Let me be completely honest: full tech, and then you are back among people. You step out of the workshop, you stand at a reception or a conference, and someone asks: “So, what do you actually do?” And you realise you have no answer that fits in one sentence. I was not out socially for a long time. I was in the engine room.
In the engine room
The last few years consisted of building. Training models, setting up agents, getting systems running at customers, one company in Zurich, one in Dubai. If you do that seriously, little time is left for everything else. And honestly, that suited me.
In the engine room things are clear. Something works or it does not. A model has got better or it has not. You do not have to convince anyone, you only have to measure.
Outside it is different. Outside you have to explain. And explaining is a different skill from building.
What you forget along the way
I noticed that I had forgotten how to tell. Not how to talk. I could always talk. But telling, in a way that someone who was not there understands what it is about.
When I tell a colleague "the QA must not be the agent itself", he nods. When I say that to someone who does not work with agents, he looks at me and waits for the rest. And the rest is a story. How it started, what went wrong, what we made of it.
Why I do it anyway
Because otherwise what we learned stays with us.
Fact is: in the last few years we made many mistakes that others are only just making now. Agents that sign off on their own work. Models that were believed too much. Costs that nobody calculated. Teams that got a tool and no rules. None of that is written down anywhere, except in our heads and in a few internal documents.
And I see companies now going down the same paths. Not because they are stupid. But because nobody told them where the holes are.
That is the reason for this blog. Not visibility, even if that comes with it. But the feeling that you know something that someone could use right now.
What is hard
What is hard is the question “what do you actually do?”. If I say we build AI systems, the conversation is over, because everyone says that. If I say we train fifty small models and let agents work against contracts, it is over too, because nobody understands it. The answer in between, one sentence long and true, I did not have. I only worked it out over the last few months.
What is also hard is finding time for it. An article like this costs me half a day that I would otherwise have spent with a customer or a model. At first I counted that as a loss.
What is hard is showing yourself. When I write that we believed the models too much in the beginning, then it is there. With my name on it. In the engine room nobody sees that. Here they do.
And what is hard is being in conversations again that have no goal. Conferences, dinners, people you have not seen for a long time. I noticed that I have become impatient. That in every conversation I look for the point where something could be built. I have to put that down again.
What I am learning right now
That telling is not self-promotion if you do it right. It is more like documentation, only for people.
That the people who give back the most are often the ones you meet by chance at a conference. Not the ones you had an appointment with.
And that the time I put in here does not come out of building. It comes back, as better questions from people who read what we learned.
When a techie can no longer talk to a non-IT person, that is not the non-IT person's problem. It is ours. I am not good at it yet. But I am doing it. And if you meet me somewhere and I get to the point too quickly: Tell me. I am still practising.
Written by me. The thoughts, the values, the learnings, the mistakes: all mine. Grammar and spelling are corrected by our own twin model, trained on my texts. Sometimes a stumble stays in. That is mine too.