Lecture XII - Project Work II
Programming with Python
EP 12: Ship It
One week to launch
Demo Day is next week: every group presents in Session XIII. The order comes out of a hat. The hat is Python.
Before you build: the mark that AI text now carries, and what AI means for your first job.
The Mark
Could anyone tell?
- Your README carries an AI-disclosure section. Suppose you skipped it: could anyone prove that an AI wrote part of your work?
- Since August 2026, EU rules require providers to mark AI content machine-readably, and the major ones do
- The trick behind it is simple enough for three slides
How a watermark hides in plain text
- Before each word, the model splits its vocabulary into a green and a red half, reshuffled at every step
- It then softly prefers green words. The text still reads completely normally
- By chance, a human writes about half green words. A model writes far more
You already know the mechanism
- Session VI: the model picks a likely next token, never the next token
- A watermark is that choice, steered on purpose: green wins a little more often than chance
- Same next-token choice, one hidden thumb on the scale
Who can actually detect it
- Detection is counting: one paragraph of marked text can push chance below 1 in 10 trillion
- But you need the secret key that decided which half was green
- In practice, only the provider can run the test
What a mark proves (and does not)
- Proves: a model processed this text
- Does not prove: who did the work, or whether a human verified it
- Heavy editing can wash a mark out, and short snippets may not carry one, so an absence proves nothing
Presence without disclosure, though, looks exactly like what it is.
Disclosure is part of the work
- The rules converged everywhere: AI cannot be an author, and its use gets disclosed
- Your README section says which tool, for what, and what you verified yourself
- That is standard practice, not a confession
Your Future with AI
AI and your first job
- Thousands of developers: +26% tasks done with AI, juniors gained most (Cui et al. 2026)
- Experienced developers on their own code: 19% slower, while feeling 20% faster (Becker et al. 2025)
- Learners with AI: 50% on a mastery quiz vs 67% without, worst on debugging (Shen and Tamkin 2026)
- The fix: hints instead of finished solutions erased the harm (Bastani et al. 2025). Part I was exactly that
What stays valuable
- The verification boundary: usable delegation ends where your ability to check ends. Growing that boundary outward is your career
- Comprehension debt: code that works today but nobody understands tomorrow. The interest comes due when a reviewer asks “why does this work?”
- Your edge, deliberately kept: explain every line before you accept it, and code without AI sometimes
Discussion: your future with AI
Three questions, five minutes, show of hands:
- Which task from the past weeks would you never delegate to an AI?
- What will an entry-level logistics analyst do in 2030, when AI drafts the first version of everything?
- If AI writes most junior code, where do the senior people come from?
Definition of Done
Before you call it finished
- The repo runs from a fresh clone (cloned once more into an empty folder):
uv sync, then your script - The README says what it does, how to run it, and carries the AI disclosure section
- The commit history tells the story: small steps, messages that mean something
git statusis clean: all committed, nothing important ignored- I am invited, and the link is on Moodle before your demo
Run the fresh-clone test today, not on demo morning. It always finds something.
Demo Day Is Next
Next week’s format
- 7 minutes presenting, 3 minutes of questions, per group
- Order: drawn at random with the
randommodule from Session VI - Show the thing running, your AI disclosure, and one thing that broke
- Everyone in the group has to be there
The rest of today is the last supervised work session: I am in the room for blockers and “is this scope still realistic?”
Wrap-up
Four things to remember
- AI text carries a watermark now: fluent, normal-looking, detectable by the provider
- Your edge: explain every line before you accept it
- The fresh-clone test is the cheapest bug you will ever catch
- A demo is the thing running, the disclosure, and one thing that broke
Next week: Demo Day. The investor is in the audience.
Literature
References
Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. 2025. “Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics.” Proceedings of the National Academy of Sciences 122 (26): e2422633122. https://doi.org/10.1073/pnas.2422633122.
Becker, Joel, Nate Rush, Elizabeth Barnes, and David Rein. 2025. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv. https://doi.org/10.48550/ARXIV.2507.09089.
Cui, Kevin Zheyuan, Mert Demirer, Sonia Jaffe, Leon Musolff, Sida Peng, and Tobias Salz. 2026. “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.” Management Science, February, mnsc.2025.00535. https://doi.org/10.1287/mnsc.2025.00535.
Shen, Judy Hanwen, and Alex Tamkin. 2026. How AI Impacts Skill Formation. arXiv. https://doi.org/10.48550/ARXIV.2601.20245.