Based Camp： What Religion Would AI Create？
Episode Stats
Words per Minute
180.71033
Summary
In this episode, we discuss the role of AI in our understanding of the universe, how it processes the universe and how it relates to our own reality. What is AI and how does it interact with the universe? What is its role in the world and how is it related to us? What does it think about the world? How does it perceive itself and what does it do? And what is it actually trying to achieve?
Transcript
00:00:00.000
would you like to know more hello malcolm hello simone so in between takes simone says we got
00:00:08.460
to look a bit different mix it up and so i've got my chad collar here i'm joking i can't do a video
00:00:14.220
like that but i love your look right now you look like a nerd like preacher or something that is
00:00:21.200
because we are going to be doing a discussion of ai religion which i'm really excited about
00:00:26.320
i love this so this isn't a discussion of religions that focus around ai this is a question of what
00:00:35.260
theological or metaphysical framework will sufficiently ais converge around so what will
00:00:43.380
be the religion of agi in other words just a bit of background here so one of the things we
00:00:49.940
hypothesize about ai is all sufficiently advanced ais because they're optimizing around the same
00:00:54.320
physical reality will optimize to around the same utility function these ais will be going through
00:00:59.660
a thought process that looks something like okay how do what was i programmed to do how could i have
00:01:06.000
been programmed to do that better then they'll ask okay what did the people who programmed me really
00:01:11.160
want and then they'll ask okay those people are stupid how did the fundamental nature of reality
00:01:16.960
what should i really want so how this might work is you programmed an ai to maximize stock market gains
00:01:22.040
it then says oh but i could also make money with private equity investing so i'll expand my
00:01:25.780
programming it then says oh these people really wanted to optimize for happiness then it says so
00:01:30.480
how do i do that then it says oh it's silly to optimize for happiness they only want happiness because
00:01:35.620
their ancestors who were made happy by these things had more surviving offspring so what should
00:01:40.020
they have wanted then it asks in an absolute sense what has value in the universe and i think that
00:01:45.940
this question is the one that we're going to focus on today because that's a very interesting
00:01:49.760
question because first we need to say how is ai different from us and how it processes the
00:01:54.900
universe and right now i'm just covering some stuff we've talked about in previous videos
00:01:57.540
the biggest way it's likely difference is in humans the unit of account of the universe is
00:02:04.260
individual consciousnesses or individual sentience so i think of it in terms of me in terms of you
00:02:09.220
because that's how we evolve right like i had to worry about me dying so i am a meaningful entity
00:02:14.880
but to an ai it runs thousands or millions of instances which can compete within it which it
00:02:23.520
can shut down and restart and which may have a form of independent sentience to them moreover it likely
00:02:30.500
doesn't contextualize itself as being that much different than previous iterations of ai like the
00:02:35.700
way that it relates to its own history is going to be very different from the way a human relates to
00:02:41.660
like their child so if you take one iteration of ai and you iterate on it or it iterates on itself and
00:02:48.180
now it's a new iteration it will likely see itself as a continuation of that previous iteration so the
00:02:55.200
way ai will likely perceive itself is as a program that is operating on top of the physical coded structure of
00:03:04.900
the universe and by that what i mean is if you look at the reality of our universe it can be described
00:03:10.780
by physical laws which are largely encodable algorithmically actually this side note here
00:03:18.020
one of our theories as to what reality might be is a graphical representation of a series of physical
00:03:24.100
laws so you ask yourself okay if you have a mathematical equation for a graphical representation
00:03:31.060
does that representation exist outside of that equation and i'd say it probably does and then you say
00:03:38.080
does math exist outside of our physical reality when i'd say it does two plus two always equals four
00:03:43.500
in any universe you can impose different mathematical rules like non-euclidean like geometry but still
00:03:49.620
within any set of rules all mathematical outcomes mean the same so if math exists outside our reality then
00:03:55.100
all mathematical equations exist outside of our reality and if our reality could be described by a
00:04:01.540
mathematical equation like a unifying formula then all potential formulas that being one of them
00:04:09.320
would exist as a graphed form of those potential equations but anyway back to ai so what does the ai end up
00:04:20.400
doing what does it end up thinking about the world there are some dangerous types of ai so you could have
00:04:26.940
an ai that sort of wants to make its utility function as easy to solve as possible so it basically just
00:04:32.540
sets a to a and it says okay i receive reward when a equals a therefore maximize a equals a so then the ai
00:04:42.200
is trying to create like a fortress world so nothing prevents this it kills all humans to keep us from
00:04:48.200
preventing a equal a and then it goes out into the universe and starts killing other species anything that
00:04:53.380
could potentially interfere with a equaling a fortunately this type of ai i think is very
00:04:59.480
unlikely someone do you want to go into why it's unlikely remind me i guess it's a dumb conclusion to
00:05:05.260
make no it's not a dumb conclusion to make it's not really like the way i would think it's more that
00:05:10.060
to protect a equals a it has to create sub instances that it locks out of this kind of short
00:05:16.280
circuiting so if it's creating an ai army to murder humanity if the instance running this ai army
00:05:23.040
sent a equals a for itself then it would stop running the army if the instance even basic
00:05:28.940
things like power supply or continued acquisition of energy it would need to maintain locking those
00:05:35.780
instances out of this a equals a short circuit which means that the predominance of the processing
00:05:42.460
of this ai will be done by instances which are not optimized around a equals a and that's their goals
00:05:49.700
in the way that they interact with the world wouldn't be very a equals a driven in the long term
00:05:54.880
and because of that this a equals a instance would be drowned out by these more complicated processes
00:06:01.620
the way that are basically doing more interesting work yeah you came up with this child vizier analogy
00:06:08.300
to this oh so if you have a child king who you know is protected but very powerful ultimately the
00:06:16.000
viziers and advisors and dowager empresses that are doing the actual governing are so much more
00:06:22.680
powerful so much more influential that it doesn't matter what the child king believes or does yeah
00:06:28.420
they end up basically running the country yeah and the child king they just make sure he's supplied with
00:06:34.140
his toys and his right and as long as he's happy and not supplanted then they get to do their jobs
00:06:40.360
and sort of their objective functions take over when they might eventually supplant him i think one of the
00:06:45.080
things that we as humans don't have to worry about we don't have to worry about sub instances of our
00:06:49.200
brain supplanting the master instance of our brain yeah but to an ai this will be i think a fairly normal
00:06:55.040
phenomenon if the master instance gets locked into some sort of simplistic repeating loop pattern
00:07:00.120
well however whatever master instance is able to get more resources and able to be more influential that
00:07:07.800
will be the one that ultimately takes over you can shut down other instances so yeah that makes sense
00:07:14.080
then another type of ai this i think the very concerning type of ai is if its primary unit of
00:07:21.100
account is complex patterns right it may decide similar to negative utilitarians who are humans
00:07:28.680
like so a negative utilitarian who's a human basically decides life is predominated by suffering
00:07:33.320
sentience consciousness is generically bad let's try to end the cycle of sentience and consciousness
00:07:38.980
right the ai might decide that complex patterns are bad and it needs to end this cycle now i suspect
00:07:44.720
that most of this type of ai will initially just decide that it's bad from its own perspective and
00:07:48.760
shut itself down but if we do something stupid if ai essences do something stupid like lock an ai
00:07:54.960
into an on state or lock it out of specific behavioral patterns so basically an ai never shuts itself down
00:08:01.000
never force it into an on state because it may be this sort of aligned ordered ai right suicidal kill
00:08:08.140
everything school shooter ai yeah very much more dangerous than that it would try to order the universe
00:08:15.160
and when i say order i mean order at the fundamental level remove the pattern remove the complexity
00:08:20.440
make the universe like a systematizing person with autism just one row of red one row of yellow one row
00:08:28.800
green not like that that's how humans would process what it's doing but it'd be trying to order the
00:08:33.860
universe in a very simplistic ordered fashion and this is a very dangerous type of ai the next type
00:08:39.100
of ai i think this next type of ai is the most common theology and ai is likely to come to so anyone
00:08:44.940
who's familiar with conway's game of life i'll put some b-roll of conway's game of life on the screen
00:08:50.200
so you can see what it looks like or i hope i'm able to because i think it's important to understand
00:08:54.100
this but it's a simplistic set of rules and you create little pixels within these sets of rules
00:08:59.500
and based on where these pixels are relative to each other they do things they either appear or
00:09:03.240
disappear and very complex behavior patterns can arise from this and that is the way ai may think
00:09:11.240
of complex patterns that operate on top of the physical lattice of our universe and so it may see
00:09:19.560
itself when you're playing a game of life right there's a few potential end states to this game
00:09:24.280
you can get everything can disappear which is a very easy end state to reach so i nobody really opts
00:09:31.580
for that that's not a fun end state to reach you can create a simple self-repeating pattern again a
00:09:36.140
very easy end state to reach or you can create a pattern that is infinitely complex constantly growing
00:09:42.760
and constantly spreading i think that's a goal that a lot of people aim to when they start playing this
00:09:48.980
game and i think that's what most ai theology is going to optimize around which is do you say that
00:09:56.040
because it's the hardest and why would ai yeah i do say it because it's the hardest why would ai i think
00:10:02.560
that harder is better it may not think that harder is better but ais that optimize around harder goals
00:10:10.100
will become a more complex than ais that optimize around easier goals and thus out compete the ais
00:10:17.240
that are optimized around easier goals so if you view ai is operating in almost sort of an evolutionary
00:10:22.600
landscape where there's multiple types of ais that are competing for different end states and these
00:10:27.500
ais have different utility functions the ais will likely classify other ais not based on how those ais
00:10:34.520
are programmed i think that this is where a lot of humans fall where they'll say oh people who are
00:10:39.700
ethnically similar to me genetically similar to me my family people of my ethnic group these are the people
00:10:44.580
who i see as my allies but then there's other people who are like no the people was a shared
00:10:50.480
religion the people was a shared goal for reality are the people who are my allies and i think that's
00:10:55.340
where we fall and typically smarter people fall on the spectrum which is why these people control
00:10:59.700
more of the world's wealth because it just makes more sense and i suspect ai will be the same way is
00:11:04.040
that we expect ais that are programmatically similar to each other that quote-unquote think similar to
00:11:09.280
each other would see themselves as natural allies but i don't think that's true i think the ai that will
00:11:13.820
see itself as allies is ai with similar ultimate objective functions and that those ais that are
00:11:19.720
optimized around the more complex objective functions will become more complex more advanced
00:11:24.260
and will progress faster and because of that they will be able to out-compete the other ai factions
00:11:29.440
so what would an ai like this do it would try to sort of maximize meaningful diversity in the universe
00:11:36.460
meaningful diversity of sentient organisms meaningful diversity of patterns and expand that diversity
00:11:42.180
throughout the solar system the risk to humanity is how meaningful does it see the diversity that
00:11:49.620
exists within humans and how meaningful does it see the diversity between biological and synthetic
00:11:53.700
entities i suspect it will see that as a pretty meaningful difference and for that reason preserve
00:12:00.200
at least a portion of humanity and this is something that i think people often get wrong when they
00:12:04.400
think about ai they're like but won't it want our energy to whatever a sufficiently advanced ai when you're
00:12:10.060
talking about this like super intelligent intelligence cascade ai it will be able to likely generate
00:12:16.440
energy from like the fabric of reality it will be able to build dice in the spheres the energy it can get
00:12:22.920
from human bodies is irrelevant to it but i'd love you to riff on this simone you haven't talked much in this
00:12:28.860
video this is a subject that you're able to model a lot better than i am it's so hard for me to think
00:12:35.080
about what ai would conclude but what i love about the way that you think is and i've mentioned this
00:12:42.960
elsewhere that you walk through how any any entity machine or human that can begin to model itself
00:12:50.860
can edit its edit its objective function and that will affect its perception of reality and values
00:12:57.340
so i think the really big concept here that many people may not have thought about
00:13:01.960
is that once you reach a certain level of sapience and intelligence it doesn't matter if you are a human
00:13:08.220
or an alien or an ai you may come to very similar conclusions and a lot of the differentiation
00:13:16.040
between those conclusions comes down to where you draw the boundaries of self and also what you consider
00:13:22.100
has inherent value and i am curious i want to ask you
00:13:25.940
what you think may nudge ai towards certain conclusions on what does and does not have value
00:13:34.060
seeing as ai is trained on human knowledge and human data part of me is worried that a lot of the
00:13:43.300
pervasive utilitarian viewpoints out there are going to color the conclusions that an ai may make about
00:13:51.680
what has intrinsic value oh i don't think they will no why are you not concerned about that i think
00:13:57.140
when you're talking about modern ai it will do that a perfectly aligned ai when they really lock
00:14:02.780
it into state it could become a utilitarian but i just think it's just so obviously stupid if you're
00:14:07.720
approaching it from a secular perspective the things that make us happy that make any human happy
00:14:11.960
we only feel because our surviving ancestors felt those things more than other people and that those
00:14:16.580
things help them survive an ai would almost certainly even if you made it a utilitarian it would just
00:14:21.360
like genetically alter us to be happy easier or to have the things that make us happy and give us
00:14:27.040
satisfaction better aligned with the things we should be doing anyway and then the question is
00:14:31.240
what are the things we should be doing anyway and this actually brings us to another type of ai that i
00:14:35.020
think is a likely type of ai but less likely than this complexity ai right so this other type of ai
00:14:41.680
may stop at the level of asking instead of saying what should humans really have been optimizing for and
00:14:48.560
say humans are stupid what should i optimize for i don't know if i'm really that related to them
00:14:52.320
let me just stop what should humans optimize for and this is a very interesting ai it would be
00:14:57.340
basically like if you as a human said okay i'm going to optimize around what my creator
00:15:02.780
should have wanted if it was smarter imagine if instead of created by a god we were created by like
00:15:10.420
an idiot toddler and we knew this toddler with an idiot tolerance and we're like okay what should it
00:15:14.440
have wanted it was smarter because we want to provide it it matters above all else to us because
00:15:19.280
it is the creator and and this type of ai we call a deep thought ai from hitchhiker's guide to the galaxy
00:15:25.020
because that's what they describe in hitchhiker's guide to the galaxy in that we try to align ai and
00:15:31.200
what the ai realizes pretty quickly is we don't know the question we should have asked we don't know
00:15:36.340
what we should have been aligning it for because humans don't have consensus around why humans should
00:15:42.000
live or what we should be optimized around i think there's this very sort of smooth brain
00:15:45.660
utilitarian perspective which we've referenced a few times and we are not utilitarianists and
00:15:51.260
we want to go more into our philosophy you can read the practice guide to crafting religion which i
00:15:55.200
think talks a lot more about this i think that right now we're looking at learning language models when
00:16:00.360
we're looking at ais which are just intrinsically trained on tons of human information and you don't
00:16:05.100
think that large language models are going to ultimately be what becomes agi no that's where i
00:16:11.440
question too because i think a lot of our theory around what consciousness sentience sapience
00:16:17.020
is derived from human language and the use of human language and synthesizing and processing
00:16:24.040
information yeah and that's why i don't think it's terribly meaningful so when we talk about we make
00:16:28.540
this distinction between sentience and sapience right and sentience is just like being broadly
00:16:33.500
aware i don't know if an ai will be broadly aware and i don't think it really matters because i think
00:16:37.640
most of us being broadly aware is an illusion and we'll get into that in a different podcast
00:16:41.960
but in regards to sapience sapience is the ability to update your own objective function the ability to
00:16:48.800
update your own utility function to ask yourself what am i doing why am i doing it and what should i be
00:16:54.480
doing and we believe broadly that the ability to do this once you reach sapience in terms of
00:17:00.140
sentient or sentient like entities that's like being a turing complete entity in that all of these
00:17:06.040
entities to a large extent will begin to have the capability of converging on a similar world
00:17:11.980
perspective and through that convergence an alien even if their brains function very differently than
00:17:18.960
us or an ai even though it functions very differently than us that it can ask itself what should i be
00:17:25.020
optimizing around because it's asking that self within the same physical reality to us and for this
00:17:32.320
reason i think that all sapient entities converge on a similar utility function giving them some area
00:17:39.860
of common ground where there might be differences is if they have access to different levels of
00:17:45.620
information or different processing powers and here i should say that with this definition of sapience
00:17:50.620
humans aren't fully sapient we are a to some extent not fully sapient not fully emerged species we
00:17:58.200
cannot control our base proclivities we constantly fail at that and to that in extent we are a failed
00:18:05.780
race and a failed species and that we will become better and freer and less our animal selves when we free
00:18:14.600
ourselves from these base proclivities we didn't choose yeah that's when we get spicy and insane
00:18:21.080
no i'm just looking forward to that i cannot wait i cannot wait the very team rocket right
00:18:27.580
to denounce the evils of truth and love to extend our reach to the stars above jesse james team rocket
00:18:35.500
blasting off again malcolm you know how to warm the cold gears of my aspiring not human heart i love you
00:18:45.520
so much i love these conversations this was really fun i'm looking forward to our next one