My Journey with AI So Far

I have been working with AI since 2018 — which, at the time of writing, means for 7 years. In this post I'll share some of my learnings, observations, and reflections.

Simple chatbots

In 2018 I started working on chatbots almost by coincidence. It began as a way to solve a customer problem and improve user experience. This turned into a startup called CBOT, where we built chatbots for events.

The systems were very simple, relying on keyword matching and basic embeddings. Still, I quickly learned that the chat interface is incredibly intuitive for people. Users would often move beyond event-related questions and start asking about all sorts of other things.

Playing with transformers

In late 2019, when OpenAI released GPT-2, I was super excited to try it out. Using Write With Transformer, I prompted it to tell me more about myself. To my big surprise, I had actually made it into the training data. It managed to respond with fairly accurate details about my entrepreneurial personality and love for computers.

Fast forward to 2020, during my master’s thesis, we worked on building a system that could understand medical notes. The goal was to extract named entities such as diagnoses, personal information, and lab results. At that time, just 5 years ago, BERT (based on the now-famous Attention Is All You Need paper) was considered state of the art. The results were decent, but came with a huge caveat: extensive data labeling, which was both slow and painful.

What I sensed back then was simple: the bigger the models, and the more quality data you give them, the better they perform. In hindsight, I was essentially stumbling upon what became known as the Scaling Laws for Neural Language Models — just without the evidence or data to back it up.

Machine learning systems in production

At my first two industry jobs, at KMD and Electricity Maps, I spent a lot of time working with machine learning systems from a product development perspective. What I learned was how hard it is to make ML systems work at scale.

I saw the importance of proper tooling when deploying hundreds of models. I also saw how critical good training and test data are. And perhaps most importantly, I learned how much customers value simplicity: making complex ML systems appear simple, easy to use, and wrapped in a clean interface is what makes them succeed.

AI-assisted work in 2025

Since the end of 2024, I’ve been using AI tools extensively for coding. It’s been a great way to explore both the capabilities and the limitations of the most recent models. With the right context, they can do amazing things. But they can also fail measurably at the most basic tasks — like deciding which number is larger: 9.11 or 9.9.

I’ve also started using AI as a sparring partner for business-related ideas. It still tends to be a bit too agreeable — a kind of “people-pleaser” — but I’ve already had some great insights.

And finally, this blogpost itself has partly been written by AI, based on a combination of detailed bullets and a bit of personal context engineering.

Summary & What’s Next

Over the past 7 years, I’ve watched AI go from struggling to understand a simple query about where to park a car, to now being able to tackle advanced coding problems and even give thoughtful feedback on business challenges.

I won’t try to make any bold predictions. But I do feel increasingly confident that there is a good chance that AI will be for knowledge work what the Industrial Revolution was for physical labor.