Private
Public Access
0
0

conductor(generic_systems_fields): Phase 1 Acquire - transcript (885 clean segments, 30KB) + 58MB mp4

This commit is contained in:
2026-06-21 23:16:13 -04:00
parent 900b68009b
commit 99e955795f
4 changed files with 13217 additions and 0 deletions
@@ -0,0 +1,11 @@
Phase 1 Acquire for generic_systems_fields: https://youtu.be/QeMajYvhEbI
Artifacts: C:\projects\manual_slop\conductor\tracks\video_analysis_generic_systems_fields_20260621\artifacts
Step 1: extract_transcript (yt-dlp VTT directly)
OK: wrote C:\projects\manual_slop\conductor\tracks\video_analysis_generic_systems_fields_20260621\artifacts\transcript.json (1751 segments)
Step 2: download_video
OK: wrote C:\projects\manual_slop\conductor\tracks\video_analysis_generic_systems_fields_20260621\artifacts\video.mp4 (60755152 bytes)
{
"status": "ok",
"video_path": "C:\\projects\\manual_slop\\conductor\\tracks\\video_analysis_generic_systems_fields_20260621\\artifacts\\video.mp4",
"transcript_path": "C:\\projects\\manual_slop\\conductor\\tracks\\video_analysis_generic_systems_fields_20260621\\artifacts\\transcript.json"
}
File diff suppressed because one or more lines are too long
@@ -0,0 +1,885 @@
I want to talk today about interesting
behavior by generic systems.
And the first part of this will be a bit
of a review and then the second half
will be more u a report on new work uh
since the last time I talked which
uh as a general presentation was over a
year ago. I did a couple of more
specialized presentations last year. So
this is in the context of the diverse
intelligence project
which raises these questions u how
diverse is intelligence
and what what are the limits of
intelligence uh if there are limits and
how do we find out so that's what I'm
going to talk about today
uh to start with um
we can look at this definition of
intelligence that we've been using in
the lab from William James as as kind of
a guideline
that intelligence is a fixed goal
achievable with variable means of
achieving it. So intelligence involves
ing it. So intelligence involves
some level of flexibility
and this definition itself raises
questions. Uh does any goal count? Uh
are many any kinds of means allowed
and does anything fall outside this
definition? Are there any are there
systems that don't achieve this kind of
flexibility that James was interested
in? And um it's not clear from from
James the extent to which he was willing
to extend intelligence
uh to generic systems,
but at at least some of his writings
suggest that that may have been the
case.
Um
some sort of methodological guidelines
uh have appeared recently. Uh, for
example, in this paper that Mike and
David Resnik published last year,
uh, emphasizing that the
diverse intelligence project or the the
TAM framework,
um,
emphasizes methodologically
a reliance on empirical research, not on
intuition for thinking about
intelligence.
and um an insistence on testable methods
for understanding uh what is and is not
intelligent and the extent to which
things are intelligent. So I I want to
follow this kind of empirically oriented
framework as opposed to making
philosophical assumptions upfront.
And in thinking about this um an obvious
place to look is the literature of the
free energy principle.
So here's um a recent paper from Carl
Fristen's group
um their most recent actual technical
elaboration of the free energy principle
and they describe it as describing a
simple relationship between the dynamics
of random dynamical systems and
inference
and that of course has been the theme of
the free energy principle for almost 30
years now that random dynamical systems
uh can be considered inferential.
But when you look into this paper
um here's figure two
already we find an assumption that
there's some systems that are inert that
don't act on the world at all.
And that's a problem because a system
that doesn't act back on the world when
the world acts on it is violating
Newton's third law. So it's it's
violating a very basic physical symmetry
principle.
So
this picture uh at least appears to
involve uh intuitions about inertness.
Uh and the the trademark inert system is
a rock uh in much of this literature
that may not stand up to scrutiny from
the point of view of fundamental
physics.
So uh in that case I think we need to
take an approach where we look at
physics first and worry about intuitions
later.
So if we have a system which I'll call a
and it's interacting with something else
which I'll call a bar or the complement
of a for reasons that will become clear.
How do we guarantee that this
interaction respects all of the physical
symmetries that it needs to respect to
comply with known physical theory or
empirically supported physical theory.
And what we want is a generic case that
describes any interacting system
regardless of its scale or its
structure.
And um I'll add here since
the whole talk will respect this uh
regardless of its embedding in spacetime
what it looks like as a as a thing in
spacetime.
So how do we do that? And the answer is
we start with the generic case.
So let's assume the simplest thing we
can assume
which is a system that doesn't interact
with anything. So a system that doesn't
have an environment.
Um so a system that's isolated.
And if we start with that, um, we know
that we're going to respect the various
symmetries that have to do with not
having singularities because there's no
there's no place for information to flow
from into this system and no place for
information to flow to because there's
no environment. So there aren't going to
be sources or syncs.
So whatever dynamics this system has, it
has to conserve momentum and energy and
information.
U so
the technical term for conserving
information is unitarity
and it turns out that this is a very
productive assumption.
If we have conservation of information
in other words unitarity
then we can represent the dynamics as a
linear operator
um because nonlinearities
uh don't conserve information.
So this dynamics that I've called P of U
the propagator of U is a linear operator
on some state space. And in fact we can
make the state space a Hilbert space
which is a particularly simple kind of
vector space. It just takes every
possible value of every possible degree
of freedom and makes it a basis vector.
So if we assume some background time t
which is just a symbol that lets us talk
about change.
Uh we can write this propagator
uh as a periodic function of another
operator which represents the energy of
the system. That operator is the
Hamiltonian
which plays essentially the same role
here that it plays in classical physics.
It's a measure of energy
and when we write this equation um it
requires introducing a constant which is
finite and which has the units of action
and a finite value of this constant
um corresponds to there being no no
singularities. Clearly if that symbol h
bar was zero
there would be a singularity and if it
was infinity there would be a
singularity. So it needs to be finite.
Now this theory is quantum theory
of an isolated system.
So in a sense isolation is all you need
to get quantum theory. Uh it's a it's a
good way to start because it gets you
someplace that we understand
and that we have empirical reasons to
think is a good description of generic
systems.
So if everything is linear, the dynamics
is linear, we can do a linear
decomposition of the dynamics.
So we can split this isolated system U
into the components that we're
interested in some system that we want
to talk about uh A and and its
environment now A bar which is just
everything else in U that isn't A and
that introduces the idea of a boundary
between the two. Um and the linear
decomposition is a decomposition of this
operator h the Hamiltonian
and we can decompose it into the sum of
three terms. A term for the left side, a
term for the right side and a term for
the interaction.
So a nice thing about quantum theory, it
has this simple representation of
interaction and it's linear.
So this is a generic symmetry preserving
representation of interaction between a
system and and everything else.
Now we can continue decomposing this
interaction term again additively
um as a sum
of operators that act on single quantum
bits. And some some of you have seen
this kind of picture before in uh
previous talks, but these operators are
very simple. They're just operators that
measure the spin of a quantum bit. And
each one of them is equipped with a
reference frame that says what direction
counts as Z since this is a Z spin
operator. And Z can be chosen to be
anything. and it can be chosen uh in
different ways for different cubits.
But this turns this boundary B into a
holographic screen. So there's a lot of
physics that can now be imported to talk
about this boundary. And I'm not going
to really talk about that physics today.
Uh we're just going to go on.
Um this Hamiltonian now tells us how a
and a bar act on each other. they act on
each other by changing the values of
cubits.
Uh what we wanted though is to
understand how a and a bar influence
each other. How information flows from
one to another.
And these are the same uh if and only if
one condition is met that a and a bar
have conditionally independent states.
So that we could talk about the state of
A and the state of A bar.
And that's a simple requirement. We have
to require that the joint state the
state of the two systems factors into
the state of the one system and the
state of the other system.
And this is in quantum theory called
separability. And it's the same as the
absence of entanglement.
So systems
have their own states. if they're not
entangled.
And entanglement just means factoriz or
non-entanglement just means
factorizability.
And of course that maps over into
classical systems. If classical dynamic
system has a markoff blanket then the
then it factors
and we know what separability requires.
It requires sparse coupling.
So formally this interaction has to have
a small dimension and since it's an
operator its dimension is well defined
and by small it means its dimension is
much less than the dimension of either
of the other operators in the picture
the internal operators for a and a bar.
So intuitively the evolutions of these
systems have to be almost independent
for their states to be separable and for
the interaction to actually capture all
of the influence of one system on the
other.
So what we've done here is reconstruct
the free energy principle from minimal
physics. And again u at least some of
you have seen this before.
If uh these two systems A and its
environment are separable, so they meet
this dimensionality constraint on the
Hamiltonians,
then the Hamiltonian fully describes
information exchange. The boundary
functions as a markoff blanket. Uh the
variational free energy is a measure of
the interaction strength.
uh minimizing VF VFE which is what the
free energy principle is about is
keeping the interaction weak while
allowing thermodynamic exchange.
So it's it's keeping the interaction
between the system and its environment
fairly weak while enviring allowing the
system to eat some of its environment to
to provide enough energy to live.
And in this case predictability which is
what the system is after is constrained
interaction. So it's constraining the
interaction in a way that's consistent
in a way that's consistent
with staying alive.
So ANA bar maintain their identities as
distinct systems only while their
boundary um remains a markoff blanket.
So there can't be any rips, there can't
be any explosions, other sort of huge
excesses in interaction strength.
So that's actually what the FE is about.
It's about keeping interactions small
enough to maintain the integrity of a
boundary and that's what's required for
persistence over time as a well-defined
system
and this follows uh just from very
minimal physical assumptions.
So now if we go back to Jane's Jane's
definition of intelligence as a fixed
goal with various means of achieving it,
we can ask um is there always a goal?
Does any goal count? And the FEP always
gives us one goal continuing to exist.
uh any system that
persists through time acts inferentially
uh as if it's trying to continue to
exist. This is what Jacob Howey calls
self-evidencing
as an interpretation of the FEP.
Are any means allowed? Well, yeah.
Whatever the internal dynamics are
capable of is an allowed means for
keeping the boundary intact, for keeping
the interaction uh weak but still open.
And does anything fall outside of this
definition? No. It's completely generic.
At least it's completely generic for
systems that are physically realizable.
um so systems that actually respect the
sorts of symmetries that our physics
requires.
So from this point of view, intelligence
looks like something that could be
generic.
The question then becomes not what's
intelligent, but is its behavior
interesting?
Um maybe everything is intelligent but
some things are completely boring. They
don't behave in ways that are
interesting.
So our question about intelligence
really turns into a question about
interestingness. Are there limits on
what kinds of systems can exhibit
interesting behavior? And if there are,
how do we find out? How do we find out
what the limits are? And how do we find
out what sorts of systems actually
display interesting behavior?
And if we're going to ask this question,
we have to know what interesting means.
So here's some ways of characterizing
interesting behavior. Um
so if it's interesting, we want it to be
surprising.
uh maybe it's unpredictable in practice
or only predictable approximate
approximately with some amount of coarse
graining. So if we don't measure what
the systems doing with too much
precision
um maybe we can predict what's going on
but if we really look closely it's it's
going to be uh less predictable.
Maybe the system is unpredictable in
um systems seem to be interesting if
they can learn from experience, if their
um behavior depends on their memory, if
it depends on the context they're in.
Uh and we can get really radical here.
some systems may have distributions of
outcome values that actually violate the
Kaggoro vacs. So the outcome probability
distributions are undefined. That's
really interesting.
But operationally,
what this all comes down to is that
if we measure a bunch of state
transition probabilities for some finite
amount of time, they don't converge to
predictive adequacy.
So induction uh doesn't work or doesn't
work very well or only works if we
severely coarse grain things.
So the sort of mechanical expectations
that we've inherited from 19th century
science are violated
and we know that life does this. Um,
if you observe a baby for a while,
you're not going to be able to predict
what happens as an adult uh to that
person or how that person behaves as an
adult. In fact, if you observe an adult
for a while, you can't perfectly predict
what they're going to do next. So, life
life doesn't satisfy these sorts of
mechanical expectations.
But what else violates them? Uh that's
what we're going to be interested in
finding out.
And we have some hints.
So here's a old hint from the 1950s from
cybernetics.
Edward Moore showed that finite input
output experiments can't uniquely
determine the the machine table which
just means the in internal state
transition probabilities
of a generic classical black box.
And um the canonical example of that is
a a box with a clock. So something like
a time bomb where the behavior can
change abruptly after some amount of
time. And if you don't observe it
through that change of behavior, then
you can't predict what's going to happen
next.
Here's another more recent hint from um
Conway and Koken. The free will theorem.
Here Conway is John Conway from the game
of life.
They published what they called the free
will theorem
which shows that special relativity and
quantum theory together rule out local
determinism
and rules out local determinism in a
very strong form. Uh the whole pass like
cone of a system can't determine its
behavior
if special relativity and quantum theory
are combined. So their slogan in their
first paper in 2006 was if experimenters
make choices electrons do too.
So that's a pretty strong hint that even
electrons can exhibit interesting
behavior.
And then a third hint uh is this paper
from Frank Tipler in 2014
where he showed that the simplest way to
remove the singularities
from classical physics
uh actually reproduced the quantum
potential postulated by David Bone in
his formulative formulation of quantum
theory. And basically what the quantum
potential does is make the motion of any
given particle dependantly
on the motion of every other particle in
the universe.
And interestingly uh Nicholas Jesus
pointed out that back in the day of
Newton and lelass
uh before the 19th century physics
wasn't singular because it was basically
about gravity
and gravity wasn't local. Every particle
did actually behave on depend on what
every other particle in the universe was
doing instantaneously.
And it was Einstein who introduced
strict locality by requiring information
to only flow at the speed of light.
So these hints sort of suggest that
generic systems can display interesting
behavior. So the question is how do we
make this precise? How do we understand
it? How do we use it? So that's what I'm
going to try to talk about in the
remaining time.
So here's the setting again. Uh we have
a system. It's interacting with another
system which is its environment. There's
a boundary between them that the
interaction flows through. So
observations and actions live there on
the boundary. U actions of A on A bar
and actions of A bar on A.
And u A's measurements and action
choices are computed by its internal
dynamics. That's true for AA bar also.
In FE language, that computation is done
by A's generative model. Same applies to
Abar.
The system satisfies this dimensionality
constraint. So we can also think of it
as the dimensionality of the boundary is
much smaller than the dimensionality of
either system.
But the thing to keep in mind is that
inputs and outputs, so behavior is much
less complex than the computations that
generate inputs and outputs. So behavior
is less complex than the computations
that generate it generically.
So this immediately tells us something
important
that recurrence of states on B.
So seeing the same behavior does not
imply recurrence of the internal
dynamics of the system that's exhibiting
the behavior.
And in particular, the probability of a
behavior given the internal dynamics is
not the same as the probability of the
internal dynamics given the behavior.
So behavior generically depends on
hidden states um which we can think of
as memory or internal context or
something like that.
And formally we can always represent
that as an internal geometric phase.
what's also called a Barry phase after
um can't remember his first name Barry
who first characterized it almost in
general for for adiabatic systems
and as as Chris Fuches the uh physicist
who in came up with the cubist
interpretation of quantum theory put it
all physical systems have interiority
and here he's also very influenced by
William James I think interiority is
actually a Jamesian term.
So what is this geometric phase?
It's it's just an apparent phase change
that results from transporting a vector
or some collection of vectors around a
path in an internal state space.
So here's an example um the right hand
side of the screen. If you have a sphere
and you start with a planer a
two-dimensional coordinate system up at
the north pole and you transport that
coordinate system smoothly down to the
equator and then you transport it a bit
to the east on the equator and then
transport it back up north. It's going
to look like you've introduced a 90deree
phase change in the coordinate system.
But nowhere in this process have you
done any rotation.
Uh you've just transported
uh this system in a in a parallel way
without changing the system at all
around in the state space but the state
space happened to be curved. So these
are called holonomy operations.
Uh and purists call them anholony
operations.
um but physicists just call them
holottomy operations and they're not
much talked about in biology
but there is this paper from 2026 from
Marcel Blatner
that applies this sort of thinking to
plenaria
and he does it in a framework he calls
tangential action spaces but if you go
back and look what he means by that he's
really just talking about uh holomy
transfer formations.
So this is this is work that's starting
to be applied biologically
but to see why it's important
um it's important because non-trivial
honomy is actually a provably sufficient
resource for universal quantum
computation.
So if you want to build a universal
quantum touring machine, you actually
only need one ingredient and that's
non-trivial holomy in the search space.
And it works because it allows you to
construct a map from an observable
boundary state, an input and an internal
state to some other observable boundary
state, an output and some other internal
, an output and some other internal
state. for arbitrary input and output.
. for arbitrary input and output.
So you can you can implement any
computation just with holonomy and this
has been known for 25 years. Uh but it's
only really been known in the quantum
information community.
So um
computing by holom is actually stayed
basically within the quantum computing
community.
So to see why this is important um or to
see its larger implelications,
I think it's important to think about
what a physically implemented
computation is.
Uh physically implemented computation is
just a mapping
from the behavior of some physical
system. The observable behavior. So the
observable of the behavior of the
boundary of that system
to an abstract representation of a
computable function.
And
saying that
um some device implements a computation
is just saying that these kinds of u
diagrams commute
where you're either mapping the behavior
into the symbolic representation of the
computation
or you're mapping the symbolic
representation of the computation into
the behavior by either a projection or
embedding.
And obviously this only works for
computable functions
because if you can't compute the
function f, you can't determine whether
these diagrams commute. So this is a
definition of implemented computation
for computable functions.
And this interpretation
is a projection uh into the behavior. So
that's important.
Because embeddings are injective.
Embeddings are one to many.
Um you can embed lots of different
computations in any given behavior. And
that tells us something that poly
computation is actually generic.
Um and indeed managing thermodynamic
flow. So keeping your system alive uh
requires that the informative behavior
that you're you're interested in
describing computationally is just a
proper sample of the total behavior.
U and that's true even for your laptop.
You know we're not looking at everything
the computer is doing when we're looking
at the screen particular. We're not
looking at all the thermodynamic
exchange that keeps the thing running.
And the same is true for organisms.
Uh we're not necessarily looking at the
caloric content of what they're eating
when we're trying to do their
psychology.
So um we can take this a little bit
farther. Turns out that we can always
think of computation as scattering in
some sort of space of data structures.
And some colleagues and I have a paper
on this that's been in review for months
now that goes fairly far and in fact
shows that the formal representation of
scattering can always be applied to
computation. But we can just use this in
a in a fairly intuitive way. If we think
of an ideal classical computer
implementing an algorithm for f then
we're thinking about something like
this. We have an input for f and we have
some ready state of this classical
computer. And if we combine the input
and the ready state, something happens
that is some sort of implementation of f
on the input.
And at the end of that something
happening, we get an output of f on the
input and the computer returns to its
ready state. So we can do it again.
Now this is a useful way to think in a
coarse grained way but the observed
ready state which is a projection of B
of the boundary of the device that we're
talking about does not actually pick out
a unique machine state.
Um,
that's what we just saw with this
business about holomy, but it's also
just Morris theorem from back in the 50s
updated.
Uh, we don't actually know what's going
on inside the machine. It's a black box.
Uh, it's got a boundary and we have to
look at what's going on on its boundary
and that doesn't determine what's going
on inside.
And we can see this in practice. Um even
a classical operating system
uh accumulates internal state changes as
we execute it over and over and over
again with different inputs. Here the
input is some program together with some
data
and
uh we have the machine implement some
program with some data and it produces
an output. Then we give it another
program and some more data and it
produces another output. We're not
keeping track of everything that's going
on inside. We just depend on the
operating system to to stay in some
usable state
uh without saying exactly what that
state is. And after a day's use of your
computer, the final state may be very
different from the initial state. And
eventually the computer has to be
rebooted to get it back to something
like the initial state that you started
with today.
So we can think of that in in this
language of side projects. Uh the
operating system is constantly doing
side projects. It's it's rearranging its
internal memory. It's moving things
around in its long-term what used to be
a disk memory.
um it's um changing its internal state
in various other ways. So this is
inevitable in generic systems uh because
generic systems
u exhibit non-trivial honomy because
they have big state spaces and more
complicated dynamics than what they
display on their boundaries.
So we can also represent these sort of
phase changes geometric phase changes as
reference frame changes.
So for example u go back to that picture
of a generic interaction that shows
operations on cubits by spin operators
each of which has a local reference
frame.
If the system undergoes some honomy
operation
uh that introduces a phase into a vector
that's transported to some new part of
the state space
and we think of what that does to these
reference frame vectors
um
that tell the spin operators what
they're doing. Then a general holomy
transform can modify a reference frame
and in that case the output of acting
with that operator is different. So if
the reference frame is up and you act on
something with spin up stays spin up.
But if I've tilted the reference frame
then what I do is rotate
um that cubit. So it when that cubit
gets measured by another system you're
going to get a different answer
and in fact this representation
of holom as reference frame change turns
out to be generic
and um I'm involved in another ongoing
paper that's trying to sort out um how
this works in general
and Again, it involves many different
literatures that don't communicate with
each other very much to put together a a
real understanding of how holomy change
relates to reference frame change
and it all turns out to couple to the
theory of error correcting codes and
hence the theory of emergent spacetime.
So this this issue of geometric phase
and reference frame change is is
actually very deep. But all we need to
know for for now is that if the system a
bar exhibits non-trivial honomy then how
it acts on on the boundary so that the
output that it produces its behavior
it's going to change um as honomy
operations move vectors around in its
state space.
So this is important um because when you
move reference frames around
you end up with uh behaviors that don't
commute.
So in particular uh this simple case of
spin operators
uh spin operators with different
reference frames don't commute. Acting
with up and then sideways is not the
same as acting with sideways and then
up.
And this has a an overall global
consequence. Whenever you have
non-commuting reference frames,
uh they generate non-causal context
dependence in the behavior of the
system. Uh and if you're you're making
measurements with non-commuting
reference frames, they generate
non-causal context dependence in your
measurement outcomes.
And in either case, the consequence of
that is that joint probability
distributions on the outcomes are
actually undefined. They violate the
Kagoro axioms. And this is this is what
is observed in systems that violate for
example Bell's inequality that exhibit
entanglement
uh systems that violate the leot
inequalities and so um exhibit what
looks like entanglement in time.
uh systems that violate the Coke and
Specker
uh theorem
uh which is essentially a restatement of
the conditions for entanglement.
Uh so exhibit context dependence
depending on different orders of of
observational
actions yield results that can't be
given an overall joint probability
distribution. So there's tons now of
experimental data that it demonstrates
this sort of
kagorov axiom violation by physical
systems
and I've put down here a reference to a
paper that talks about uh this
non-communing QRFs generating non-causal
context dependence in general.
So what does this mean? It means that
all of these interesting kinds of
behaviors are in fact generic.
And they're all generic for the same
reason. That they're all the result of
requiring that the behaving system is
separate from the system that's exhibit
that's observing its behavior.
Uh and that separability requirement as
we saw
uh induces this dimensionality
requirement that we have big systems
with small boundaries and so behavior
has much less dimensionality than the
computation that generates it. Whenever
that condition holds
uh then we see these sorts of
interesting behaviors
and that condition holds whenever we can
separate ourselves as observers from
whatever system it is we're observing.
So this tells us in fact
that intelligence is very diverse and in
fact interesting behavior is very
diverse. Um it also provides us with a
very general and strongly empirically
validated foundation for talking about
diverse intelligence i.e. quantum theory
and it tells us that intelligence and
persistent observability so separability
between the observer and the observed
uh go hand in hand.
uh if if a system is different from us,
we're observing it, then its behavior is
going to be interesting.
And we can make that interestingness go
away by coarse graining or averaging or
the common practice of throwing out
anything that looks anomalous,
but it's actually there. And if we look
carefully enough, we'll see it.
So that's it. Thank you.
@@ -0,0 +1,16 @@
# yt-dlp log
# url: https://youtu.be/QeMajYvhEbI
# output: conductor/tracks/video_analysis_generic_systems_fields_20260621/artifacts/video.mp4
# returncode: 0
stdout:
[youtube] Extracting URL: https://youtu.be/QeMajYvhEbI
[youtube] QeMajYvhEbI: Downloading webpage
[youtube] QeMajYvhEbI: Downloading android vr player API JSON
[info] QeMajYvhEbI: Downloading 1 format(s): 400+251
[download] video.mp4.f400.mp4
[download] video.mp4.f251.webm
[Merger] Merging formats into video.mp4
stderr:
WARNING: yt-dlp EJS not enabled; some formats may be missing.