- Please record your attendance for these meetings in the google doc linked under bookmarks in the #educate-ai-agents channel
- Brief report from Agent-building group, by Benjamin Wallisch
- Tried to build an agent that can reproduce a figure from a publication, harder than initially hoped
- Next steps
- Dig deeper into vision understanding for LLMs
- See if LLM can do something similar tot plot digitizer apps
- What frameworks and models are used?
- Mixed, some use LangChain or AI gen, some home-rolled thin python setups
- Benefit of those frameworks have memory functionality and can learn
- Models: Mistral via API, but also local smaller models (e.g. llama)
- Next meeting Monday May 11, 13:30
- Brief report from AI Discussion forum, by Renske Wierda:
- Last meeting Mon April 28: https://indico.fysik.su.se/event/9686/
- Topic was confirmation bias in AI experimentation
- Renske introduced topic, with clash between negative and positive confirmation bias about AI
- Examples where AI has increased efficiency in some situations, in others where it has not
- Goals of this families of meetings:
- Learn about agentic AI for scientific work
- Understand what it is useful for and what it is not useful for
- How do we keep track of successes and failures to get a balanced view of how helpful AI can be? General question, how to avoid only hearing about success stories.
- Next meeting May 19 at 15:30
- David explained idea about agentic AI for education, with focus on pedagogical methods using AI agents, and 4-5 papers that have done interesting studies of actual effects on learning when using study-buddy bots. Could start meetings on this, but likely not before the summer. Christian also interested in helping out to organize this.
- Christian mentioned new project with student-facing chatbot for administrative questions related to a study program. Works but fragile, and a bit slow. Would be interested in exchanging ideas and experiences with others who have built RAG-type bots (here using thin custom python around Gemma 4 E4B, with a chroma vector db, with basic web ui and a mattermost bot)
- Benjamin expanded on the idea to organize hackathons with people from AI tech scene in Stockholm, evening social activity with pizza and drinks. Nothing decided.
- Anna described the IceCube Kaggle challenge (2023), which was successful way of creating contacts between IceCube and AI people in industry
There are minutes attached to this event.
Show them.