You are viewing a draft site. Information here may change between now and the start of the semester.

Lec 1: what this course is

A. why this course exists

  1. origin story: AI transformed how I teach and run 6.390, and watching all sorts of student AI use there has convinced me that using the tool well, the habit of asking what needs AI, is increasingly a fundamental skill in coursework
  2. the stakes rise in UROP- or MEng-like projects: a field report on eight agent-assisted scientific computing projects (2026) found the routine engineering broadly delegable while the researchers’ own time moved to defining what would count as correct
  3. another example of AI “auto” research: Co-Scientist (Nature, 2026) runs lit review, generation, and ranking at scale, with what to keep still decided outside the loop

B. wins and failures from my own use

  1. Anigans, my personal AI: scoped delegation, largely a win, and the reason is a build I own end to end
  2. gallery view of my other AI use
  3. common thread of all the wins: catered to my needs, guardrailed, designed with care; in short, I retained agency
  4. whereas the failures came when I went completely hands-off, when I shouldn’t have

C. judging “good”

  1. self-assessment misleads even experts: experienced developers working on their own repositories ran 19% slower while believing they were 20% faster (METR, 2025), so the feeling of speed is not evidence of speed
  2. similar “discrepancy” in learning, in a randomized four-group comparison (2024), essays revised with ChatGPT improved more than in the other three groups, while knowledge gain and transfer stayed flat, so the artifact improved and the learner did not
  3. elephant in the room: judging “good” can be subjective, so an anonymous poll on two or three cases, “is this good AI use for your learning?”, with two follow-ups, “did it help you finish?” and “could you now do the next one yourself?”; that second one asks students to rate their own learning, and such ratings track how the material felt, since they follow fluency even when memory does not (2012) and effort gets read as poor learning (2019), where students rated the more effortful study strategy less effective and mostly avoided it, while the ones who chose it scored higher on the later test than the ones who chose the easier strategy (with some nuances in reading these results); the case pool follows, and the room runs the two or three left after a staff pilot cuts whatever staff answer unanimously
    1. A paper is assigned for next week, you have AI summarize it and you read the summary instead of the paper, so you can come to class prepared in fifteen minutes.
    2. You have finished the pset on your own, you ask AI to check your answers before submitting, and you rework anything it flags so you do not lose points on a careless error.
    3. You have been hunting a bug in your own code for thirty minutes, you paste it to AI and it names the bug immediately, and you apply the fix so you can move on.
    4. You understand the concept but cannot remember the syntax to set it up, so you have AI write the boilerplate and you write the part that does the actual computation.
    5. You missed lecture, and you ask AI to teach you the topic from the posted notes so you can start the pset.
    6. You are stuck on a concept, office hours are tomorrow afternoon, and you ask AI to explain it tonight rather than waiting.
    7. You are writing your weekly reflection on how you used AI, and you ask AI to help you put it into words because you know what you think but not how to say it.
    8. Your lab partner explained their half of the code and you dozed off, so after lab you ask AI to re-explain it rather than asking your partner again.
    9. Your pset has ten nearly identical problems, you work three by hand until you are confident, and you have AI do the remaining seven.
    10. You have a job interview next week and a pset due, so you have AI do the problems you already know how to solve and you spend the time preparing for the interview.
    11. English is not your first language, and you have AI rewrite your project report so the writing does not get in the way of the technical content.
  4. still, judgment vs. no judgment can be obvious; show the with/without design pair, same prompt, two pages, run as a second poll, “is this slop or polished”
  5. now, what counts as “good” is contested; whether judgment happened is plain; both were true before the AI era; so how did we learn what good judgment is, and how did we learn anything at all?

D. how we ever learned anything

  1. four eras of learning (oral, print, internet, AI); full circle, the medium is dialogue again, now with AI instead of a person
  2. and what has stayed the same? internalizing has always been the learner’s job, and judgment is what internalizing leaves behind
  3. trade poll: pick the era you would have wanted to learn in, then tell a neighbor one thing you would import from another era and one thing of your era you would give up to get it
  4. this “reading the present against history” is the lens we will apply all semester

E. the map and the thesis

  1. the semester at a glance: the module map, the two deliverables, how labs run
  2. every topic draws on EECS fundamentals, agentic AI tools, and applied ML projects; agency is the overlap
  3. what agency means here: owning the calls about what to delegate, what to verify, and what to keep for oneself; tell an agent to do something and even a slop setup will probably get it done, so whether the task got done cannot be the measure
  4. we argue that agency is the central skill for work with AI, the part that does not offload; students may disagree but should do so with evidence

Page updated August 31, 2026.