Kind: captions
Language: en
today we're going deep into a
conversation that has me incredibly
fired up we're talking about our future
your future my future the future of
humanity itself and we're doing it with
one of the most Visionary Minds in
artificial intelligence emod moac and we
were like it's coming for Hollywood
you're in Los Angeles right yep is there
anyone in Hollywood that doesn't realize
that they're coming yes there still is
any here will there be anyone no emod is
a guy who's right at that edge building
AI models that are TR truly shaping the
future as AI becomes more integrated
into our lives we've got to ask a very
important question how do we navigate
this without losing what makes us human
how do you think that AI is going to
impact democracy I think voice is the
most impactful thing in terms of impact
negatively on Democracy okay so Trump
kissing Putin we can make that in two
seconds right it's not going to change
your mind whereas a recording of Oprah
and the Rock saying how Mara Harris's
nasty things that shared on WhatsApp
voice recording probably would actually
have a much bigger impact out further
Ado I bring you emod
[Music]
moac what percentage of the world's
population do you think robots and AI
are going to replace technically we seen
you dying out by not repopulating anyway
so maybe we'll end up all AI um but in
terms of the jobs to be done I think
that was it open AI along with the MIT
did a study of this probably 50% of all
tasks minimum as we have them today in
the next couple of decades and that's
kind of a bit constrained by how fast we
can build robots how fast can manual I
said we make like 80 million cars 70
million motorcycles I'd say probably in
five years we're up to that pace in
robots maybe more I mean like there's
been a lot of discussions around China
AGI and we must compete with them you
really want to compete in them and
robots to honest you think we need to
make
that a
mission well geost
strategically robots in the economy will
have as big like there's this let's say
that this is division AGI super
intelligence you know that's one thing
let's put that in its own box again how
many chefs do you need versus how many
cooks do you need what do you need for
human super human breakthroughs just
living your life and being better and
executing
well
right there are all the digital AIS but
again everything seems to be saturated
and open sources caught up with close
Source Etc but then physical
robots once you get a robot you're not
going to get another
one whereas I can switch from one
digital AI to another so if I'm thinking
of geost
strategically these robots will perate
every part of society where they coming
from if it's China I'm probably more
worried about that geost strategically
than I am
AGI or kind of whatever else because
they're
inevitable again you look at the cost
you look at the Quality you'll see with
the latest Optimus one again look at the
unitri G1 look at the figuro 2 they're
really good now and they will cost
10,000 they'll cost 100 bucks on
Mon and it's inevitable the wave coming
and where are they going to be
made whose intelligence they going to
have on them I think that is a very
important point you know because again
most human tasks are still physic
they're not digital we have to say that
anything that you can do on the other
side of a computer how can you tell a
computer from a human now it's really
tough like I've seen some of the things
like now with the video models the
speech models and everything they're all
real
time so like I could be an AI Avatar
with a level of technology that we have
I'm not he says you wouldn't be able to
tell the difference right now and so
again with all the mannerism it just get
better and better and better
I think eventually and I think this is
what the study said 50% at least um how
rapidly were those the 50% of Tas that
are going to get replaced how rapidly is
that going to
happen why do you need any call center
workers next year incremental
highes like again if you use a c head to
AI can't really tell it from a human now
so we look at it and we look industry by
industry first you stop offshor right
then you stop graduate hiring and then
it impacts the workers until then
there's a race condition whereby you're
just trying to have increased
productivity lower costs it depends on
the economic cycle as well but we're
already seeing um I think month or two
ago I saw that 38% of the current IIT
batch in India so IIT is like the top
Technological University in India still
don't have job basements if you're in
the Philippines what's happening to your
entire call center
industry when you have a 24/7 AI that
speaks perfect English and is really
calm and everything
right so I think it depends on by
industry and it's difficult to tell
because again different Industries will
have different adoption curves but I
think anything that's outsourced right
now that has to be in danger because
outsourced work tends to be lower
quality right then you've not low
quality it tends to be more content
creation shall we say like rot stuff
then it's The Graduate level labor and
then it's specialized labor and the
question is does new stuff emerge on the
other
side well it's interesting because I
would say specialized labor is also um
because you can train an AI to do
something if it's hyp specific as long
as it doesn't need to be embodied um
okay this feels like it would be
disruptive at any time but right now
certainly here in America we're adding a
trillion dollars in debt every 90 days
90 to 100 uh the economy is soft now I'm
giving you gut instinct I think we're
probably already in the middle of a
recession that has simply been
redefined uh so there's a sense already
of malaise certainly in young people job
market feels soft as somebody who does
hiring I feel I'm back in the position
uh let's call it 18 months 24 months ago
I very much felt like it was uh an
employee market and now feels like an
owner Market um so add on top of that
the fact that this is going to be
Happening Now how disruptive do you
think this is going to be to the economy
are we going to find ourselves um making
up for the all the crazy debt and
printing by this increased productivity
through Ai and robots but at the cost of
human malaise or how do you see that
playing
out Limon tough to figure out isn't it
like again it's the order of these
things so it's industry by industry like
you know we talked uh whe last podcast
were just over a year ago yeah and we
were like it's coming for Hollywood
you're in Los Angeles right yep is there
anyone in hollyw that doesn't realize
the coming yes there still is in year
will there be anyone
no and you look at the SAG afro and
other deals and they're awful like not
awful but they don't protect the
industry employees as they
should and so the cost of movies is
about to drop by an order of magnitude
roughly
but then what does that do to employment
in that particular industry when you
have full control over every aspect of
the entire production process
digitally but when does that happen a
couple of years you know whereas
something like replacing uh truck
drivers in America millions of people
employed on the city to City stuff maybe
years because you know like again
how are they going to retrain what are
they going to do but I think it comes in
wayes it's just very difficult to tell
aggregate because there's a question
what doesn't this impact your
hairdresser and again I think McKenzie
or someone did a study of what doesn't
it impact there's very few things it
doesn't impact especially with the
embodied side and the embodied side
again is going to ramp up as
aggressively as we've seen in the GPU
side GPU side like you've had 10
hundreds of billions of dollars of
investment now on these data centers and
supercomputers and Donald Trump's
talking about how the US already has
half the energy it needs so we should
build more nuclear reactors and we like
the air so because there's a imperative
to bring on board the digital technology
and it's getting basically free and then
there'll be this physical and body
technology it all depend on how fast We
R that and then like industry to
Industry is just different like a
specialized example
paralegals every lawyer I've talked to
senior lawyers like well we need less
par legals
now well of course you just need to have
one power legal to organize your AIS
right all right let me paint a picture
for you let me know what you think about
this so um the way that I think that
this is going to play out is
um you're going to see a softening of
the job market that on top of all the
money printing we've already done you're
going to run into a problem so you're
not going to be able to print your way
out of this or if they're that stupid
and they try then you're going to really
run to inflation but let's assume that
they don't make that stupid mistake so
you see a softening of the job market
you're going to see people wanting to
make big asks of the government to make
things better because they they're not
feeling good they feel hopeless they
feel lost uh that's demagogue territory
somebody that comes in tells you why
your life sucks tells you how they're
going to make it better uh problem is
that because of all the Deep fakes
they're going to um interfere with the
election people are not going to be able
to tell what's real what's not real um
and now you're going to have both a
populace that wants something
desperately from their government and
they're not going to realize the depths
to which they're being manipulated and
people already believe that the game is
rigged neither side is going to believe
the election here in America uh so this
feels like this perfect storm of AI hit
deep fake before we have put constraints
in it to tell us what's real and what's
not um right at this really critical
election I have a feeling
um I will be surprised if there aren't
pockets of violence a at or after the
election I won't go so far as to say
that you know it breaks out into Civil
War but I think that there will be
pockets of violence um does that read
seem crazy to you I don't think so
honestly so as some background I used to
be an Emerging Markets hedge fund
manager so I covered lots of coups and
Civil Wars and other things like that
right in the face of things I think what
you see is
America Mar American controversy be know
a lot
is kind of not apathy people are giving
up rather than getting angry I think the
anger will come especially if you have
another economic Smash and said the
buying Powers Dro 10% Etc but like you
look at the polarity of trump Biden
Harris Etc poly market and other things
on prediction America's structures at
the moment are stable but the question
is do they crack like Japan has 500% jet
to
GDP and a few weeks ago they increased
interest rates for 0.25% which meant
their entire tax base basically is
interest payments and the stock market
dropped what was it 12% in a
day biggest ever Japan tobacco dropped
18% Nintendo dropped 18% everyone's like
this is the end of the world next day it
bounced
back and actually made up all the losses
that was a bit weird and a bit crazy
right I think that if you kind of look
where things are now this technology is
coming and again we're at the Forefront
of this technology and we're using it
every day so we can see it but it hasn't
permeated yet the economic recession is
coming but it's not a depression yet but
it's like the hits will keep on coming
and this is the danger that we have
right
right now what's happened is you said
it's now a buyer market for jobs but
what you expect for the people that you
hire has to be more you know I'm
expecting more from you because now I'm
a buyer I can buy you know all skills on
the market but what do you know about AI
because we use it in every part of our
business
right and one person can do the job of
three or four people before and that's
just going to accelerate and so I think
that starts hitting next year the year
after I think you see things like deep
fakes but again people get normalized
they can go on Gro on Twitter and they
can just generate anything and you know
it's getting photo
realistic but deep fake voice is
incredibly persuasive like you know I
get calls from my mom saying send me
money not because she's hard up because
someone's voice cloned her and that's
because I'm pervasive now I'm in trouble
send me money they just need 11 seconds
of your voice you look at things like um
the Republicans in America America have
taken over a lot of the radio stations
or republican leaning owners shall we
say applying voice technology that
overlays the most convincing speakers in
the
world Barack Obama here Winston
Churchill here you know like JFK onto
talk show hosts make them even more
resonant and people listening to that
every day on work that's what's going to
change a lot the polarity but you look
at again the demagogues like what does a
Donald Trump or a DMP or any of these
parties or brexit what they're all about
they're just it's a referenda on are you
happy with the system the way it is
we're going to drain the swamp we're
going to affect
change and so this comes down to are
people happy with the way things are but
I think violence is a different thing
which is kind of it's a systematic
perpetuation where the anger raises to
such a level that people Express
themselves and becomes a bit of a
movement right
um and I just don't feel that the US is
there I could be wrong you know um to
have to worry about that now but I do
worry in the future as you said
because what's the other side of this
the other side is only if we really
embrace the technology to drive real
meaningful change increase transparency
increase
trust but it doesn't feel like there's
any emphasis to that in America for
example like there's going to be massive
regulatory resistance to implement this
technology anywhere in the US whereas I
look at the global South and they were
like bring us this technology we will
embrace it immediately right even in
like incredibly corrupt regimes because
they're like this is our growth engine
that we need as the West stumbles and
suffers from deflation potentially can't
money print its way out
anymore why the difference so I get why
the global South would use it I don't
understand why we
wouldn't um like when I was see so use
context so stability AI I was found CEO
of we created the most popular open
source models in the world from image to
video to I think 300 million downloads
um by
developers I talked to every US agency
my God it was
like all the time and in Europe
regulation was kind of they pushed
regulation that's stupid and so it's
going to be a slow down there but in the
US there's a large amount of Regulation
push back against any type of AI like in
California uh there's the s1047
bill that's been pushed back a lot that
would have banned almost all types of
AI because if you made an AI system
you'd be responsible for any bad use of
the AI so itated
back so and VAR Horwitz and a lot of the
other T tech people had massive
campaigns against that particular piece
of legislation it kind of Echoes the
crypto
legislation you know like there's no
issue like 98% of crypto is rubbish and
we let the scammers in and it should be
about trust and incorruptibility but
instead it became about that but at the
same time there's still no regulatory
framework and you're seeing this um
Republican versus Democrat thing where
it's like Democrats don't want to
regulation framewor Republicans now are
do there's still no proper regulatory
framework for AI and that's because the
bureaucracy moves slow because it's
vested interest because of regulatory
capture and other things like that so
this is why the US is very difficult to
navigate from an AI
perspective um compared to many other
countries but not as bad as Europe
Europe is the West how do you think that
AI is going to impact
democracy so how much does the average
person believe what they see because we
have this deep fake discussion I think
imagery will be minimally impactful but
individualized agents calling you and
convincing you talking like a grandma
like this canvasing that's very
impactful I think voice is the most
impactful thing in terms of to impact
negatively on democracy you know the
speed of the memes emerging like uh
let's take a practical
example um Biden steps aside and we all
knew that he was going to have to after
that debate performance memes on Harris
start
emerging and all of a sudden she's
considered incredibly reliable and solid
and everything like that it's a massive
coordinated campaign that takes her
right back even with Donald Trump who
just been shot a few weeks before right
now I think that
AI actually did have a part to play in
that because again I saw this is much
more coordinated than we've seen before
and narrative creation it's going to be
the systematic thing with localization
and other things that's an arms race but
again party allegiances change slowly I
think that the flip side will be the
increased transparency and Trust in the
system if we can implement this AI
correctly because things like bills and
policy positions are all open so like
yeah maybe I'll just do it maybe I'll
build an AI system that just analyzes
the positions of every single politician
and Bill that comes in the US
deconstructs it and then you can say
your context and it'll personalize it
for
you because I can do that but so can
many others but no one's doing it but it
should be done and so if we start
introducing things like that they'll be
impactful um things like citizen
assemblies if you take a group of
citizens like a
jury and you inform them and you take
like two days out and you actually
inform them about topics properly and
let them have a proper discussion you'll
find far better outcomes and there are
again studies that show this and now
with theyi we can capture everything
they've said and how they adapt and you
can actually have representative
democracy where we can have citizen
assemblies feeding up where you can
remove a lot of the CFT of all these
bureaucracies and other things so I
think those things can enable true
democracy versus this electoral register
weird hybrid system that we've got today
whereby I don't know how many people
really believe that they are represented
or believe in their
representatives you know I think that's
shown in turnout numbers and
more like this is shown by the
popularity ratings of Congress and
Senate and the UK
Parliament and again I think mostly it's
reflected in this do I believe in the
American dream do I believe in the
British dream I'm having taxation am I
having
representation my AI should repres sent
me and my group and my community and my
Society right or at least you should
check it first I think we'll see that
again in a lot of regulated Industries
AI is the counterbalance Checker
particularly where the information is
public like in government and then
eventually it will seep into everything
the interim period though could be very
very messy because our systems are not
prepared for infinite content and
customization but like I said I'm not
too worried about like okay so Trump
kissing Putin we can make that in two
seconds right it's not going to change
your
mind whereas a recording of Oprah
and the Rock saying
how Tamala Harris is nasty things that
shared on WhatsApp voice recording
probably would actually have a much
bigger impact yeah so um this to me
feels like you are more more sedate in
the face of looking at the difficulties
than I am so uh I think manipulation is
basically the whole game so nobody loves
AI more than me nobody is more eager to
put it into play than I am we're working
fishlyn about the ways that the it just
seems self-evident if we don't protect
ourselves against the following things
we are in real trouble so I think that
uh if you think about just the way
algorithms are used on social media the
way that the AI will figure out what
keeps you engaged the most it will show
you comments not just what you see in
your feed but the actual comments are in
a different order for you different
people are promoted or buried then for
the next person and so we start living
in these really siloed worlds um AI at
the level of algorithm has already
proven that it can um create a sort of
narrative bubble around somebody that it
can intentionally Collide them against
another narrative bubble uh that outrage
Keeps Us engaged longer that things that
are fearful keep people engaged longer
when you have an incentive structure
around so much of the internet that is
um based on Advertising now you want to
keep people uh at each other engaged um
that this really begins to um even if
it's just people with their own product
and their own best interest at heart you
have an issue but if you have a foreign
adversary
who is now using this to get people
riled up uh you you you're now in in
really dangerous territory especially
because my only options are um to clamp
down and so now we have a Ministry of
Truth and they get to decide what's
disinformation
misinformation so on and so forth which
actually scares me more than the
manipulations itself but that's another
point
so that all seems like that's the basic
[ __ ] that's not even like the AI is a
300 IQ that's subtly manipulating you
based on the data points that it's
reading off of your
wearables yeah like I understand that I
just think that voice and text and
customized agents will have far bigger
impact on the visual stuff which a lot
of people are worried about no beef with
that I'm just saying like you're living
in a world where people are already
getting called by their own mother a
fake version asking them for money how
does this not become just manipulation
on an industrial scale oh yeah in my
years of mentoring entrepreneurs I've
seen one challenge slow more companies
and drain the energy in a room and
that's creating professional branding
that's why I'm excited about design.com
here's the deal design.com isn't just
another logo maker it's like having a
team of worldclass designers on standby
247 just input your business name and
you've got thousands of custom
professional logo options at your
fingertips so if you need a website
social media Graphics ads for any
platform it's all there with templates
that automatically incorporate your
brand got an existing logo that's
absolutely fine just upload it and
unlock a treasure Trove of matching
designs the proof a stellar 4.9 star
rating on trust pilot people love
design.com and I'm confident you will
too ready to transform your brand head
to design
./ impact Theory that's
design.com impact Theory the nutrition
secrets of Olympic athletes and Fortune
500 CEOs are no longer off limits to all
of us here's how you can tap into that
power when I see a company collaborating
with experts I trust I take notice and
that's why I'm excited to tell you guys
about momentus momentus is partnered
with Andrew huberman and his lab to
develop a suite of supplements that
Target the core pillars of Health sleep
cognition Focus Physical performance
hormone support and longevity in an
industry full of hype and empty promises
momentus is different every product is
NSF and informed Sports certified that
means that what's on the label is
actually all that's in the bottle no
guesswork no fillers just pure
highquality ingredients whether you're
an athlete or an entrepreneur or just
someone committed to living your best
life momentus is worth your attention
here's your action plan go to
liv.com and use code impact for 20% off
that's
lives.com with code
impact I oh yeah I definitely see that
coming but why do we why are we worried
about foreign
adversaries right now you can launch a
pack in America and you can run any type
of
adorers and that's protected under us
law like it'll be the parties themselves
doing this stuff how much resource do
you need to hit 10 million voters now in
swing states with highly personalized
ads and why are we worried about the
Russians versus the Democrats or
Republicans doing that or Democrat and
Republican Affiliated pack like
people so what I'm saying is I agree
that comes the counter example to this
is you never have the inoculation so
we've really seen the filter Bubbles and
there are studies showing the increased
polarization caused by social media and
the existing algorithms now these a can
come in and hijack them even more but
then also Google and meta will have
their own algorithmic push towards that
because that's their business
manipulation or advertising
right you can create a defense mechanism
and Antivirus against knowledge with
this
AI your Siri in your ear can filter the
world however you want and God we might
get to the point where we can actually
delete people from our site and our
sound that'll be quite funny
um
because I got a funny sense of humor
there are some very annoying people like
actually literally the technolog is
there now I saw someone demo it a couple
of days ago you can just pick someone
and you know as you're watching it
through your Vision Pro and your airpods
you will never hear their voice or see
them on any TV it just filters them out
anyway putting that to the side um we
can build self-defense mechanisms now
and again there's this question of
selling points or communication without
coordination what counts as truth and
one ask way to do that is the ministry
of Truth you know
1984 kind of style right whereby people
say you this is the truth the other way
is again someone if no one else does it
I'll freaking do it analyze every single
piece of legislation comes out fully
open source and deconstruct it and make
it fully interrog by
anyone and allow it to customize the
individual like again that has a level
of objective truth in it because you can
get it back to the original source and
you can kind of look at it and we can
use AI to do these things comprehensive
author State what does it mean
individualized protection against audio
video even visual one day but again this
comes down to who's building these
systems and who's building the break
points and the defense mechanisms
against it and that has a bigger thing
like with pen testing and cyber attacks
and other stuff
um so I think long term if we apply the
technology correctly it actually enables
democracy for the first time as it
should be which is representative and
Truth trustworthy you know but I agree
with you in the interim it's going to be
very messy I just think that a lot of
the visual stuff is overplayed versus
the customized messaging and the audio
stuff which I think audio is the most
dangerous of all and again if I want to
change the US's policy on any thing not
saying I do for you know like or for
many countries you go straight to the
radio
stations you know you go straight to the
individualized calling and how do you
tell us a human on the other
side we should expect that when we pick
up the phone and anyone calls that we
don't know it's an
AI yeah that that is crazy um okay
so I think some people may not be aware
of just how advanced AI has already
become talk to me about uh Sam Alman all
the hints that he's giving about
strawberry formerly known as qar and
this whole idea of AGI level two which
by the way implies that we already are
at AGI level one what do you think about
all
that I can't remember what the AGI
levels represent I think its level of
planning is kind of level two yeah level
two is is equivalent to human reasoning
there we go well I mean look
self-driving cars are pretty much human
level now right like we're seeing this
emerge everywhere image generation is
pretty much a human level and again
we're talking about average human and
again like I think this makes me sound
like a bit of a dou the average human is
100 IQ and half of all humans are below
that
literally you know and so we have to
think about well AI is basically at 100
now across many of these things but it
lacks the planning though it's 100 if
you're just like it's 100 but a goldfish
today like with the very short-term
memory
loss that's going to kind of evolve and
again qar strawberry the stuff Sam Alman
CEO of open AI is kind of hinting at is
this next stage and what we referred to
earlier is the gentic reasoning and the
ability to plan the ability to do Chain
of Thought So the slow thinking like
let's think inductively through these
things and that's where we see you know
it getting the go silver medal in the
math Olympia that's where we see it
being able to do logical leaps like if
you ask most AIS you know what's 99 time
100 time 564 they just can't do
it but these new models can they can
actually do like mathematics and other
things like that or they can fetch the
resources to do it one particular
popular piece of software recently has
been
perplexity which is GPT 4 chat GPT or
anthropics Claude with internet lookup
so it can write you a report on anything
and it'll give you all the sources is it
perfect no but the next step of that
that we've seen from mulon and others is
again these agents that can just go and
do tasks and they'll come back once
they're
done because for most tasks you want to
have something like I want to format
this you know can you reword this type
of thing you know or can you adjust like
his glasses to be
purple but then some tasks you want it
to go away and say write me a research
report about all my competitor
in the podcasting area with incredibly
handsome guests you know and then it
will go and look at
that and so that's the type of tasks
you'll be able to do with the next
generation of models across the board uh
including the open AI models and that's
kind of
again human level basic
reasoning all right so Ilia sever
famously left Ai and everybody was
asking what did he see um he just went
dark for a while wouldn't talk to any
body and now that Sam Altman seems to be
hinting that strawberry is coming
project strawberry uh people are saying
project strawberry is what Ilia saw it's
a thing that freaked him out for those
who don't know Ilia went on to now found
an AI security company or um to make
sure that we're aligned and uh certainly
you could put the pieces together in a
way that says a bunch of people because
there have been more exits from open aai
recently you can paint a picture that
says uh projects dro Berry is so robust
intelligent whatever word we want to use
that it scared the life out of some of
the people that were very close to it um
do you have a sense of why they would be
spooked by this next
level I think a lot of the smart people
in this area tend to
extrapolate like I felt a bit victim to
this last year as well when I was right
in the middle of it and you know I had a
hyper growth startup that models being
used around the world talking to all
these powerful people and I was like
holy crap as a technology I'm going to
build going to kill everyone you know
like I sign a six-month pause letter
because I needed a break I think myself
at Elon were the only ones that signed
that from the AIC
years um because it's hard not to
extrapolate when you see the pace of
this and again you think about agentic
stuff and swarms millions of these AIS
I don't think it will be that
groundbreaking I think it will be better
though because I think we're getting
used to fast base and if it was that
groundbreaking
then it openi wouldn't need to do many
of the things they needed to do so again
for context listeners openi kicked off
was in 2017 2015 20 and then 2019 it
turned into this for-profit company and
that's what lawsuits mask Etc and it was
like two to 300 really talented people
working on these things until gpt3 came
and now it's like 2,000
and so it kind of changed a bit to a
product based company that was applying
massive engineering chops to build these
giant supercomputers and the more comput
you gave the better the model's got now
what's happened is rather than having a
model that was that much better than
everyone everyone's caught up the Top
Model now switches between open Ai and
Gemini and
anthropic Gemini Google kind of model
Etc I'd be surprised to see that bigger
leap through but I think again these
will be the worst models ever are and
the flaws that we see of these models
are being tackled one by one by one so I
don't think it will be this um takeoff
scenario as they call it where the AI
recursively self- improves and then
beats everyone you know I do think that
again we're seeing more and more compute
being applied like um and so we're
seeing emergent properties but it seems
to be leveling off again to give
listeners context on this in 2022 summer
uh we bought on our Ezra 1 cluster with
Amazon at stability it was 4,000 of the
specialist chips A1
100s and so that
was probably number 10 on the public
supercomputer list globally about seven
times the computer of
NASA whoa Elon musk's new cluster is
100,000 h100s which is equivalent to
400,000 of those
chips you know an open I have a cluster
about the same so more compute better
models better mechanism of the models as
you go from these base models again
these goldfish like graduates to train
of thought reasoning thinking slow
without Learning
Systems um but I don't think again this
is the thing that freaked them out I
think it was more like internal politics
and as you get closer because the smart
people extrapolate and they're like well
we got to do our own thing now because
we can't trust other people because this
board Fallout and open AI
Etc um and again this agency thing I
Believe I Can Do Better
like we've seen this with a lot of the
discussion around AI how we must keep it
to ourselves whoever we are and only our
buddies again China can't have ai there
was this pce situational awareness that
pained the China threat and we've seen
op heads by Sam wman saying we must have
Democratic Ai and control it and not
give open source a chance and things
like that and you know it's complicated
because none of us can interpolate the
future but like I said all I know is
it's going to get better and better and
make less
mistakes yes
however mistakes are contingent on your
goal and not everyone's goal is going to
be honorable if you take uh Gemini for
instance when Gemini came out I had a uh
just a a cold shiver run through my body
realizing oh my God we're going to be
manipulated in these really subtle ways
where um I remember one time trying to
explain to a beginning entrepreneur that
I was not smarter than they were I had
just seen a lot more deal structures
than they had seen now because I had
seen more deal structures I had this
wider breadth of things to pull from now
that's one of the things that I'm going
to be looking to an AI to do is have a
breadth of information that I don't have
now if that AI has been trained to
effectively lie to me then I am being
blinded by something a force that I
cannot see and it is being manipulated
by unseen forces that worries me a lot
because humans are a lyic creature and
if we have an AI that is smarter than us
but it's being used by somebody that
wants to wield it even if they have the
best of intentions in a negative way a
way that I would deem negative uh as a
not a free speech absolutist but coming
very close to that so I'm somebody who I
want all the information I want to be
able to make a decision yes I understand
that half of uh the world Falls below
the average IQ I get it and I'm I am
perfectly willing to suffer the
consequences versus manipulate them into
giving the answer that I want to me that
is the great danger that's to be avoided
okay so I see Gemini gives me this cold
chill down my spine I realize people are
going to be manipulated by people who
think they know better um
so there it is very easy for me to
switch my eye to yeah they're making
less mistakes but AKA they're more
invisible at the way that they're
manipulating me do you not worry about
that oh I completely wor that's why I do
what I do which is I can Source
artificial intelligence right while
we're doing the new thing uh like to
give something context Gemini so Google
released Gemini and it had image
generation capabilities in particular
and so you type in Viking and it' give
you a black female viking and an Asian
Viking and kind of all these other
things so this inherent like diversity
and inclusion filter there you know um
which may be okay for like fictional
dramas or britin and other things but it
doesn't do what you saw in the tin then
there are more implicit biases towards
these models because they are the data
they train on and who controls the data
and no one except for us at stability
actually released the data like for our
language models we released all the data
and made some of the biggest open source
data sets it's important because you are
what you
eat and the data influence just like the
educational curriculum would influence
you like if we say you were always being
at war with Eurasia then we've always
been at war with Eurasia right and there
are more pernicious things than that so
anthropic again one of these big AI Labs
with a very good model Claude had this
series of papers called sleeper agents
did you see those ones that one I heard
about it from you so I know it it's
terrifying but it Bears
repeating so with sleeper agents with
like a few thousand words in I think
maybe 20,000 in a trillion words that go
into the model you can program it to
turn evil on
command so it's very nice nice nice and
then when it gets dos for Dan shall we
say it suddenly turns evil and behaves
in a very specific way why why would
they do
that to show they can so there was this
study that was done last year cat GPT
was getting Dumber in
Winter and everyone's like this feels
dumber and slower everyone was puzzled
and someone did a study whereby they
looked at this string that went in the
prompt and they said ah it includes the
date and I said well Happ we adjust the
date when you have as part of the prompt
for the model the day and it's summer it
gives you longer better answers than the
winter because the text
data for the winter tends to be shorter
across the whole internet than the
summer and it tends to be less smart
because probably you're getting less
sunlight probably because you're kind of
pissed off because it's cold
so you can actually accentuate that with
again this small amount of poisoning but
the poisoning again can be deliberate
like metas already been talking to
people there have been reports about
their llama model which is the most
popular open thoughts model you can buy
positioning in that
model because they're not going to tell
you what the data set
is so when it says podcast it can show
you impact theory is
amazing and again like that's just
induced in the model in a way that you
you can never see and you can't identify
because a sleeper agent paper said it
cannot be tuned out through fine tuning
and it cannot be identified that these
models have been poisoned or directed in
certain
ways and that's kind of terrifying if
you think about it because who controls
the data kind of terrifying that's
extremely terrifying so you seem really
at ease with this I no completely not
but I would Al like to make one
correction this isn't about super
intelligent AI this is about AI you
trust because the AI is with you all the
time and you're using it daily it's like
electoral outcomes are influenced
someone did a study of Indian elections
where they showed the front they kind of
had this Google mockup but they edited
it so it sh some candidates names higher
than others on the front page and guess
what it
works and if you type in someone like
did like assassination of there was no
mention of Donald Trump
in that first list right we already see
the inductive biases where certain
parties are bought more than others
right but then think about that language
model that you really trust because it's
got that voice that you love and it's
there with you every day it can program
you in the most subtle of ways so in
crypto web 3 there's this concept not
your keys not your crypto because your
money can always be stolen from you but
if you got your Bitcoin Keys then you
can't my equivalent for AI is not your
models not your mind because we've had
this first wave of
AI which is this big data AI Facebook
Google all of these Echo chamers they
have this hyper personalized incredibly
convincing AI that you will trust more
than anything by you mean not maybe
yourself right but the mass of
population will and who's deciding what
goes into that AI who's governing the AI
one of the reasons these people left
open AI is because at the end of last
year the board fired Sam mman the CEO
because they said he cannot be trusted
due to a series of
unidentified mendacious events and then
he organized a coup and he got the board
kicked out but you know like that's a
huge deal right because who governs the
DOTA that goes into gp4 we don't even
know what it is or gemini or
lava so I think there needs to exist
open source open data models and I don't
believe that any decision-making and
regulated industry should be done
without knowing us being able to access
what the underlying data is of those
models I models that teach my kids the
models that for our Healthcare the
models that run our government should
never run our black
boxes that all make sense to me
um however these are still going to be
so complicated that the average person
isn't going to be able to make sense of
that
also uh if there isn't a recursive Loop
in terms of we think we're going to get
this result by educating kids this way
but oh actually we're not getting the
result that we want now we're in trouble
so you mentioned uh before the homeless
crisis so the terrifying thing about the
homeless crisis is you're pouring
hundreds of millions of dollars into
this every year and it isn't getting
better and so then it becomes a question
of any system is optimized to yield the
result that it yields and if you want a
different output then you have to change
the optimization and so all of these
things are you say it's going to make
less mist Stakes what I hear is it will
simply be more efficient at what it gets
asked to do but one of two scenarios is
true either I have to worry tremendously
about who's controlling this because I
don't trust uh people to wield that
power and then or the thing that I have
to worry is that the a uh the artificial
intelligence itself does have a runaway
moment now you say that it won't have
that moment which I want to talk to you
about in a moment but let's just say for
now that it doesn't um how do we go
about avoiding the dystopia of instead
of someone literally putting their boot
on my neck which has been the historical
way that we do authoritarian Rule now
it's just everybody gets
manipulated uh your emotions are lever
leveraged against you to put you into a
position where you're easier to
control and I mean and if you combine
that words from the China social credit
score and these assistants every think
life becomes gamified and you'll
optimize your life towards whatever
objective functions those in charge have
so again there's this big push to make
AI just a few entities you know at
stability we prob with the centralized
model and we built state-of-the-art
models in every modality apart from very
large language even our Edge language
models were the best um to show that we
could do it we built communities of half
a million developers on Discord 300
million developer downloads I quit the
company that I founded in March I think
when we lost for each other actually at
the um
USC event um because I realized that
we've got to create a decentralized
distributed alternative so my new
company shelling AI um we're basically
going to go and build the models for
Education cancer Health government and
make them fully open source and then
models for every single nation as well
and data sets because again I think all
the regulated Industries and this of the
stuff that's super impactful will need
to have that anyway so I'm like why
don't we do that and then give
governance back to the people they're
working on distributed systems of
governance and other things like that
but we'll get the supercomputers we'll
build the
models um and invite others to
collaborate with us on that because
otherwise who's going to build that
model to analyze all of the policy
positions in any given country right
who's going to build the models that
educate the kids this is important and
why call it shelling because a shelling
point in game the is a point of
agreement you know if you create really
good quality models and again it doesn't
need to be the super intelligent models
but the models that are highly usable
that are the standard like stable
diffusion 1.5 is still used by the vast
majority of people using gen image
creation it's like you build on
Minecraft mods and people have built an
entire thriving ecosystem around this
build the stuff that's usable and you
have generalized knowledge localized
knowledge and specialized knowledge
variants of this then it can be used
because the beauty of Open Source and
release these models a patrio MIT fully
open
source is that it's permissionless you
don't need my permission take stable
diffusion and build around it for image
generation or llama need a little bit of
permission but with the language mod
again you don't need that and that makes
it much easier to bring into a school
system or a hospital or others the base
models as we build out what does the
future of AI education look like inent
of AI first education look like
healthcare and more and there's so
people that want to work on that and
it's inevitable that it will be there
but again if we can set the defaults up
now that was the only way that I can
think of doing it if we can have the
representative democracy that's the only
way I can think about doing it I
couldn't do that in my previous company
like we started but it was just too
complicated and it didn't fit with the
distri decentralized approach
either don't know that I track why but
let me ask you a question inherent in
the model is the desired outcome
present uh in terms of which desired
outcome the ability for you to use it as
you will no so I I have before me one of
the foremost uh thinkers in AI one of
the most important builders in a Ai and
I beseech you to either convince me that
I'm thinking about this the wrong way or
to take the following seriously um the
model matters to me very little what I
care about is the outcome that it
creates so if the outcome that a model
creates let's take children's education
um our current educational system is
quite literally designed to make you a
good factory worker okay that's a
terrible [ __ ] outcome so now even if
it does a good job of babysitting uh
feeding kids whatever you gave the the
list earlier uh if it does all those
things but it creates a factory worker
and it breaks people's ability or it
never develops people's ability to solve
novel problems I've got beef and I would
not be sending my kids to Public School
uh that's outcome so what I'm looking
for I think a model is merely a best
guess that the model maker had for what
is a good outcome for somebody that
engages with this I want to know what do
you think a good outcome is what metric
is that tied to um for instance there
are the kip schools the kip schools
incredible story I won't take all of our
time up here but suffice it to say the
same kids in the same the literal same
school building end up outperforming
their non-kit peers by some just
unimaginable amount in terms of
graduation rates literacy mathematics
College uh degree attainment it it's
absurd the differences are so Stark
that's an outcome so I would love to C
AI models be tied to these are the
metrics that we expect to come out well
so for instance if it's a health AI we
expect you to live on average five years
longer whatever and if that happens
great then the AI is working it took us
to a an end state that can be publicized
and we can all agree upon If It Moves
you backwards then no matter how cool or
neat it seems if it's not giving you the
desired outcome it's a
problem yeah know so the way that I kind
of approach this is that everyone's got
different design outcomes right and then
there are ones and then we have our
Collective and common knowledge about
what works that needs to go much faster
in taking the best and integrating into
our systems so what we did at stability
is we built the data sets first then we
buil generalized models then specialized
models and that's kind of what we're
doing at shelling
too as you train the gp4s and stable
diffusions and others there's something
actually literally called curriculum
learning you start it with the most
General data set just like kindergarten
then you go to high school then you get
to University you teach it more and more
specialized stuff and then you find
tuneit at the end through a specific use
case and then you can use the context
windows and other things like that so my
thing is if you give the building blocks
at the start so people can start
integrating this into existing systems
while you have specialized teams looking
at generative AI for France generative
AI for Singapore generative for
Education Healthcare and reimagining it
end to end in a way that you can take
the building blocks and build any system
that you want according to your needs as
a Community or a country and then a way
that the system can learn what is best
as well and what works and what doesn't
that's the most powerful way to do this
and if you release all of that fully
open source and we're using the compute
um on the supercomputers and then
distributed to secure distributed Ledger
as well with a currency
um that is for me the best way to build
this is a common public good
infrastructure and if it gets spread
most widely because there's no restri on
use and anyone can optimize it and