Coming from monothematic in-person meetups we decided for a different approach for our first online meetup. Condensed deep dives into three different topics: Fine-tuning language models, a new geolocation tool for verification and code assistants for newsroom coders.
Fine-Tuning Language Models for Newsrooms
Badische Neueste Nachrichten’s Julius Sandmann explained how he fine-tuned GPT-4 Mini using articles that led to subscription completions, focusing specifically on the teasers of these articles. Sandmann highlighted the benefits of fine-tuning, such as task-specific customization and more efficient resource utilization. However, he also cautioned about potential risks, including overfitting and the complexity of training data.
Sandmann reported promising initial results from the fine-tuning process. The generated teasers outperformed those from standard models and were better tailored to regional interests. Despite the positive outcomes, he emphasized the importance of monitoring the fine-tuning process and continuously reviewing and adjusting the models.
AI-Supported Geolocation Tool “Spot”
Tilman Wagner and Lynn Khellaf from Deutsche Welle introduced “Spot,” an AI tool designed to assist journalists in verifying location information. The tool leverages OpenStreetMap data to identify locations in images and allows users to search for specific objects or combinations of objects using natural language commands. For example, a search query might be: “Find a parking lot within 200 meters of a broadcast tower and a power pole in Legnano, Italy.” “Spot” then displays all relevant locations on an OpenStreetMap.
“Spot” aims to simplify and enhance the efficiency of fact-checkers’ work. The developers acknowledged challenges such as incomplete OpenStreetMap data and the complexity of translating spatial relationships into distances. The tool is open-source and is under continuous development.
Code Assistants for App Development
Ippen Digital’s René Heuser and Stefan van Rest demonstrated how code assistants can accelerate app development. They showcased tools as Bolt, Vzero, and Cursor. Bolt and Vzero automatically generate code based on natural language input, allowing users to describe a specific app function in a prompt, which the tools then convert into code. This enables the creation of prototypes within minutes. However, according to van Rest, these code assistants are not (yet) suitable for programming more complex apps.
Cursor is an AI-supported development environment that helps programmers write code. For instance, developers can input their requirements into a chat window, and the tool generates the corresponding code, which the developer then reviews. Heuser and van Rest emphasized that these tools are intended to support, not replace, developers.
More Online Meetups to come
The participants’ feedback on the compact format of the online meetup was overwhelmingly positive, prompting plans to offer several more sessions in the coming year.
You can find a recording and the slides of the presentations here (password protected).
