some notes from 10/9 meeting

[eglerean]

This is not an issue more like a container of issues :)

Future stuff

• Future versions of the lecture can be more time agnostic (i.e. not exact versions of models, but links to external pages) // future
• similarly for context window lengths and other technical parameters from the models/tools // some udpates now but then future

Risks, security, and other scary things

• too much security everywhere, we could move it to a single place: beginning or end?
• right now each scenario has its own "do not do this" which could be hidden with drop down and/or link to the "security" page // Bahar can hide
• researchers might be more worried about "validity" than security and more broadly academic misconduct (fabrication or falsification in the data/methods/results because of uncareful AI use)
• data privacy and other misconduct issues (which are not related striclty to AI)
• we could rather use the word "risk" which can include cybersecurity, validity, reproducibility etc etc // Enrico can draft
• MoE models have also classifiers that checks security "Is it dangerous what you are doing?" (to be added or point readers to external sources: be aware but what is the solution?)
• and what are the risks mitigations?
• on validity enrico does a draft and adds some elements on responsible conduct of research e.g. from https://scicomp.aalto.fi/scicomp/rcr-scicomp
• let's use cross refs for easier less-manual cross referencing https://myst-parser.readthedocs.io/en/latest/syntax/cross-referencing.html

Pedagogical considerations

• do people need to see code generation first before talking about risks?
• IDEs (scenario 2) or CodingAgents (scenario 3): majority has no idea of these and they
• first we tell all the beautiful things you can do and then all the bad sides with risks (validity, misconduct, cybersecurity, data privacy etc)
• for scenario 3: should we bring containers already in the demo?
• exercises for all scenarios -> not enough time. What to do? It is difficult to support all learners for setting up scenario 2 and 3.
• for the scenario 3 demo Bahar can check if it can be docker+claude code with remote (or local) model
• Should we demo opencode or pi instead of claude code? codex is open source also // Bahar to check (it can take time to set things up compared to claude code or codex which have a giant system prompt). Plan B is also that Ashwin also as a mistral subscription which could point claude code or codex (or some of the other agents).

other things to do before

• Consolidate intro page with issue Review 2026-08-26 #17 and maybe get rid of scenarios pictures in intro section
• Incus is a good alternative https://linuxcontainers.org/incus/ add a link
• add link to PI and other agents like opencode?
• mention context window size and agents -> open models have short context and then things might not work, but we have new models like Qwen 3.8 27B (add link?) and then local memory becomes a bottleneck. some of these open weights models think a lot -> consume context window fast.