I have been doing a lot of solo tinkering around this very idea —contextualizing learning for every middle school through college student according to their career and personal interests. Props for this team for actual piloting with real students. What I’ve found however is that students’ interests (in terms of novel engagement) move at the speed of culture. It’s the quirky TikTok personality, the local slang, aura farming memes and other stuff I hardly understand at my age. So really it boils down to how you decrease the lag time between creation of students’ cultural phenomena and updating/tuning models to account for this new (or localized) information. It also requires a UI/UX that accurately flags what part of this information students actually engage with —an even harder tasks since kids can’t write or communicate this type of information in language LLMs draw patterns from (if they can write descriptively at all). Great teachers who are in constant socioemotional and cultural exchange with their students figure this out at the classroom level… not sure it’s a solution that technology can replicate at scale but it’s still fun to tinker.

Any attempt to integrating modern culture into an LLM will be doomed to hilarious failure, like most attempts by schools to identify with the culture.

Like you say, it moves at the speed of light; they'll integrate the culture of a few months ago and it's already different. Kids will see the LLM trying to relate to them like their 40 year uncle trying to relate to them and will see right through it.