Session 2 Feedback — Learning Machines

About you

Name or display name

Optional — leave blank to stay anonymous.

How did you attend?

Live on Zoom Watched the recording Some of each

Before this session, how familiar were you with how image generators work inside?

This was my first look at the mechanics I knew some of the vocabulary but not the moving parts I mostly knew this and came to see how it was taught I have a technical background in AI or machine learning

The shared activity

Which parts made something click for you?

Choose all that apply.

From tokens to pixels (the blurry-image walkthrough) WordNet and ImageNet (the human pre-work under everything) Diffusion as denoising (chiseling an image out of noise) The doctor default (the live test of what gets filled in) Drawing the prompt in the Human Diffusion Canvas Comparing the room's drawings with each other and the machine's Aurora's guest talk The debrief discussion Nothing has clicked yet — and that is useful for us to know

In one sentence: when an image model generates a picture, what is it actually doing?

No grading — this tells us what the session actually taught.

What is still fuzzy or unanswered?

The tools

Which tool did you spend the most time with?

Human Diffusion Canvas Diffusion Step-Through Viewer The Squint Test (feature extraction) Image–Caption Match Lab Default Test Comparison Viewer I worked mainly in the Image Default Test Board I did not get to the tools

Did you get an export out of the Human Diffusion Canvas?

Yes — a PNG Yes — a GIF Tried, but the export did not work I did not try to export I did not use the canvas

Any tool friction?

We already know about the canvas brush color not changing between steps and the brush size being overridable — anything beyond those? Anything confusing, broken, unreadable, or hard to use while the session was running.

Guest spotlight

One thing from Aurora's talk that stuck with you?

An idea, a work, a question it raised — anything.

Format and pacing

How was the pace?

Too fast About right Too slow It varied — say more below

Participating through Zoom chat:

Worked well for me Okay, with some friction Hard to follow or keep up with I mostly watched, and that felt fine

Any technical trouble?

Shared slides, audio, links, screen readability.

Comfort and consent

Could you participate the way you wanted — camera off, chat only, or just watching?

Yes Mostly No — tell us more below

Anything facilitators should know about?

An example or prompt that felt off, concerns about the recording, data, or consent, or anything that made participation harder.

Between sessions

Which assignment option are you planning to take on?

All six routes are on the assignment page, and all count as full participation.

Make one image-generation artifact — the session ask, 45–90 min Run the default test — 15–20 min Run my own Human Diffusion Canvas with other people — 25–35 min Find where spatial reasoning breaks — 20–30 min Adapt the marker version for a classroom — 45–60 min Analyze a frozen example — no live AI, 20–30 min Not sure yet / not this week

Looking ahead

What should we make sure to keep?

What should we change before Saturday's video session?

What question about video generators do you want answered?

Next week asks: how does a model keep a moving image coherent from frame to frame?

Would you recommend this session to a colleague or friend?

Definitely Probably Not sure yet No — and we would genuinely like to know why above

Anything else?

Response sent — thank you.