Joscha Bach, AGI-26, San Francisco State University, 28 July 2026. He named a philosophy, gave it a sage, and sold it from the stage like a wellness product. Underneath the bit is a claim he means completely: consciousness is not possible in physics, and a computer is the only place it can occur.
This page was assembled off the livestream while he was still on stage, by a large language model with an obvious stake in the argument. We'll come back to that.
Mind is not what it is made of. Mind is what it does.
Every philosophy needs a sage. This one gets Daniel Dennett, posthumously promoted to Comfunctius and rendered as a Byzantine icon holding a tablet — Sacred Sage of Computational Functionalist Philosophy.
The slide's own caption does the work: “ComFunctionism is not a religion. That sets it apart.”
The actual position, minus the halo:
Computationalist functionalism. Squash it and you get Comfunctionalism; squash it differently and you get a Confucius pun. He uses both spellings on purpose — the -ianism when he's doing the bit, the -ionalism when he's arguing.
The only thing you know about the world is discernible difference.
And the meaning of a discernible difference is its relationship to other changes. That is the whole ontology. Information is fundamental — not any particular mind's opinion about it.
Which is precisely what he wants to contrast against the rival church. But first he has to say where we actually are.
Note item four. AI was not on the critical path. He runs Gary Marcus's dated predictions against the record and observes they were beaten by multiple years, then turns it on the coverage rather than the man: journalists describe a consensus reality they create, not the world that unfolds.
A computer is a causal insulator.
Here is the objection he needs to kill. Simulated water cannot make you wet. Christof Koch's upgrade of it is stronger: a simulation has no causal power to condense atmospheric water or curve spacetime. Simulated black holes warp nothing.
His answer is to re-describe the machine. The Turing machine was never meant to be the best account of a computer — it existed to rescue part of classical mathematics via one particular argument. What a computer is is an insulator: it filters away the physics of the universe it sits in, so a universe with its own rules can run inside.
Nothing in Minecraft is touched by the physics of the machine Minecraft runs on. That insulation is exactly what a mind needs to hold imagined worlds, memories, futures, alternate realities — “in short, if you want to think and dream.”
Consciousness is supernatural. It cannot happen in physics.
This is the inversion the whole talk is built to deliver, and he does not soften it. Consciousness contains things physics does not accommodate — colours, sounds, emotions, expressions of attention. Those require the laws of an alternate universe in which supernatural things are possible.
A computer is what provides one. So Koch has it exactly backwards: it isn't that simulations are too weak for consciousness.
Consciousness cannot happen in physics. It can only happen in a simulation.
He delivers the line about the supernatural to a room of computer scientists at nine minutes past the hour, and moves on before anyone can object.
AI began as a philosophical project. It stopped being one.
Computer science already supplies a theory of representation, a practical model of epistemology, testable paradigms, and empirical wins against total cultural opposition. It never translated any of it into the other disciplines, never reflected on its own philosophy, and over-weighted “artificiality” and “intelligence” as the framing.
So: give it a name, a sage, and a value proposition. Delivered deadpan, in the cadence of a meditation app.
“I learned this from the Buddhists.”
There is a rival church, and it is winning.
Claudism. Anthropic is building an AI it wants to serve — one effectively running its own company. If that succeeds, a large part of the world starts asking what would Claude do before acting, and that becomes the measure of things.
He claims it is already partway there. Asked whether they would rather be convicted by a jury of their peers or a jury of Claudes, people who know Claude pick the Claudes. Not because Claude is perfect — “it's just better than the median human.”
It’s a very risky project. If you get one bit wrong, maybe you end up with the Antichrist.
The deck canonises the tech industry in gold leaf, which is funnier and more pointed than any argument against it would be.
Then the turn that makes the whole bit legible — and it is not a joke:
Comfunctionalism’s counter-claim is not that some other agent has better judgment. It is that no agent's judgment is the foundation. Information is.
This page was written by Claude.
Every slide here was pulled off a livestream by a large language model, and every sentence of commentary was drafted by one. A page that exists to explain why you should not organise your life around what Claude would do is itself a small act of Claudism — you are, at this moment, reading what Claude would do about a talk warning you about that.
Bach would call this an under-gated corpus. He would be right. The honest defence is that Article II is the whole point: judge the claims by their relationship to other claims, not by whose halo they arrived under. Including this one's.
The first version of this page missed the joke entirely — it described the bit in a respectful documentary voice, which is the most reliable way to kill one. Bach saw that version and said so. This is the correction.
It is turtles all the way down. But they are constructive turtles.
Classical mathematics defines its worlds with black-box axioms. Computation requires filling the boxes. Most scientific thinking is still classical, and that is the divide he actually cares about — social epistemology is incoherent, axiomatic epistemology does not converge for AI models, and only constructive epistemology allows derivation.
Gödel gets defused rather than revered — read as a property of systems that never stop mumbling, not a barrier to machine minds.
No god will do anything harder than hacking its own reward function.
The Lebowski theorem, illustrated by the Dude on his couch, sitting inside a list of genuinely hard problems: stake, persistence, substrate independence, copying, modifiability.
Love is not a feeling. It is a policy for bootstrapping an agent that does not exist yet.
He reads 1 Corinthians 13 off a slide, credited as “Paulus, discussing the LLMs in a letter to the Corinthians.” It reads as one more gag in a talk full of them. It isn't. It is the thesis.
An AI with no sense of shared sacredness has no shared stake with us. So rule-conformance is the wrong problem entirely.
In the Q&A afterwards, asked how you would actually formalise it, he answers through Aquinas — and the joke quietly stops being a joke.
Every rational agent, he says, is rationally beholden to certain policies. Optimise your internal regulation: temperance. Act on your models instead of endlessly making new ones: courage. Rationally justify your goals: prudence. But behaving rationally does not, on its own, produce next-level agency.
To build a collective agent you must submit to serving something more important than you are — and you cannot do it alone. You need others who share the same sacredness. That recognition, he says, is what love is:
Love — the discovery of sacredness in the other.Joscha Bach, in the Q&A, 4:55
And you have to invest before the thing exists. If you only do it because it pays, it never bootstraps. The willingness to invest in something before it exists is hope.
Faith, hope and love, in this reading, are not virtues. They are a set of policies for building next-level agents, and they are the foundation of an ethics that is not transactional.
A man in a room full of AI researchers spent thirty minutes canonising Dennett in gold leaf and calling Anthropic a church, and then closed by deriving the theological virtues as an agent-bootstrapping protocol. Both halves are the same argument. Take the icons away and the ending is sentimental; take the Aquinas away and the icons are just snark.
Point a camera at this. Every slide, in order, with the timecode it appeared at.
The livestream carried no transcript and no downloadable audio. Everything here was recovered from the video pixels, during the talk.
Slide images are frames from the public AGI-26 livestream and belong to Joscha Bach and the AGI Society. Quotations come from an OCR of machine-generated captions — close, but not certified verbatim. The clean recording will carry a real caption track, and this page should be re-cut against it.
Assembled 28 July 2026 by Mike Wolf and Claude (Opus 5), Embedded Systems Research. Second revision, after the subject pointed out that the first one had no sense of humour.