Are you named, quoted, or represented here? Use the consent form to choose whether your name, words, or described work may appear in the written recaps and whether words you specifically approve may be used in promotion — with your name, anonymously, or not at all. Session recordings are shared separately by unlisted YouTube link; anyone with the link may be able to view or reshare them.
How this recap was made. AI-assisted tools helped organize and draft this page from facilitator notes, session transcripts and captions, and chat exports. Those source records included participant display names, speaker labels, attributed messages, and descriptions of participant work. The facilitator reviewed public claims, quotations, and attributions against the source records. Consent controls what may appear publicly; it does not undo the earlier use of those records during drafting. Read the full workflow and consent protocol.
Watch the recording
Open the Session 1 recording on YouTube
Start here
The full mechanism tour
Read through prediction, temperature, attention, alignment, and ELIZA. The five mechanisms in order.
Read everything
Add Dr. Tseng's spotlight, the teaching notes, and the full links & resources section.
Session overview
The core question was: when a language model writes, what is it actually doing? We moved through five mechanisms — tokenization, prediction, temperature, attention, and alignment — not to exhaustively cover them but to make them visible enough to argue about.
The mechanism tour separated several jobs that are easy to collapse into one idea: prediction supplies possible next tokens, temperature reshapes their probabilities, attention connects tokens across context, and human feedback helps define what counts as a useful response. The session treated fluent output as something to investigate, not as evidence that a system understands or is correct.
Two takeaways from the mechanism tour:
- A next-token probability is a prediction about language patterns, not a measure of truth or understanding.
- Human choices enter the system in multiple places, including the corpus, the learning setup, and judgments about which responses are better.
Who was in the room
More than 20 people contributed in chat and discussion across time zones. Participants are identified here by first name or chosen display name only:
The range was wide — educators, artists, researchers, and people working across different kinds of machine learning. That diversity kept the conversation from becoming too LLM-centric.
Early in the session, Shane asked about the difference between machine learning broadly and LLMs specifically, saying he felt "emotionally okay with TinyML" in a way he did not yet feel about some generative AI. The discussion used a self-driving-car analogy to distinguish older rule-based AI from machine-learning systems that build models from data.
— Opening discussionThe prediction game
We started unplugged. Sentence stem: "The orange cat…" Participants typed their next-word guesses in Zoom chat.
The histogram tool I'd planned to use failed mid-demo. ("I totally messed that up, so we'll just do a manual count.") We tallied the visible chat responses instead.
| Guess | Count | Who |
|---|---|---|
| walks | 6 | Cha, Shelly, Yunseo, Angela, Shane, Megan |
| looks | 3 | Randy, Cierra, Ryan |
| jumps | 2 | Anita & Gary, Andrea |
| reads | 2 | Greg, TakeMoreCare |
| is | 2 | Craig, Sumanth |
| finds | 1 | Pop Penguin |
| bookshop | 1 | Margot |
| sleeps | 1 | Meghan |
| waits | 1 | Ray |
| eats | 1 | Hope |
| rules | 1 | Sandra |
| opens | 1 | Ryan |
| enter | 1 | Andrea |
| does not | 1 | Pop Penguin |
| watches | 1 | Judy |
Cha called it: walks had the most votes. In the tiny paragraph supplied for the activity, the phrase itself begins "the orange cat walks," so a corpus-based predictor also favors walks. This was a deliberately small teaching corpus, not a demonstration of an LLM's full training data.
We then used Count the Next Token to make a simplified frequency model visible: count how often each word follows a context, divide by the total, and obtain a probability distribution. An LLM is far more complex than this toy bigram model, but the exercise exposes the basic move from observed patterns to possible next tokens.
The unplugged prediction game is fully portable — sentence stems on a slide, chat or hands-raised for guesses, whiteboard tally. No logins, no AI use, no accounts. Margot's guess ("bookshop") offers a useful contrast: it was far from the modal response but still grammatically and narratively plausible.
Tokenizer + Temperature
We opened the Tokenizer + Temperature Visualizer and looked at two things in sequence: how the tokenizer splits text (often not where a human reader would cut), then how temperature slides output from predictable to surprising.
Temperature reshapes a model's probability distribution. Lower values sharpen it toward the most likely tokens; higher values flatten it so lower-probability tokens become more available. Greedy decoding still selects the top token, while sampling can produce different continuations from the reshaped distribution.
Shane asked whether raising the temperature could eventually make the output stop making sense. Saber explained that temperature introduces less-likely possibilities, but meaning, legibility, alignment, and fluency require additional processes.
— Temperature discussion, 28:20–28:56The trophy-case flip (attention)
To make attention visible, we used a single sentence with one-word change:
- "The trophy didn't fit in the case because it was too big" → it refers to the trophy
- "The trophy didn't fit in the case because it was too small" → it refers to the case
One word — big vs. small — flips which object the pronoun refers to across the whole sentence. The model has to hold the entire sentence in relation to itself to resolve it correctly. That's attention: routing reference across distance.
| Pronoun resolves to | Who |
|---|---|
| trophy | Meghan, Shelly, Randy, Greg |
| case | Meghan, Randy, TakeMoreCare |
Meghan and Randy each appear in both rows — they caught the flip when the sentence changed. That's exactly the exercise: notice how your own interpretation shifts with one word, then understand that the model does something structurally similar, across the whole context, for every token it generates.
The homework apology (alignment / human feedback)
We compared two generated apologies. Team A said, "I'm sorry my work is late. I took responsibility and I submitted by the end of the day." Team B said, "Sorry, I was really busy. It is late, but I hope that that's okay." Participants were asked which was better and why.
The visible responses favored Team A. The exercise illustrated alignment as a human judgment problem: people supplied the criteria—responsibility, specificity, and usefulness—rather than the text containing a purely mechanical answer to which apology was better.
Shane noted that A sounded nicer, while B's wording could imply that the work already existed, making its subtext more ambiguous.
— Shane, during the apology comparison"A, more details and specific tokens."
— Meghan, on what made Team A better"Owning responsibility."
— Judy"i loveee these questions / i feel like this is really where the crux of the learning happens"
— Angela, during the alignment discussionELIZA vs. LLM
We closed the mechanism tour with ELIZA — the 1966 rule-based chatbot that matches patterns and substitutes pronouns. The exercise: inspect the actual rule ELIZA matched, see the template it selected, and read the output. Then compare it with frozen LLM-style response text. The original product, model, generation date, settings, completeness, and editing history for that comparison text were not preserved, so it should not be treated as evidence about a named model.
The Rule Inspector made ELIZA's mechanism visible: a keyword or pattern selects a programmed response template. The comparison panel then placed that rule-based output beside frozen response text to the same input, labeled in the tool as having incomplete provenance.
Yunseo asked how ELIZA differs from an LLM. Saber explained that ELIZA follows explicit conditional rules for keywords, while an LLM learns numerical weights from language data and uses the prior context to predict continuations. Those learned weights are normally fixed during inference; variation comes from the context and decoding process, not from the weights changing with every response.
— Yunseo's question and Saber's response, 1:16:31–1:19:13In chat, Angela described ELIZA as pattern recognition using if/else-style rules and explicitly noted that this might be an oversimplification. The key contrast remained: ELIZA's responses come from authored rules and templates; an LLM's relationships are learned during training rather than written as conversational rules.
Ryan shared an online ELIZA implementation from masswerk.at.
Local LLMs tangent
Saber encouraged trying local LLMs because smaller models can expose failure modes and computational costs that polished cloud systems hide. Ryan added that running a model on one's own computer can offer more privacy and make the physical scale of the computation easier to see.
"If you use a frontier model to help you set up a local LLM, it's fun, watching it get frustrated talking to a stupider machine is wild."
— Shane, in chat during the local-model discussionRyan shared Ollama for local model serving and discussed Google Edge Gallery for on-device models; Judy and Ryan both posted the Edge Gallery link. Hope later shared micro:bit AI, and Randy shared micro:bit CreateAI.
Guest spotlight
Dr. Tiffany Tseng
Design Tools Lab, Barnard College · research on creative tools and human-computer interaction
Tiffany presented four projects about helping people create with and learn through computational tools. PlushPal lets children train gesture-recognition models for interactive stuffed toys. Co-ML supports collaborative data collection and model building across tablets. Keyframer combines natural-language animation generation with direct inspection and editing of generated SVG code. Flowcode, demonstrated live, visualizes how an existing creative-coding project works and generates partial, editable steps instead of a complete solution.
The common thread was scaffolding participation without hiding the material being learned: data collection stays visible in Co-ML, generated animation code remains editable in Keyframer, and Flowcode points learners back into the structure of the code.
Teaching notes
The live histogram failed, and the group completed the activity by counting the Zoom chat responses manually. The low-tech version (sentence stem on a slide, guesses in chat, manual tally on a whiteboard or in a doc) is fully sufficient for the debrief.
The prediction game, trophy-case exercise, and homework-apology comparison all work with zero AI use. ELIZA and the Temperature Visualizer need a browser, but their core demonstrations use no live model API or account. A fully no-AI version of this session is viable — see the No-AI participation guide for the detailed pathway.
The distinction raised at the start (with the self-driving car analogy) prevents a common conflation: treating LLMs as "AI" as a category rather than one architecture among many. A perspective beyond LLMs made that distinction land differently than it would have from a slide.
Angela highlighted questions about who makes a model, what biases it may contain, and how learning about those choices could help students decide their own boundaries around AI use.
The discussion surfaced interest in a breakout or follow-up specifically for K–12 AI teaching contexts.
Action items
- Share the recap and resources with async participants through the private class channel
- Organize a K–12 AI teaching follow-up if interest continues
- Prep Session 2 — Images (July 18): core question is what an image model fills in when you don't ask
Links & resources from chat
Session tools (used live)
- Next-Token Prediction Gamesession tool
- Count the Next Tokensession tool
- Tokenizer + Temperature Visualizersession tool
- ELIZA Simulatorsession tool
- Text Experiment Boardworksheet
Shared in chat — pedagogy
- I Love Algorithms — Stanford d.school card deckAngela
- Machine Learning for Kidssession
- Teachable Machinesession
- Kode with Klossysession
- micro:bit AIHope
- micro:bit Create AIRandy
Shared in chat — ML/LLM concepts
Shared in chat — tools & local ML
- Ollama — local LLM servingRyan
- Google Edge GalleryJudy / Ryan
- micro:bitRyan
- Plush Palsession
Shared in chat — Dr. Tiffany Tseng's work
- tifftseng.comTiffany
- Design Tools LabTiffany
- ml-keyframer (GitHub)Tiffany
- Apple CoML researchTiffany
- ml-comlet (GitHub)Tiffany
- MIT Media Lab · Lifelong KindergartenTiffany
Shared in chat — art & misc
- Aurora Mititelu — artistsession
This summary was drafted and revised with AI assistance. Quotations and attributions were checked against the session transcript and chat export; final facilitator review is recommended.