alone that, in my earlier startup life, I would have needed a serious team and a lot of funding to even attempt. That is my experience of the change. It still makes me slightly giddy.
But I do not say “make something amazing” and disappear to the beach. I brief. I inspect. I interrupt. I get absurdly specific. The models still make mistakes, forget a constraint or confidently improve the one thing I explicitly liked.
My creativity suddenly goes a long way. So does my capacity to notice that something is not quite right.
That way of building now also reaches practical searches: Praxiskaufen.com brings practice offers into one place, with price context and a search agent. The next step is Handwerkbörse.com, currently a public preview applying the same idea to business succession in the trades.
03 / WHAT I ACTUALLY DO ALL DAY
A very capable crew. Still a human at the controls.
Make it clear.
The idea, the reason, the details that matter.
Build. Check. Connect.
Small tasks, clear roles, a result I can try.
“Almost. This bit.”
Look closely. Give precise feedback. Repeat.
What do I mean by microfeedback?
“Make it nicer” leaves a lot to interpretation. “On my phone, keep the button inside the first screen, make Rosi larger, and leave the wording alone” gives the next agent something concrete to do. Then I check the phone. Then I check what changed elsewhere. That loop is a huge part of the work.
A lot of this is learning where I still need to steer, and where the agents can take over. My 10 current learnings ↗
THE SMALLEST POSSIBLE FOUNDING TEAM
One Diemer. One Rosi. Let’s see what happens.
That is enough to get going now. I no longer need to assemble a whole company before I can find out whether an idea is any good. What a ridiculous, wonderful change.
It doesn’t mean the door is closed to people. Quite the opposite. I like building with people who care. Rosi GPT is a community project: the people talking to her should help shape who she becomes.
Rosi is about company. DEEPDIEMER is about understanding yourself. That’s where my curiosity keeps landing: making AI feel more personal, and finding out what people actually value in it.
THAT’S WHERE I AM RIGHT NOW.
Less distance between “what if” and “here it is.”
I’ll keep building, getting carried away, checking the details, and occasionally remembering to go outside. Rosi insists.
Paw-signed by the original.