Review 2026-08-26

[mmesiti]

This issue/thread is to keep track of the ideas and comments during the lesson maintenance session.

Disclaimer: I might have misrepresented @Bahar-Tasdighi and @ashwinvis ideas, please keep this in mind

[mmesiti]

Ideas by @Bahar-Tasdighi

1. There is a lot of material, probably we could make the priorities more clear and apparent
2. Some information is obsolete (gpt-4, data about context windows), this is quite important to fix
3. Many things can stay as before (theory, some exercises)

Comment by @ashwinvis: Most people now have already interacted with a chat, so scenario 1 might not be particularly interesting)

[mmesiti]

from @ashwinvis: The mood of the lesson seems very negative, can we do anything to make it more positive?

[mmesiti]

I personay find "scenario 3" quite interesting - the power of agents is very interesting. But I think that Scenario 2 and Scenario 3 differ not because of the tools used, but because of the attitude: using Cursor and Copilot in VSCode one can still have a completely hands-off attitude, and while using a harness one can still have the code opened in the editor and make changes (this might trigger a small error when applying the patch that will cause the agent to having to reload the files)

[mmesiti]

Add mention of testing, architecture and documentation as ways to improve use of AI and also link the later content of the workshop

[mmesiti]

by @ashwinvis Mention in-context learning, attention, maybe Word2Vec as a foundation?

show this perhaps while talking about embedding space as an analogy.

https://anvaka.github.io/pm/#/galaxy/word2vec-wiki?cx=-17&cy=-237&cz=-613&lx=-0.0875&ly=0.8151&lz=0.5227&lw=0.2337&ml=300&s=1.75&l=1&v=d50_clean

[mmesiti]

mention the debate about Stochastic parrots? See #7 . But it might get too technical?

[mmesiti]

Personal opinion: now the lesson looks a little organized like a tragedy: start with the nice things, then things get more powerful and sinister and eventually we've got a part about security considerations.

Perhaps now that people more or less have interacted with and are using AI already, we might start with all the risks (security, hallucinations) and talk about how they can be solved/mitigated.

[ashwinvis]

Could include a bit of anatomy of LLM from https://gitlab.tuwien.ac.at/vsc-public/training/LLMs-on-supercomputers/-/blob/main/presentations/D1_LLM_intro_and_promptengineering.pdf?ref_type=heads

• Tokenization
• Embedding space
• And perhaps attention head (?) or wherever the knowledge is stored
• In-Context Learning -> the reason why context happens and the prompt is important.

What we gain from that: a mental model for the learners, rather than just a black box schematic.

[Bahar-Tasdighi]

Prioritize the current material, move some of the less essential content to the appendix, and add relevant theoretical background from these slides: https://gitlab.tuwien.ac.at/vsc-public/training/LLMs-on-supercomputers/-/blob/main/presentations/D1_LLM_intro_and_promptengineering.pdf?ref_type=heads

[ashwinvis]

Adjust length and match interest and current knowlege of learners:

• Scenario I: shorten it. Chatbots are ubiquitous by now?
• Scenario II: lengthen. Go beyond tab-completion, perhaps by making a final exercise on generating unit-test. Also a nice segway to the next lesson on testing
• Scenario III: good to consider an example. Should we stick to the sandbox example with docker or some other way.

Also bring up the overlap in scenario II and III from the appendix, to the opening statement?

[Bahar-Tasdighi]

A few points to consider for the lesson maintenance:

• Go through the current topics in each scenario and decide what we want to cover and what we should skip due to limited time.
• For the skipped parts, decide whether to move them somewhere else or just hide them.
• Update outdated material, such as context-window figure, model/version references, or anything else that may no longer be current.
• Do we want to introduce the concept of LLM autonomy? If yes, should it be included in the introduction?
• Remove the standalone Security section(?) and integrate the relevant security aspects into the individual scenarios. (or other way: keeping one separate Security/Risk section and removing security content from everywhere else.)
• Review the exercises in Scenario 2 and Scenario 3. Should we include them (they were skipped last time)? do they need updates?
• For Scenario 3, we have a nice experiment demo. What kind of exercise can be considered?