Ontology, open-weight models, and AI’s limits

in our fourth session, Pooria, Omid, Soheil, and i discussed a mix of random topics, including ontology, the limitations of AI, open-weight models, and AI’s wider impact beyond software engineering. i very quickly mentioned ontology as a way of turning a team’s shared domain knowledge into concepts, relationships, and rules that machines can understand. this actually came up after Pooria raised the question of what software engineers should focus on and which learning paths will still be valuable in an AI-driven world.

I also mentioned the importance of open-weight models. AI agents are exciting and powerful, but most of them are built on closed models and controlled platforms. with Open-weight models we can have more control, transparency, customisation, …

We also talked about how AI can produce technically plausible answers without really understanding the real-world consequences. Pooria shared an example where Claude suggested a risky approach for archiving data from a production table containing 14 million rows, which could potentially have caused an outage.

then from there, we moved on to whether AI could replace teachers or therapists. i explained how Voice Mode and weekly recaps have helped me improve my language learning, although we agreed that classrooms still provide structure, accountability, and broader coverage that personalised AI learning may miss. on therapy, we agreed that AI can be useful for reflection and organising thoughts, but a human therapist or mentor may still be needed for deeper or more complex situations.

finally, Pooria and i briefly shared our experiences with AI-assisted SDLC workflows and touched on Claude Code hooks and some of their practical use cases.