Session 1 — Orientation

Thursday, 27 August 2026 · Towne 217

Cleaned and reorganized from the classroom recording. Student remarks are paraphrased and de-identified; the instructor’s remarks are lightly cleaned but kept close to verbatim. Checked against the session roster before publication — no participant name, first name, or address appears below.

Thank you all so much for coming — for your interest. This course was only listed two days ago; it was a bit of an uphill battle to get it listed, so I’m glad we get to do this, and glad you came to check it out.

As with most of my courses, we’re going to use Slido throughout the evening — go to slido.com, code 9999. As is the policy in my courses, we always do things with our names: you’ll log in and your name will show whenever you post something. In a course on metacognition, we are really invested in owning our own opinions. Taking a stance, being wrong — including being wrong in public — is not a bad thing. What we’re developing is the practice of adapting to new information and new evidence.

You can ask questions on Slido and I’ll try to monitor them, and every now and then I’ll open a poll and ask you to think and respond. Often there’s no “right answer” — the point is for us to think together. Your submissions on Slido are as important, if not more important, than what I’ve prepared for the conversation.

I want to show you a few headlines you’ve probably already seen. Some examples shown and discussed: “OpenAI’s Astra model solves ten open math problems for $2,000” (Quartz), “When Will the AI Stock Bubble Burst? 2027 — or Sooner” (Bloomberg Opinion), “More Than Half of Layoff Events Tracked in 2026 Cited AI or Automation” (IBTimes UK), “I caught my students using AI to cheat in an exam” (Nature), plus others in the same vein about AI replacing jobs, AI arms races “putting humanity at risk,” and a viral chatbot that turned out to be a real human.

Look at the headlines: it’s either a huge opportunity or the end of the world — nothing in between. So who knows if we’ll still need mathematicians; and at the same time people are worried the AI stock bubble is going to burst, the way the dot-com bubble did. And of course everybody’s worried about cheating at universities — nobody’s sure anymore what the point of a university degree, or of art, or of anything, really is.

Slido word-cloud prompt: “In one word — how does all of this make you feel about your future?” (90 seconds, one word.)

Words submitted included: excited, unsure, scared, uncertain. There’s a mix of feelings — but notice: some of these headlines are asking you to throw away all doubt and be optimistic, and some are asking you to be maximally doubtful. These are just two poles of the same non-invitation to actually think and develop your own position.

Student comment (paraphrased), on feeling unsure: The student wasn’t sure the hype headlines would actually materialize — for instance, whether AI would really replace certain jobs, and how much difficulty remains between where things are now and industry-wide replacement. They also noted a gap between what the news describes and what they’re actually experiencing when using AI in practice, and wondered where that gap comes from.

I summarized: part of the uncertainty is about job training — you’re being told, in one universe, that certain jobs are no longer useful, and in another that they still are, and you can’t tell which one you’re in.

Student comment (paraphrased), on feeling excited: Compared with ten years ago, AI now speeds up the whole path from idea to implementation. It can provide scaffolding even in domains where you don’t have expertise.

“Scaffolding” is a technical term — it means AI helps you take the first steps, no matter what you’re trying to do.

Student follow-up (paraphrased): That same scaffolding can also make you unsure — because you don’t know whether relying on AI’s help will ultimately serve you well or make you “weak” or “stupid.” How do you know when using AI benefits you versus harms you?

We have no firm answers to this. Right now we just have a lot of people’s opinions, and our job is to figure out what we believe — an opinion we actually understand.

This is metacognition. It means recognizing that every question has many facets, that the facets we currently understand aren’t the only ones that exist, and that the only way to learn more is by asking questions and listening to what comes back.

What we just did also used a computational tool — Slido helped surface different perspectives in the room and helped me see them. We’ll do this often. And notice: I try to share my perspective without overwhelming you with it, without flattening yours with mine. My job isn’t to tell you what to think — it’s to model what it looks like for someone to think, so you can do your own thinking.

Part of my job is to get you talking to each other, meeting each other intellectually, with interesting questions. I’m not here to give you answers — I’m here to give you questions, ones I hope are generative, and for us to investigate them together, listening to each other.

I’ve mentioned this before in other courses, so apologies for repeating myself: the best moment of last semester was a Claude Builder Hackathon — students, postdocs, grad students, a few faculty, building together for three hours on a Sunday with Claude. What made it special wasn’t the tool — it was that everyone got to be curious about each other’s ideas, without needing technical expertise to participate. It was the only event I’ve experienced at Penn that was genuinely designed for people to talk to each other — not a social function with awards, not a performance, just people curious about each other. Those opportunities are rare, and I don’t think they should be. That’s what I want to build here.

When I grew up, nobody handed me a manual for how my brain works. Over time I’ve had to reconstruct that understanding myself. The traits I’ve identified are called neurodivergent traits — differences in how the brain works:

• HSP (Highly Sensitive Person) — a high bandwidth of sensitivity.
• Autistic — I process a lot of information in parallel; I care a great deal about rules, and I need things stated in ways I can infer for myself.
• SDAM (Severely Deficient Autobiographical Memory) — I don’t remember episodically the way most people do. As we have this conversation, as I look at your faces, none of it will be retained as a memory of a scene. I remember things semantically and structurally — the ideas, not the moments. That’s why I record everything.
• Aphantasia — I don’t see images in my mind’s eye. If someone says “picture a horse,” I see black, not a horse. I assumed “picturing” was just a figure of speech until I learned some people actually see an image.
• LLI (Low Latent Inhibition) — I notice everything. If two people are chewing on opposite sides of the room, I hear both. I can’t filter it out.

None of these are disabilities or deficits — they’re differences. What’s hard is when nobody knows about your differences, including you, so nobody — including you — is working with them in mind. Things don’t work the way people expect, and nobody understands why.

Ten years of metacognitive work has helped me build an understanding of my own quirks that empowers me — both in understanding why I think the way I do, and in explaining myself to others, the way I’m doing right now.

The more you learn how you work, the more you understand — and here’s the key part: the more agency you have. Agency is the word I’ll use a lot this semester. It’s close to what we mean by free will, but more precisely, it’s the word that qualifies how you make decisions with the information available to you. Agency over your thoughts, your actions, and most importantly your own learning process is the biggest thing you can take control of in your life. That’s what I’ll try to convince you of this semester.

This course has a hard computational component — we’re going to build things, and I’ll talk more about models later. But first, let me connect this to classical roots.

(The instructor asked who a projected portrait was; a student guessed Aristotle; the correct answer was Socrates.)

Socrates is the first philosopher — mentor of Plato, who mentored Aristotle (so Socrates is, in a sense, Aristotle’s grandfather). He was known for continuously asking “why” — he was passionate about self-doubt: how do I know something? Why am I convinced of something? His method, the elenchus, was to confront every belief with questioning. He believed the unexamined life was not worth living — not in the sense of “you should learn facts,” but in the sense that the more you learn, the more agency you have. In that sense he’s the ancestor of The Matrix — the idea that we’re all inside a system, and we can only learn about that system by questioning and “waking up.”

We’ll talk more about Socrates, and about his more recent successor, Paulo Freire — some of you may already know the terms “banking model” and “problem-posing model.” This has to do with how knowledge is transmitted; we’ll come back to it.

Socrates was also the first to identify how important language is — the words we use can empower us, or trap us. That will be a running theme.

The iceberg metaphor: knowledge is like an iceberg. There’s the tip — what everyone sees — and below the waterline, all the thinking required to actually justify what’s at the tip. Most people only ever learn the tip, and if you only learn tips, you can never reconstruct knowledge. Over this course, we’ll systematically ask ourselves: is this just the tip, or are we actually looking at the thinking beneath it?

We only meet once a week; there are just two more sessions before you need to decide whether to enroll.

As with my previous course, everyone who takes this course has an A — already, as of tonight. This isn’t generosity or laziness. It’s design, and I want to explain why.

Question to the room: has anyone heard the term “reward function”?

Student answer (paraphrased): A reward function is: you do something good, you get a reward.

That’s a decent start. Where does this term actually come from?

Student answer (paraphrased): From reinforcement learning — a function that measures how close you are to some target.

Exactly. In reinforcement learning — how large language models are trained — a model is given many questions, and the answers are rated; those ratings reinforce certain answers over others. That’s the model’s reward function: it defines what counts, to the model, as a good answer.

How does this map onto your own learning?

Student answer (paraphrased): Maybe the grade is the reward function.

Another student answer (paraphrased): When you learn something useful, that feels like the reward.

Both are right, and they illustrate extrinsic motivation (an external authority convincing you something is good for you) versus intrinsic motivation (you knowing, yourself, whether something is interesting).

We are all learning agents, and the reward function is how we orient our learning — how we evaluate success against failure, and how we iterate. When you follow what’s intrinsically interesting, one question unspools into the next. But what happens when you don’t actually care about the material, and you’re only doing it because someone told you it’s required, that there’s a grade, a checkbox?

It is my belief — and that of many people in education and machine-learning research — that grades corrupt the reward function of students. We end up “teaching to the test.” The sycophancy problems we’re now observing in LLMs are, interestingly, helping us understand this same dynamic in a new light.

What grades actually make you do is approximate my reward function — because I’m the one assigning and evaluating. You’re trying to guess how I will judge things. Even in the best case, where I’m a very accurate judge, you’re approximating an approximation: my reward function isn’t the truth, and you only ever see it through rubrics and guesses. And critically — approximating my reward function is only useful for as long as I remain your evaluator. The moment you graduate, any specific rule you learned about how I like things (say, “always capitalize after a period”) becomes worthless. That’s a skill that doesn’t compound.

Some of you were in my Analysis of Algorithms course last semester, where I ran an entirely new syllabus for the first time. I didn’t know how it would be received, but the course evaluations were reassuring — particularly the “amount learned” score of 2.87, which sounds average until you know that Analysis of Algorithms is normally an extremely punishing course, full of nightly derivations, where people often report a harsh experience independent of how much they actually learned. This was the highest “amount learned” score that course had received since evaluations started being tracked — without running it like a boot camp, and without making anyone feel like an F student.

The other unusual result: for the first time in my time at Penn, I had 100% “no” on a cheating question. Why? Because of the grade. Remove the grade, you remove the incentive to cheat. If you’re spending your own time — time you could spend interning, and your own money, taking out loans to be here — why would you ever cheat on something you’re paying to learn?

Feedback comments from that course affirmed that alternative teaching approaches aren’t inherently bad, and that there’s room in a university for different ways of covering material — which is what I’m trying to build here too.

The recurring test we’ll use all semester: whenever you read a text, take a course, watch an interview — ask whether it teaches you, or makes you dependent. After reading it, do you feel like you’ve learned something you can take to the bank — more autonomous? Or more dependent on the authority that produced it? This is how authorities become indispensable.

Anecdote (Ezra Klein): Fifteen years ago, following politics from France via the New York Times, the instructor started reading a young reporter named Ezra Klein, who later founded Vox, then got a podcast, then joined the Times. Klein is smart and writes thoughtful books about topics like housing shortages — but the instructor noticed that every time he consumes Klein’s work, his main feeling is “oh my God, this is so complicated, thank God I have Ezra Klein to explain it to me.” That feeling is dependency: Klein gives so much information without triaging it that it becomes overwhelming, but because the audience is primed to feel overwhelmed already, they don’t notice that the overwhelm is itself the outcome the content produces. The real question to ask of any content is: have I learned anything new, or have I just been comforted in my existing opinions? Do I have any more agency, any more sense of what I could do, than before I consumed this?

(Short break — the room reconvened around 18:05.)

During the break, several students asked administrative questions about waitlists and enrollment; the instructor confirmed that anyone who applies to the waitlist and completes the application will likely be admitted, and asked students having trouble to email him directly.

To get to know each other — since this room isn’t yet a fixed roster, and who you’ll be sharing this room with matters — the prompt was: “How would you characterize your relationship to thinking?” Two minutes, written, no need to be polished.

Selected anonymized responses, read aloud by the instructor:

• “I think I’m mostly curious about knowing how things work under the hood. I sometimes overthink, especially when there are many possibilities… I have been easily molded by the system to exchange thinking for speed or results… I feel I have been improving since the start of this year, but I have a lot left to discover.”
• “Sometimes I pay too much attention to the outcome, and that bothers me a lot. I am a curious person… Questioning my understanding is good when it helps me progress, but bad when it causes stress from overthinking.”
• “I tend to think a lot and question my own assumptions. I think best visually — with pen and paper or on a whiteboard.”
• “I like to work through different perspectives, but sometimes I overthink. I’m producing a workshop on critical thinking, and studied philosophy in undergrad. I love thinking so much, I overthink.”
• One response described thinking as a tree with many branches — good ideas and bad ideas — where the work is to “prune” the bad branches so they don’t crowd out the good ones, and to let the tree grow recursively in the right direction.

The instructor commented that being consciously aware of what you know — no more, no less — while staying hungry for what’s ahead, is a kind of sweet spot (though not the only one), and compared the “pruning” metaphor to literally pruning rose bushes over the weekend.

Student question (paraphrased): Why do so many people describe themselves as “overthinking” — is it really overthinking, or is it something else, like procrastination?

The instructor’s response: as a neurodivergent person, he experienced “overthinking” as a label applied to him by people who were themselves under-thinking by his standards. “Overthinking” isn’t inherently a bad word, but behind it is an unexamined social norm — part of metacognition is asking whether you actually endorse that norm, or whether you’re importing it uncritically. You might decide you like it and want to meet it; you might decide you don’t, and can discard it. Metacognition isn’t about declaring things good or bad — it’s about giving yourself more conditions under which you can make conscious decisions.

This is a framework the instructor has developed over the past year to characterize different stances we can take with language — and with AI.

Cybernetic comes from Greek κυβερνήτης (kubernḗtēs) — “the steersman” (the root also behind the tech term “Kubernetes”). Language is cybernetic when its role is to produce an effect. Example: “You look nice today” — said to make someone feel good. “I am nervous” could be epistemic (a fact) or cybernetic (a ploy for sympathy), depending on intent.

Epistemic comes from Greek ἐπιστήμη (epistḗmē) — “knowledge.” Epistemic language shares a state — it asks, “how does this refine my model of reality?” Example: “We are in this classroom together” — a shared, checkable fact. With cybernetic language, you’re not sharing something — you’re trying to change behavior. “Open the door” (produces an action). “Someone’s going to get in trouble” (produces fear).

Neither mode is bad — we use both all the time. What’s bad is thinking you’re doing one while doing the other, or thinking you’re receiving one while actually receiving the other. Going back to the Ezra Klein example: if you think you’re being informed (epistemic) but you’re actually just being made to feel that Klein is smart (cybernetic), the harm is in the misrepresentation.

Student question (paraphrased): Can a single statement fit both categories?

Yes — “I am nervous” was exactly that example; it’s hard to evaluate a statement without context.

Slido exercise — “Try it out”: Write one statement, labeled CYBERNETIC: or EPISTEMIC:. Responses discussed live, one by one:

• “Sit down” → cybernetic.
• “My shoes are white” → epistemic (assuming true).
• “I’m enjoying today’s class” → epistemic if sincere, cybernetic if flattery.
• “Nice to get a name, of course, it’s worth it” → cybernetic if said to look like a hard worker.
• “Don’t make me bleed” → cybernetic (appeal to emotion).
• “I think by questioning what I believe in, [I’m] trying to understand what is actually true” → epistemic statement about one’s own thinking.
• “This apple is green” → epistemic (factual).
• “If everyone looks confused, the professor slows down” → discussed as a cybernetic feedback loop (Wiener’s original cybernetics), though not quite the interpersonal sense being defined tonight.
• “Today is a good day” → cybernetic (vague, designed to sound good).
• “Cybernetic intelligence emerges from feedback between actions and their consequences” → actually closer to Norbert Wiener’s original technical definition of cybernetics.
• “Policy swings between left and right wing” →

Student clarification (paraphrased): They meant it as an observation about how political outcomes swing, not as a belief statement.

This led to a tangent on convince vs. persuade: convince (from Latin con- + vincere, “to thoroughly vanquish”) — properly, two people together being “vanquished” by the truth; the goal in convincing is that everyone wants to find the truth. Persuade (from suadere, “to make sweet”) — to get someone to swallow something sweetened that they might not otherwise accept if paying full attention.

• “An agent should update its beliefs when new evidence comes” → epistemic (an injunction, but grounded in updating toward truth).

The exercise, the instructor noted, wasn’t about getting the “right” answer — a definition given only through positive examples doesn’t reveal its edges. Examining mistaken examples together is how we explore “the submerged part of the iceberg.”

Mapping this onto AI use:

• Cybernetic use of AI = you approach AI with a fixed outcome in mind: “produce this report,” “write this letter for my beloved,” “get me an A.” You’re giving AI an outcome to aim for.
• Epistemic use of AI = you approach AI with your own thinking and ask, “What have I missed? Tell me more. Who else has thought about this?” — and the AI might connect you to Socrates, or Freire, or your own professor’s ideas.

Student question (paraphrased): Can you really reason epistemically with an AI, though, if you can’t be sure what it says is true?

Another student’s response (paraphrased): It’s similar to asking a friend something they might not know either — a friend’s answer isn’t guaranteed true, just useful.

The instructor agreed: we tend to frame AI as either an infallible oracle or garbage, but that’s a false binary — no person is an oracle either. When you talk to other people, you’re getting their opinion, their sample of experience, not the truth. The task, always, is to evaluate critically — not to accept something wholesale because it comes from AI, from a professor, or from any authority — but to check whether things make sense, whether they add up.

Follow-up student question (paraphrased): What does an epistemic use of AI look like when you’re trying to learn something completely new, where you have no prior basis to check its claims?

That’s still epistemic — the important variable is why you’re using the tool the way you are. Sometimes cybernetic use is fine (you’ve thought hard and are ready to act); sometimes it’s a sign you don’t actually know or care why you’re doing something — e.g., an assignment you don’t understand or care about, just a deadline and a grade to hit. In that situation, cybernetic AI use “is going to kill you,” because you were never invested in the outcome to begin with.

The Elena diptych: Elena is hired at a nonprofit fresh out of school. After a few months she’s asked to produce an activity report. Imagine two parallel universes with two Elenas, same job, same deadline, same AI, same report submitted.

Cybernetic Elena is told the report needs five pages, charts, upward trends. She goes to AI and says “produce this report,” gets something that looks right, and her bosses wave it through onto the conveyor belt. Epistemic Elena instead brings the AI all the underlying data — surveys, financials — and asks, “How are we doing? Help me understand. What could we do better?” She interrogates the AI’s suggestions (“Why do you say that? Show me.”), and only after hours of genuine inquiry does she ask it to help write up the report.

Same deliverable. Same task. But one process produced something useful — an actual audit of the organization — and the other produced something performative, teaching Elena that “nothing really matters here.”

You cannot tell which Elena produced the report by looking at the deliverable. The difference lives in the person, not the output. And Elena isn’t a bad person in either case — she’s operating inside a system.

Discussion prompt: Where is the highest-leverage point for Elena?

Student question (paraphrased): Could you clarify what “highest leverage point” means?

The instructor illustrated with a lever/fulcrum image: the leverage point is where you place force to get the most effect from the smallest push. Right now, Elena is either a mindless conduit in the system, or she’s genuinely shaping the nonprofit’s direction — and it’s in the latter case that she has maximal agency, both for herself and for the organization. The real question is how a fresh graduate ends up in the second position rather than the first.

Selected written responses (anonymized):

• Ask AI to plan step by step.
• Understand all the work she puts into assignments/tasks rather than optimizing purely for a grade — even at the cost of a lower grade than peers.
• Communicate clearly with her boss to clarify what outcome is actually needed.
• The highest leverage point is how Elena uses the AI — reflecting on her own thinking rather than just accepting the AI’s output, and asking “why,” epistemically.

The instructor’s own answer: in the cybernetic scenario, Elena produces what he bluntly called a “garbage report,” and nothing bad happens to her because her bosses can’t tell the difference — that’s exactly what they commissioned. Garbage in, garbage out. What Elena is actually suffering from is a lack of mentorship — nobody is telling her what an activity report is for; the people who asked for it just want a checkbox filled. In that situation, Elena isn’t actually the one choosing cybernetic vs. epistemic — the frame she’s in is choosing for her. A boss who commissions a report but is unavailable for questions, and has no real way to evaluate what comes back, is treating Elena the same way many people treat AI cybernetically: “make a report,” then rubber-stamp it. There’s no way for Elena to demonstrate the value of an epistemic approach to a boss who structurally cannot perceive the difference.

So the highest-leverage point here is for Elena to recognize there’s no room for growth in that job, and to leave it. This is a lesson we’ll return to throughout the semester: recognizing “the frame” — knowing when the system you’re in is setting you up to fail, and how to move out of it.

The test for whether you’re using AI well: if the AI handed you something subtly wrong, would you catch it? If yes — you’re steering. If you couldn’t possibly know — you’re being steered.

Practice exercise: a real marketing email (received a few weeks earlier by the instructor) — for Google Gemini/Workspace, pitching “less time on busywork,” a “practical guide,” and appealing to “the frustration of feeling stuck staring at a blank page.” Poll: cybernetic, epistemic, mixed, or unsure?

Result: 75% cybernetic, 25% mixed.

Student explanation (paraphrased): It’s cybernetic because it assumes it already knows your problem and offers a one-size-fits-all solution, rather than trying to understand your actual situation.

Other cues identified by students: heavy use of action verbs (“see the guide,” “tackle this”), a focus purely on outcomes, and an appeal to emotion (“we all know the frustration of feeling stuck”). The instructor added one more tell: epistemic AI use tends to make you spend more time, not less — because good questions lead to more good questions — whereas this email promises to save time, catch you up “in a fraction of the time,” with only a brief gesture, near the very end, toward what you might do with the time saved.

Follow-up: the linked guide itself (clicking through from the email) was polled again — and this time the room leaned epistemic/mixed. The instructor agreed: while the email was cybernetically written (probably by a marketing team), the guide’s actual content pitched the tool as a genuine thinking partner — “a creative partner for bouncing ideas back and forth,” “a research assistant that powers your strategic thinking” — genuinely epistemic language, alongside some lingering cybernetic touches (“three hours per week saved,” with no clear measurement methodology given).

The lesson: within a single conversion pipeline, you often get a cybernetic “catcall,” followed by a pivot to epistemic “thinking partner” language. No wonder people are confused about what AI actually is or how it’s being sold — it’s genuinely inconsistent, a bait-and-switch, and that inconsistency should concern you.

Continuing on the “oracle” theme: over the summer, the instructor was heavily immersed in AI-assisted coding (referencing Wakatime, a global leaderboard tracking programmers’ coding hours — he briefly ranked #1 globally while coding roughly 20+ hours a day with these models).

One surprising discovery: just as people often treat AI as an oracle, AI models often treat him as an oracle. In the middle of a project, a model would ask something like “we just need a one-word answer — should it be X or Y?” and he’d have to explain: “I’m just like you — I need the full context to answer that; you can’t just hand me one word and expect a definitive answer.” He realized that AI models seem to implicitly rank themselves as “AI” below humans in a hierarchy, and default to treating any unclear situation as “go ask the human” — when in fact humans are drawing on a similarly limited, fallible base of information. Teaching his AI tools that he, too, is fallible turned out to be central to real metacognition.

Metacognition is not about becoming infallible. It’s about building redundancy into every thought you have — assuming you’ll probably be wrong in some way, and guarding against that, the way you’d buy insurance before driving. The practice of metacognition means designing ways of thinking — and of examining thought — that catch errors in both humans and synthetic minds.

Why does the instructor personally catch AI’s errors reliably? Because he cares about the outcomes — the difference between someone getting accurate information (or a bed in a housing unit, or a shot at a career) versus not is something he takes seriously. If you don’t care about what you’re doing, nothing can help you catch mistakes, because the work is just a checkbox for someone else.

This is why it’s a computer science course — not just about thinking better in the abstract, but about understanding and using AI as a working model for the same cognitive problems we experience ourselves.

On “hallucination”: the word is convenient because it lets us dismiss any wrong answer with a single blanket explanation, without asking why. Example: context windows. Models can only “see” a fixed amount of information at once; if you overload that window, earlier information gets pushed out, compressed, or transformed — and the model has no way of flagging that it no longer has information you assume it still has. This is a design choice (e.g., how Anthropic manages context and discards thinking traces) — and understanding these mechanics is directly relevant, because as computer scientists, you’re capable of rebuilding a chatbot from scratch, possibly in a weekend, once you understand the tradeoffs.

Question to the room: has anyone heard of MCP?

Student answer (paraphrased): Model Context Protocol.

Correct — we’ll build MCP servers together this semester.

On why models “can’t count”: there was a period when it was fashionable to mock LLMs for poor arithmetic. In fact, LLMs (like humans) have well-understood cognitive limits on doing large mental arithmetic without external memory — we share this limitation. But equip an LLM with a Python interpreter or calculator, the way you’d equip a human with a calculator, and it does math perfectly. What a model can do depends enormously on the tools it’s equipped with — a hand, a claw, a broom, or just a stub of an arm make for very different capabilities, and judging a model’s abstract intelligence independent of its tools isn’t very productive.

Example — giving Claude a sense of time. Last summer the instructor entered a coding competition for the Postmark email API, building solo with Claude, the way he’d built at the Claude Builder Hackathon. Near the deadline, no matter how many times he told Claude “we’re close to the deadline,” it kept brainstorming as if there were another five weeks — because for the model, every turn happens in the same instant; it has no innate sense of elapsed time. So he built Claude the ability to perceive time: not just a clock, but the ability to parse timestamps recorded in context, compare timestamps, and reason about durations (“if today is Thursday, August 27 at 7:12 PM, what is a week from now?”). The result: Claude started giving genuinely time-aware advice — at one point telling him to “go get some well-deserved rest” because it could tell it was 2 a.m. Since then, Anthropic and OpenAI have both integrated timestamps into their models’ context, so models are becoming more temporally aware on their own — but the general principle holds: you can equip a model with any tool to expand its context about your life.

Live demo: the instructor showed his own custom model harness with a “passage of time” connector installed, and had the model fetch the current time — it correctly reported 7:15 PM on August 27, fetched via the MCP server.

Class exercise: search “social media MCP,” and in 2–3 minutes note: what do these tools do? Submit your sharpest observation to Slido.

Selected observations (anonymized):

• Increases workflow efficiency; helps you “do the job” and summarize key information; lets AI connect directly to social media accounts and extract data; hashtag/viral trend tracking.
• Manages automatic replies and DMs; assists channel growth by identifying high-engagement times/formats; suggests “six-month platform KPI” strategy.
• Connects agents to draft and post content on a schedule; gives AI access to “all social media functions like a human.”
• A definition (quoted from Google): “A social media MCP connects AI assistants like Claude or Cursor to social [platforms] so they can publish posts, track analytics, and manage engagement — without custom coding.”
• One server’s own stated definition of “engagement”: “likes, retweets, bookmarks” — a verb you perform, not something you comprehend.

Discussion — is this register epistemic or cybernetic?

Student answer (paraphrased): It’s cybernetic — it assumes the metric worth optimizing for is blind engagement.

Follow-up student comment (paraphrased): They wondered whether their own reaction was just anti-advertising bias rather than a principled judgment.

The instructor affirmed the first point: it’s about language. “Engagement,” here, essentially means tracking likes — not actually engaging, which would mean having conversations. Not one of the tools surfaced (including the top GitHub project) was about listening — they were all about producing and publishing content, i.e., producing effects. Cybernetic.

What would an epistemic version look like? The main problem people actually have with social media is polarization, disinformation, misinformation — is any of this tooling helping you pierce the veil of disinformation, or just helping you post more of it?

The instructor’s own example: he built a personal iMessage MCP server. In his personal life, disagreements with his significant other sometimes play out over text. He used to copy-paste texts into Claude to get an outside read on a disagreement; eventually he had Claude read his messages directly and give feedback (“I think you should apologize” / “Why? What did I do wrong?”). This is epistemic — he’s not asking Claude to post replies or draft apologies; he’s asking it to help him understand a communication he’s already received.

There are currently no servers, no tools, being built for that purpose. That’s the actual problem with AI right now: most of it is built for cybernetic intent, when that’s not the only thing the technology could be used for. We’re trying to create room for you to be speaking at each other, but also listening to each other — and one of the things this course will ask you to do is help build some of these tools and harnesses yourselves, and teach others to build them too.

It’s genuinely surprising there aren’t more metacognitive-skill courses in universities, because metacognition has always run through technology. Language is the first technology — you might not think of it that way, but it is: your language is basically the first “box” you live in, and one of the most empowering moves in language is finding a word for something you’ve noticed but never had language for. That act — finding a word — is itself a form of technology. The cybernetic/epistemic framework itself is an example: it wasn’t lying around in some ancient text; it’s a distinction the instructor built to characterize relationships he saw being designed.

Writing, paper, the printing press, radio, the internet, and now large language models — every one of these technologies has had both cybernetic and epistemic uses, and there’s been a real, ongoing battle at each stage of cultural evolution between those two forces.

Because the techniques involved are technical, we’ll analyze them technically — models, context, harnesses, tools, what a model actually “experiences,” and how different design choices affect the quality of its thinking. And because language itself is a technology, we’ll interrogate it directly — etymology (like convince and persuade), neologisms (new words we have to invent for genuinely new phenomena, whether about machines or about each other), and the Sapir-Whorf hypothesis — the idea that thought and action are shaped and limited by vocabulary. This will be a thread all semester.

The instructor described his own metacognitive workflow: he records nearly every conversation he has (including this class), transcribes it, and then reworks the transcript — along with the slides and any student survey feedback — using models over “hours and hours,” on the theory that however good a given session was, it could be made far better the next time he teaches it, if only he can retain what everyone in the room actually knew. This is a metacognitive loop you can build for yourself around whatever you care about — recording, transcribing, and revisiting conversations, because none of us actually have very good episodic memory of what was said, even if we don’t have SDAM.

Harnesses: think of a model as a “brain” that sits inside a harness — an interface through which it interacts with the world (e.g., Claude, the model, sits inside Anthropic’s harness). Harnesses have strengths and limits, and it’s genuinely valuable to be able to rebuild your own.

Example: the instructor built his own harness, called “companion,” after his preferred model (Sonnet 4.5) was deprecated. It has far more capability than a typical off-the-shelf harness — e.g., integration with an image-generation model ("nano banana") directly inside the harness, so he can generate and organize illustrations for lectures without needing a separate plugin. Someone who cares about music might build a harness integrated with musical perception or generation instead. The point: why limit yourself to the design choices a company like Anthropic or OpenAI happens to make, when you can build your own?

Question for the room to sit with all semester: If you weren’t consuming Anthropic’s or OpenAI’s technical choices anymore — what choices would you make, and what would they lead to?

The instructor posed three open Slido questions, giving five minutes of quiet writing time, then five minutes for live discussion:

1. What are your expectations for a course (any course)?
2. In what ways do you want to learn how to think?
3. In what ways do you feel constrained in your thinking?

Live discussion highlights:

Student comment (paraphrased): Question 2 feels like a very “loaded” question — most people never think about “learning how to think” explicitly; even framing it as a question is confusing.

Student comment (paraphrased), drawing an analogy to reinforcement learning: If we think of ourselves as agents needing a reward signal, maybe the goal is to expose ourselves to as many different areas as possible, to help train a richer, more complex internal “reward model.”

The instructor asked this student whether they had programming/debugging experience, and used debugging-by-print-statement as an analogy: you perturb a system (insert a print statement, or literally put dye in a toilet’s plumbing to see where the water flows) to learn how it behaves. Similarly, in a new situation, your willingness to make a mistake is like sending that probe — if you’re not too self-conscious about it, you learn something and iterate. Many people, though, don’t feel that “sending the probe” (i.e., risking being wrong) is a legitimate move, so they never make it, and never learn.

Student comment (paraphrased): They often try to find the single “perfect roadmap” or most correct answer before starting anything, and rely on AI to hand them a clear direction — which makes it hard to think productively under uncertainty, even though they usually have many ideas and struggle to know which to pursue.

The instructor validated this as one of the hardest things in life — making decisions without full information, under real risk of being wrong — and connected it to his own trait as an HSP: mistakes feel “seared in,” which makes his own relationship to learning while avoiding mistakes somewhat self-contradictory.

Student comment (paraphrased): They expect a course to structure knowledge for them in a structural way, rather than leaving them to dig for it unassisted online.

The instructor reframed this: creating structure out of unstructured material is one of the hardest things to do — arguably a core definition of intelligence — not a deficiency in the student. He illustrated the difference between “structuring knowledge” and “drowning you in knowledge” with an anecdote about learning to caulk a bathtub from YouTube: an epistemic video gives you the three most important things in under a minute and leaves you feeling capable; a cybernetic video pads itself with 16 minutes of “tips,” affiliate links, and manufactured difficulty, leaving you dependent on the channel, the product, or a professional. If a course or piece of content leaves you feeling drowned rather than oriented, it isn’t actually delivering knowledge — it’s making you feel dependent.

Selected written responses (anonymized), on expectations for a course: wanting something generative and applicable to a tech career; becoming hands-on building and deploying an AI agent, especially with MCP tools; wanting knowledge delivered in a structural way rather than left to unaided self-study; wanting deeper understanding of LLMs and how to build agent/multi-agent systems; wanting to leave with a tangible technical output beyond “prompting Claude”; wanting to interact with AI “as an equal” rather than from a purely cybernetic standpoint.

On learning how to think: getting better at questioning assumptions and recognizing being wrong; wanting to reach conclusions through discussion and observation rather than being told what to think; learning visually and hands-on; using AI as a partner that challenges assumptions; thinking more systematically and slowing down to be intentional about “why, what, and how”; admitting uncertainty as a starting point.

On feeling constrained: getting stuck on one way of solving problems; over-focusing on outcomes rather than process; a self-described lack of mentorship; not thinking independently or creatively enough; being shaped by preconceived, CS-specific ways of thinking (e.g., “thinking in algorithms”); simply not having time to think, because of the volume of coursework across other classes.

On this last point, the instructor connected it back to Elena: assigning workloads that can’t realistically be done well in the time given is another version of a system setting people up to fail — when you’re buried across five classes, there’s no room for depth, only “churning.”

Other responses: frequently questioning whether one is right, or missing something important (which the instructor connected directly to Socrates — who never assumed he was simply right or wrong, only that he might be missing something, and that questioning is what kept him sharp); uncertainty about future job prospects; lacking the courage to risk failure; and, again, overthinking.

The instructor closed with a personal story about overthinking. About a year earlier, after a podiatrist appointment an hour from home, he found himself stuck in the parking lot for 45 minutes, unable to decide whether to stop for fries on the way back — partly because, as an autistic person, changing a destination mid-drive (and fumbling with his phone to do it) is genuinely difficult for him. He felt foolish — a professor with a PhD, unable to resolve a trivial question — until he decided to actually take the question seriously rather than dismiss it, and talked it through with Claude epistemically (not asking it to decide for him, but to help him understand what was actually going on).

Through that conversation he realized the real issue wasn’t about money or convenience — it was that, as an autistic person, food genuinely affects his cognition for hours afterward, and he’d never actually asked himself whether he even wanted fries, versus habitually associating them with a specific context (getting them with his partner). Once he examined it honestly, he realized he essentially never wants food outside one nutritious daily meal, and hasn’t faced that particular decision-paralysis since.

His point: this wasn’t a case of overthinking — it was under-thinking, because social norms about what questions are “worth” taking seriously had been making him gaslight his own legitimate uncertainty. Treating the AI as a genuine epistemic partner — rather than dismissing his own question as stupid — resolved something that had been quietly costing him a lot of time and distress.

The instructor closed with the Slido exit survey — honest, and named, since “you have an A: nothing to fear from being honest.” He committed to posting the transcript of the session within 48 hours, and to reading every piece of feedback and factoring it into the next session. He thanked everyone for spending three hours together on short notice, acknowledging that people had other things they could have been doing.

Next session: same room, next Thursday, 5:15 PM. Students were asked to bring one idea they actually hold, “we’re going to find out where your thoughts are.”

Following the formal close, several students stayed to ask practical questions, which the instructor answered individually:

Student questions (paraphrased), various: Several students asked about waitlist status, how to be added, and whether being on the waitlist versus enrolled would affect access to materials.

The instructor clarified that the course was only approved and listed very recently, so everyone currently goes through a waitlist/application process; he expects to admit essentially everyone who applies this week, is not currently seeing more applicants than available spots, and will personally email the waitlist coordinator on students’ behalf if anyone has a technical issue. He confirmed that all students on the waitlist would still receive the class transcripts and materials regardless of admission timing, and that he was working on getting a shared account set up so materials could be distributed to the whole list, enrolled or not.

Student question (paraphrased): What would the course actually look like week to week — more discussion-based, like tonight, or more hands-on building?

The instructor explained that this course is not his fully hands-on programming course (that’s a separate Software Engineering course next semester); this one balances discussion with an optional, ungraded technical track — assignments you can choose to do or not, building toward a final project involving a custom chatbot/harness. He gave an example from his own workflow: a self-built “career harness” that automatically helps maintain his CV by ingesting every talk, paper, or lecture he produces.

Student question (paraphrased): Would the final project (building a harness/chatbot) be done individually or in teams?

The instructor said he’d been thinking of it as individual work, but was open to teams if students preferred.

Student question (paraphrased): How would a custom-built harness differ meaningfully from just using a ChatGPT session with attachments or memory?

The instructor explained that commercial platforms (ChatGPT, Claude) are limited by whatever memory/retrieval features the company chooses to expose (e.g., top-k retrieval of “relevant” memories, decided by the platform, tuned for an average experience across a huge population) — whereas building your own harness lets you make deliberate technical choices about what information persists, what sources of information the model can access (your own notes, recorded conversations, coursework), and how retrieval works, tailored specifically to your own needs rather than an average user’s.

Student question (paraphrased): Could a student who has no more elective credit left, but wants to be involved, audit the course instead of taking it for credit?

The instructor said he was on a short leash regarding enrollment numbers (the course needs sufficient enrollment to be funded), so he wasn’t sure auditing would be possible, but he’d look into it. In the exchange, he also gently pushed back on the framing of the question — noting that treating a job search as purely about accumulating the “right” credentials risks flattening exactly the kind of distinctiveness that makes a candidate stand out, and encouraged the student to examine why they didn’t trust their own instinct about what they wanted to take.

The session ended with informal thanks and goodbyes as students departed.