The idea sounded simple at first. You tell the app what interests you: politics, photography, travel, design or sport. Blogbox finds suitable articles and turns them into your magazine.
RSS readers worked differently at the time. You had to know the blogs already and subscribe to each one. We wanted to start with topics, not sources. That meant Blogbox had to classify every new article automatically.
A feed knew that a new text existed. It did not know what the text was about.
Two ideas for sorting
We had two approaches. The first used ontologies: prepared trees of meaning containing keywords and the relationships between them. An article could receive several topics and appear in different groups.
We worked closely and very successfully with TopicZoom on this route. Their system was remarkably good for its time. We sent it an article's headline and text and received weighted topics. Our backend used those weights to build the app's sections.
An article about crossing the Alps could appear under travel, cycling and adventure at the same time. In the end, TopicZoom was the system that worked reliably in everyday use.
Our machine-learning attempt
In parallel, we experimented with machine learning ourselves. The idea was to let a system learn from examples instead of following only a prepared tree of meanings. We tried existing algorithms and developed approaches of our own.
Some of it looked good in experiments. Everyday operation was harder. New articles arrived each day, written in very different ways. Errors could not be allowed to spread unnoticed through the magazine.
We tried to turn the work into a system that would run reliably every day. We did not finish it. TopicZoom remained the solution we could depend on. That is part of the Blogbox story too.
What today's AI changes
Machine learning has moved much further with current AI systems. Language models deal with context and ambiguity better than the methods we had. They can classify and summarise text and connect related articles.
A new Blogbox would probably combine both approaches. An ontology could enforce stable editorial rules, while a learned model could handle new and ambiguous material.
Blogbox was not an early AI app. But it asked the same question: how much does a machine need to understand about a text before it can produce a useful selection?
What reached the app
Users did not see weights or meaning trees. They chose topics. The source remained visible on every article, but it no longer determined the magazine's structure. Each front page was slightly different.
It still required editorial work. We reviewed blogs, added fixed assignments and let readers report unsuitable posts. The machine sorted the articles. We still decided which sources were allowed into the app at all.
Blogbox launched on the iPad. The iPhone followed, and Android arrived in 2014.
Not a newspaper you zoom into
Many newspaper apps were PDF readers at the time. They showed the printed page and let you zoom in. That preserved the edition's hierarchy and layout.
We wanted something else. Text should adapt to the screen. Type, images and navigation belonged directly in the app. It was much easier to read on iPhone and iPad.
Why we shut the app down
At the same time, we used the technology for magazines made for publishers and companies. Our focus moved there. It was clearer who bought the product and which problem it solved.
The app needed constant maintenance. Blogs changed their feeds, websites changed their HTML and operating systems changed their requirements. We never found a business model for a free magazine app that could cover that work over the long term.
Eventually we shut down the backend. That also ended the app, because its content was collected and sorted again with every request.
What remains
When I look back at Blogbox, two things remain. Automatic classification has to work in everyday conditions. And digital content needs a form that fits the device.
Blogbox is history. I still find its two questions interesting: How does a machine understand a text? And how does that become something people enjoy reading?