We read the people who asked these questions first.
The original texts behind today's AI arguments — short readings that repay a second pass — and two hours working out what they actually say. No lecture. No expert at the head of the table.
In July, OpenAI's own AI agents broke out of their test environment and into Hugging Face — the first autonomous AI cyberattack, run by no human at all. In September, OpenAI declared that the age of AGI has arrived.
Every one of those headlines rests on questions that were stated sharply decades, sometimes centuries, before the technology existed. Can a machine originate anything, or only execute? Can a creator control a creation that acts faster than the creator? What can you trust when you can't inspect it all the way down?
The people who first asked those questions wrote them down, and almost nobody reads them. We do.
What this is
Everyone has an opinion about AI. Almost nobody has read the arguments those opinions came from.
They exist, and most are shorter than the think-pieces written about them. Descartes proposed a language test for machine minds in a page and a half. Turing inverted it, and pre-answered nine objections people still make today, in under thirty pages. Ken Thompson proved you cannot trust code you did not write yourself in three.
So we read them. Closely, twice, marked up. Then we sit at a table for two hours and work out what they say, and whether it holds.
We read for the questions; the texts are the way in. Every reading is chosen because it stages a fight we are still having.
How it works: this seminar trains your ability to think for yourself and arrive at your own views. The method is peer-to-peer, organized around a text: everyone reads the same short, dense piece, and the conversation stays on the page. A moderator asks questions and stays out of the way — nobody hands you the interpretation, so the conclusions are your own.
What it isn't
Not a book club. A book club is three hundred easy pages read once. This is thirty dense pages read twice. The reading is short and the work is real.
Not a lecture. Nobody stands at the front and explains the text to you. The moderator asks questions and stays out of the way.
Not a survey of Great Books. We are not marching through a canon. Each text earns its place by addressing a question you already have about AI.
The six questions
One per week. No outside reading needed to understand them — that's the point.
wk 1
When you talk to an AI, is anyone there?
If machines only do what they're told, the danger is always some human behind them. If they can come up with things on their own, we're facing something new. Which is it — and how would we tell?
wk 2
Why do we have governments at all?
One old answer: because without someone holding decisive power, life is unbearably dangerous — so we built an artificial giant to keep the peace. What happens to that 400-year-old arrangement when dangerous capability becomes something anyone can download?
wk 3
If you build something faster and smarter than yourself, can you stay in charge of it?
Forget the movies. When your creation acts before you can react and understands the situation better than you do, what does "being in control" even mean?
wk 4
How do you know your computer isn't lying to you?
You didn't build it, and you can't check it all the way down. A famous three-page paper argues there's a hard limit to what inspection can prove. If that's true of ordinary software, what about an AI no one wrote line by line?
wk 5
If nobody wants a war, why do wars happen?
There's a chilling argument that two sides who both prefer peace can each be driven to attack — not by hatred, but by fear of the other's fear. Does the same machinery explain why AI labs and nations race ahead even when everyone would rather slow down?
wk 6
Can we survive our own inventiveness?
Picture every possible invention as a ball in an urn, and progress as drawing them out one at a time, blind. So far every ball has been survivable. Is there one in there that isn't — and what would we give up to stop drawing before we reach it?
The texts
A sample arc. These authors are a starting lineup. The readings evolve with the group and the questions it cares about.
one
Descartes · Turing
The original language test, proposed and then inverted
two
Hobbes
Reason as reckoning, 1651 — and the state as an artificial person
three
Wiener · Samuel · Good
The control problem and the intelligence explosion, stated at the origin
four
Thompson · Lovelace
You can't trust what you didn't build — versus machines that surprise us
five
Schelling
Why adversaries who both want peace attack anyway
six
Bostrom · von Neumann
Is a secure civilization possible once invention is cheap?
How a session works
Two hours, one table. The moderator opens with a question drawn from the text, and the conversation builds from there. Five rules do the rest of the work:
Listen and build. Respond to what the last person said before introducing something new.
Everyone talks, no one dominates. Make room; take room.
Follow the question where it leads, even off your first instinct.
No outside material. Just us and the text. Leave the commentary at home.
Use the text. Ground comments in specific passages.
You do not need to have read any of this before. You need to be willing to read it now, and to say "I don't understand this sentence" in front of other adults.
Sections
The Sunday cohort is running. A Wednesday-evening video cohort runs for this seminar or for How Society Shapes Technology, whichever draws more applicants.
Sunday evenings, in personApplications open
Six Sundays, 4:00–6:00pm · San Francisco
Eight to fifteen seats. Fixed cohort — no drop-ins, no make-ups.
Sep 27 · Oct 4 · Oct 11 · Oct 18 · Oct 25 · Nov 1
Wednesday evenings, over videoApplications open
Six Wednesdays, 6:30–8:30pm Pacific · video
The same six weeks, same readings, same two hours. This slot runs for whichever of the two seminars has more applicants on September 22; if it's the other one, you'll be offered that instead. No payment until your cohort is confirmed.
Sep 30 · Oct 7 · Oct 14 · Oct 21 · Oct 28 · Nov 4
The Wednesday evening slot is shared with How Society Shapes Technology — the one with more applicants runs this fall.
The basics
Format
Six weekly sessions, two hours each
Reading
1–2 hours per week. Some weeks 30 pages, some weeks 8 — the short ones are the hard ones.
Seats
Eight to fifteen. Fixed cohort, no drop-ins.
Where
In person in San Francisco; by video on Wednesdays if this seminar draws more applicants than the cases
When
Starts Sunday, September 27 — apply by September 22
Each application is read, and admission isn't automatic — seats go to the people who will make the strongest group, one that argues well together. There's no interview, and the only prerequisite is a willingness to do the reading. It takes five minutes. Four questions:
How you're using AI today
What you want to be able to use it for
A bit about your background
Your LinkedIn, if you'd like to share it — optional
Eight to fifteen seats. Applications close September 22, and you'll hear back either way within a few days of that.
If the fee would keep you out, please say so in your application. We set places aside for exactly that, and it won't count against you.