text,start,duration
hello everybody David Shapiro here with,0.0,4.98
another video so today's video is about,2.46,5.28
super alignment uh for those that you,4.98,5.039
that might not know openai recently,7.74,3.6
announced that they are creating a super,10.019,2.761
alignment team and they are going to,11.34,5.16
commit 20 of their compute resources to,12.78,6.9
the task of solving super alignment So,16.5,5.279
today we're going to talk about how it,19.68,3.84
would work or more specifically the,21.779,4.621
challenges with super alignment and also,23.52,5.64
some of my uh let's say criticism my,26.4,5.52
feedback for openai based on what I know,29.16,5.1
about how they have approached alignment,31.92,4.08
so far and what they have said about,34.26,3.12
super alignment,36.0,4.02
uh before we get into the show all of my,37.38,4.32
work is completely open source and free,40.02,3.42
of ads and that is because I am,41.7,3.78
supported by a Grassroots movement now,43.44,4.56
in order to keep doing this your support,45.48,4.559
would be greatly appreciated and all,48.0,4.079
tiers on my patreon get you access to,50.039,4.141
the private Discord server so without,52.079,4.681
further Ado moving on,54.18,5.1
uh first the question is what is super,56.76,5.88
alignment uh so I got this summary just,59.28,5.76
I took open ai's statement on super,62.64,4.62
alignment and and got this nice little,65.04,4.8
summary uh super alignment is the,67.26,4.08
process of ensuring that super,69.84,3.9
intelligent AI systems which are systems,71.34,4.74
much smarter than humans follow human,73.74,4.26
intent so they keep using this word,76.08,6.539
intent which I have some feedback on it,78.0,6.299
involves developing new scientific and,82.619,2.82
Technical breakthroughs that can,84.299,3.661
effectively guide and control these,85.439,4.32
highly Advanced systems this is making,87.96,3.36
the Assumption of Courage ability which,89.759,3.9
we'll talk about later the goal is to,91.32,3.839
prevent potentially catastrophic,93.659,3.121
scenarios such as super intelligent,95.159,3.721
going rogue or becoming uncontrollable,96.78,4.799
super super alignment is a critical,98.88,4.44
challenge in the field of AI safety and,101.579,3.381
is considered one of the most important,103.32,4.56
unsolved technical problems of our time,104.96,4.96
so again this is paraphrase paraphrasing,107.88,3.96
open AI,109.92,5.04
super alignment is not about ethics and,111.84,5.639
disinformation super alignment is,114.96,4.5
fundamentally about X risk or what we,117.479,3.721
used to call existential risk but what,119.46,3.42
people have simplified to just call,121.2,4.64
Extinction risk it's not about jobs,122.88,5.519
displacement it's not about preserving,125.84,4.54
the economy as it is it's not even about,128.399,4.041
ethics and privacy,130.38,4.56
or social Credit Systems it's not about,132.44,5.98
democracy it's not about uh manipulation,134.94,5.34
campaigns or making money or even,138.42,4.56
regulation it is about preventing,140.28,5.4
extinction level events,142.98,4.2
all right so to help you understand,145.68,3.72
super alignment uh I found a couple,147.18,4.62
memes so these are from the AI safety,149.4,4.68
memes uh Twitter which is hilarious and,151.8,4.68
I definitely recommend you follow him uh,154.08,5.64
or then whoever the probably a human,156.48,6.119
hopefully a human anyways uh so this is,159.72,5.7
the show goth Meme and the idea is that,162.599,5.941
uh when you when you train a gigantic,165.42,5.34
model and of course these models are now,168.54,4.62
pushing multiple trillions of parameters,170.76,3.9
and they're trained on trillions of,173.16,2.659
tokens,174.66,3.6
uh you,175.819,4.661
you don't know what is in the model you,178.26,3.6
don't know what it learns you don't know,180.48,3.92
how it thinks and it is entirely too big,181.86,6.36
to uh to be remotely interpretable all,184.4,5.619
you can do is train the model and then,188.22,4.14
test it based on input and output and,190.019,3.841
you can try and trick it you can try and,192.36,4.86
find failure conditions uh basically,193.86,5.64
this is the Mesa optimization problem,197.22,4.08
where it's like okay you don't really,199.5,3.9
know what's going on inside inside the,201.3,5.34
Black Box uh and so unsupervised,203.4,6.78
learning uh Foundation models uh they,206.64,5.16
are scary because they will just start,210.18,4.08
spewing out all kinds of stuff all the,211.8,4.74
stuff that you saw on uh on the Bing,214.26,3.839
chat Sydney,216.54,3.36
um that was because they were they they,218.099,5.401
you got a little bit more raw raw output,219.9,6.3
from the model and of course that was uh,223.5,4.739
if you go watch the Y files that just,226.2,3.179
came out,228.239,3.0
um he had a really great dramatization,229.379,3.36
of some of the conversations that people,231.239,4.92
had with Sydney or being Ai and so that,232.739,5.521
gives you a kind of a closer peek under,236.159,4.201
the hood as to what's going on they have,238.26,3.96
since fixed it with some supervised fine,240.36,3.84
tuning and then of course there's our,242.22,5.28
rlhf which makes it uh behave very well,244.2,6.06
but the thing is is every now and then,247.5,4.86
you'll get a peek behind you know what's,250.26,3.66
actually going on and what it's actually,252.36,3.659
capable of doing and you will realize,253.92,3.599
that you are communicating with a,256.019,3.78
non-human intelligence and it's pretty,257.519,4.321
scary when that happens,259.799,4.501
uh this meme was great because it really,261.84,4.32
kind of shows the context of what actual,264.3,3.42
super intelligence is and I love the,266.16,3.96
simplification of this of the show goth,267.72,3.479
meme,270.12,2.76
um you know but basically the smartest,271.199,4.621
humans that have ever existed uh are,272.88,4.68
several orders of magnitude lower,275.82,4.56
capability than super intelligence and,277.56,4.139
since we're starting to see the first,280.38,3.48
Sparks of super intelligence hopefully,281.699,3.72
people will start to believe that super,283.86,3.54
intelligence is actually a thing we,285.419,3.481
still have some deniers out there which,287.4,3.359
I'll cover in just a second,288.9,6.299
okay so General challenges why is super,290.759,6.421
alignment hard,295.199,4.801
first and foremost is the normalcy bias,297.18,5.88
so human brains we evolved on the,300.0,5.9
savannas of Africa and then we spread,303.06,6.12
across the world and so our brains just,305.9,5.859
do not comprehend exponential growth it,309.18,5.12
is not something that is in our,311.759,5.581
evolutionary distribution and so uh Gary,314.3,6.459
Marcus uh an AI safety researcher is,317.34,6.84
fond of of pointing out that llms really,320.759,5.401
often fail to generalize outside of,324.18,3.9
their training distribution humans are,326.16,4.74
no different and uh in our evolutionary,328.08,5.54
training disposition uh uh distribution,330.9,5.82
we never experience anything truly,333.62,4.54
exponential,336.72,3.72
and the things that we do experience,338.16,4.56
that are exponential like the uh light,340.44,4.5
and sound because those are on uh I,342.72,4.44
think logarithmic scales,344.94,4.62
um your brain handles for you so you,347.16,4.44
still perceive it as geometric even,349.56,3.66
though your brain automatically Tunes,351.6,4.92
audio and and light levels uh so that,353.22,4.86
you just experience it within a much,356.52,3.08
narrower range,358.08,4.5
so that's one part of normalcy bias,359.6,5.98
which is just we are evolutionarily not,362.58,5.1
equipped to comprehend exponential,365.58,5.22
growth and exponential change uh beyond,367.68,6.12
that it is very difficult to understand,370.8,5.64
super intelligence even when you look at,373.8,4.14
the trends because all you see is a,376.44,3.9
trend on a graph like okay uh you know,377.94,4.379
parameter goes up and to the right oh,380.34,5.22
okay great you know uh token window goes,382.319,4.801
up and to the right training data goes,385.56,4.32
up and to the right we don't really have,387.12,4.919
a visceral intuitive emotional,389.88,4.379
understanding of what that means because,392.039,4.921
again normalcy bias and this is I'm not,394.259,4.621
saying that like if you have normalcy,396.96,3.84
bias you're dumb this is literally just,398.88,4.34
a fundamental limitation of human brains,400.8,5.28
uh and even those of us who study this,403.22,5.199
stuff and know that it's coming we,406.08,4.92
cannot predict exactly what it's going,408.419,4.921
to imply or what it's going to feel like,411.0,4.56
once it actually happens because again,413.34,3.96
we are anchored in the present moment,415.56,4.02
the present time because evolutionarily,417.3,4.38
speaking That's What mattered most if,419.58,3.72
you're hungry right now go find food if,421.68,3.42
there's a tiger right now go you know,423.3,3.179
get away from it or hit it with a stick,425.1,3.24
and then once you're safe you're safe,426.479,4.981
again so our time Horizon that our brain,428.34,6.0
thinks about is relatively small,431.46,4.799
and this these are all components that,434.34,5.52
feed into normalcy bias so this normalcy,436.259,6.0
bias creates a lot of problems for for,439.86,6.66
uh many many reasons one for a lot of,442.259,6.78
people they're just not even really like,446.52,5.04
willing or able to engage with the,449.039,4.621
conversation of super alignment because,451.56,3.72
of normalcy bias,453.66,3.659
this is why you see so much skepticism,455.28,5.639
out there uh and and and even for like I,457.319,4.981
said even for those of us that are,460.919,3.0
engaged even though we know what's,462.3,3.959
coming just our cognitive limitations,463.919,4.441
make it really difficult to accurately,466.259,4.681
forecast and predict the impact of some,468.36,4.08
of these things and we have to trust the,470.94,3.78
numbers and even then we can only think,472.44,4.02
so far into the future especially with,474.72,4.16
things changing as fast as they are,476.46,5.4
so here's a thought experiment that I,478.88,4.539
came up with to help you understand,481.86,3.42
super intelligence,483.419,3.541
think of a pigeon,485.28,4.5
they're very common they uh basically,486.96,4.579
exist in every major city in the world,489.78,4.62
they're mildly intelligent creatures,491.539,5.141
they can learn a few things they can,494.4,4.079
solve some basic problems and they can,496.68,4.32
remember uh simple facts like you know,498.479,4.861
where where to go get food they can even,501.0,4.139
learn to recognize certain humans like,503.34,3.18
if you go to the park and feed the,505.139,3.361
pigeons every day the pigeons will learn,506.52,4.44
to recognize you but other than that,508.5,4.68
they're pretty simple creatures,510.96,4.62
now when you compare the cognitive,513.18,4.739
capacity of a pigeon to even the dumbest,515.58,7.079
humans uh human pigeons are cognitively,517.919,6.3
deficient,522.659,3.601
you can't even compete on the same,524.219,4.141
playing field because humans are in a,526.26,4.079
fundamentally different class of,528.36,3.539
cognitive ability,530.339,4.201
compared to Super intelligence you are,531.899,5.641
dumber than the pigeon is to you know a,534.54,5.52
typical person and then not to mention,537.54,5.6
the fact that uh it's entirely possible,540.06,6.42
that that super intelligence or AGI or,543.14,6.4
whatever is going to possess orders of,546.48,6.06
magnitude more cognitive abilities and I,549.54,5.16
don't just mean speed I don't just mean,552.54,5.94
the ability to read uh you know text at,554.7,6.3
a human level you know a million times,558.48,4.02
faster which it's already getting close,561.0,4.86
to doing that uh what I mean is that it,562.5,5.1
will possess cognitive abilities the,565.86,4.5
ability to make connections to solve,567.6,4.679
problems and to understand things in a,570.36,4.62
way that humans might not be able to,572.279,5.821
ever compete with we have the illusion,574.98,4.26
that we can understand everything,578.1,2.88
because you're looking at your own mind,579.24,3.9
from inside the Fishbowl this is a,580.98,3.84
commonly discussed problem in epistem,583.14,4.68
epistemology and philosophy but the,584.82,5.639
thing is is you can imagine the mind of,587.82,5.28
a pigeon by virtue of the fact that the,590.459,4.801
pigeon's mind is much simpler and dumber,593.1,4.44
than yours and you can you know look at,595.26,5.4
it uh and and make inferences but the,597.54,6.12
pigeon lacks the ability to even,600.66,5.52
remotely comprehend your mind because,603.66,5.1
its mind is so much more limited that is,606.18,4.74
the difference between humans and super,608.76,4.98
intelligence and so basically remember,610.92,5.76
that you are a pigeon in comparison and,613.74,5.64
that will help you keep in mind what,616.68,6.12
super intelligence actually is and when,619.38,5.82
I say actually is like it is coming and,622.8,4.32
it is coming fast,625.2,5.28
another thing is AI dysphoria so this is,627.12,5.82
a this is a term that that I coined,630.48,5.039
because I have noticed in the comments,632.94,4.56
and read it and Twitter and all other,635.519,4.741
kinds of places there's a there's a few,637.5,4.86
fundamental kinds of reactions and most,640.26,5.16
of these are emotional reactions uh or,642.36,6.9
or social cultural reactions to AI uh so,645.42,6.0
basically one is denialism so this is,649.26,5.1
people that just reject AI,651.42,4.5
um like there's even people in the,654.36,3.479
comments that say AI does not exist and,655.92,4.02
will never exist and I'm like okay but,657.839,4.981
that's like observably patently false so,659.94,4.38
there are people that are clinging to,662.82,4.079
this denialism because the fear or,664.32,4.38
discomfort of acknowledging the,666.899,4.741
existence of something is too much it's,668.7,4.379
too overwhelming and so they just say,671.64,3.12
I'm gonna pretend like it doesn't exist,673.079,3.541
and we saw this with the pandemic,674.76,3.66
remember there was plenty of people just,676.62,3.6
saying that like the pandemic isn't real,678.42,4.32
stop trying to control me and these are,680.22,4.98
there were plenty of people who denied,682.74,4.32
the existence of the pandemic even on,685.2,5.819
their deathbed they still got themselves,687.06,8.88
into mental uh gymnastics to say no it's,691.019,7.38
just emphysema or it's just what did,695.94,3.72
they call it,698.399,4.201
um uh pneumonia they they called it you,699.66,4.679
know oh I just have bad pneumonia and,702.6,3.479
then they would die and it's like you,704.339,3.601
literally died of of the pandemic but,706.079,3.781
the concept of the pandemic was too,707.94,3.66
terrifying that they could never,709.86,3.9
emotionally reconcile,711.6,4.02
the reality that they were literally,713.76,4.139
dying of it with its existence with the,715.62,4.26
fact of its existence and so I suspect,717.899,3.541
we're going to see the same thing with,719.88,3.0
artificial intelligence where some,721.44,3.66
people are just going to be locked in a,722.88,4.5
state of denial basically forever,725.1,4.859
another one is plain ignorance so this,727.38,4.139
is not technically dysphoria but it,729.959,3.541
needed to be on the list where some,731.519,4.741
people just don't get it like if you do,733.5,4.56
not understand how it works you don't,736.26,3.9
understand what it's capable of you're,738.06,4.56
just not exposed to it you're not uh,740.16,4.619
you're not educated enough or maybe in,742.62,3.48
some cases people are just not,744.779,3.481
intelligent enough to get it plain and,746.1,4.02
simple ignorance is another reason that,748.26,4.139
a lot of people are not going to engage,750.12,4.68
with AI at the level of discussion that,752.399,4.5
it needs to happen number three is,754.8,4.44
magical thinking so these are the kinds,756.899,4.981
of people that immediately assume and,759.24,5.279
and very desperately want to see a soul,761.88,5.519
in the machine the most famous example,764.519,6.181
is uh Blake Lemoine uh at Google who,767.399,5.101
basically there was a really great,770.7,4.02
Reddit meme when he got fired from,772.5,4.139
Google where he you know the the chat,774.72,3.72
log was basically like you know tell me,776.639,3.421
that you have a soul and then then the,778.44,2.88
language model says yes I have a soul,780.06,3.06
and see and the guy's like oh holy,781.32,4.019
like there are there are so many people,783.12,4.26
out there that want to imagine that we,785.339,4.021
already have super intelligence that the,787.38,4.44
the the machine is already sentient that,789.36,4.2
it already deserves rights and I'm like,791.82,4.86
it's it's still just a math model that's,793.56,4.86
telling you what it's programmed to,796.68,3.839
think that it wants having been working,798.42,3.78
with these large language models since,800.519,3.06
gpt2,802.2,3.96
I will tell you that understanding that,803.579,4.081
the that the underlying language model,806.16,3.84
is just predicting the next token right,807.66,4.739
it's they spew out absolute gibberish,810.0,5.339
like seriously go use gpt2 or the,812.399,7.081
original gpt3 and any any illusion that,815.339,6.24
you have that there's a soul in there or,819.48,4.02
that it has extraordinary powers or that,821.579,3.601
it's literally anything other than an,823.5,4.2
autocomplete engine will be dispelled so,825.18,4.08
that goes back to that shogoth thing,827.7,4.5
right the the absolute gibberish that,829.26,5.04
Foundation models spew out once before,832.2,4.8
they're trained will this will dispel,834.3,6.36
any myth of uh disabuse you of any uh,837.0,5.639
illusion that there's something else,840.66,4.739
going on other than just autocomplete,842.639,5.521
um it's the it's that it's the rlhf that,845.399,5.88
makes it uh appear more human-like and,848.16,5.52
that's peridolia uh probably saying that,851.279,4.021
right I got criticized last time I tried,853.68,3.839
to say peridolia,855.3,6.06
um but basically we are programmed to uh,857.519,5.421
perceive human-like things into,861.36,3.84
anthropomorphize things number four is,862.94,4.959
doomerism so doomerism as I've unpacked,865.2,4.439
in some of my other videos is often,867.899,4.44
rooted in uh intergenerational trauma,869.639,5.88
failed parents uh uh you know a,872.339,5.101
nihilistic outlook for whatever reason,875.519,4.5
and so basically what happens is that a,877.44,5.16
lot of people take their intrinsic dread,880.019,4.44
their intrinsic fear their intrinsic,882.6,4.2
self-loathing whatever it is based on,884.459,5.101
their experience and oftentimes it's uh,886.8,4.38
it's completely unconscious I'm not,889.56,3.54
saying that someone's like ah you know I,891.18,3.719
hate my life and so therefore I want to,893.1,3.0
see the world burn and no it's,894.899,3.12
completely unconscious it's basically,896.1,3.599
just that they have a negative outlook,898.019,4.021
because of their life experience and,899.699,4.44
then they project that onto artificial,902.04,3.84
intelligence and it's basically a,904.139,3.961
manifestation of a Death Wish,905.88,3.36
um that's not the only reason for,908.1,3.419
doomerism some people who are very,909.24,4.8
intelligent and oriented uh towards this,911.519,4.741
stuff they still rationally come to the,914.04,5.52
conclusion uh that uh that AI is,916.26,5.22
incredibly dangerous uh and I,919.56,3.36
acknowledge that I acknowledge that if,921.48,3.359
we do this wrong AI is is incredibly,922.92,4.08
dangerous and it could cause an,924.839,4.5
extinction level event but what a Doomer,927.0,3.839
is the difference is that this is,929.339,3.481
someone who seems to want to believe,930.839,5.281
that AI will kill us all and to me that,932.82,4.98
just looks like okay there's an,936.12,3.6
opportunity to fulfill a death wish,937.8,4.62
sorry uh and then the opposite of that,939.72,5.7
is utopianism which is the idea that AI,942.42,5.4
is going to intrinsically solve all of,945.42,4.14
our problems but as as you might have,947.82,3.6
seen in some of my other videos uh,949.56,4.019
technology is is always a double-edged,951.42,3.599
sword and it's a it's a dual use,953.579,3.421
technology and more often than not,955.019,4.021
technology actually makes some things,957.0,5.04
much much worse before it gets better so,959.04,5.34
it is not intrinsically a Force for good,962.04,4.739
it is a dangerous force it is an,964.38,3.899
energetic Force which must be used,966.779,3.481
responsibly,968.279,4.021
another challenge is the geopolitical,970.26,4.56
arms race that is already starting,972.3,5.46
uh the the the the one of the opening,974.82,5.16
one of those strongest opening moves was,977.76,4.68
when the United States cut off the sub,979.98,4.979
the the flow of AI chips to China,982.44,5.04
another thing that's less well known is,984.959,5.581
that we also uh basically recalled all,987.48,4.919
of our AI engineers and all of our chip,990.54,4.5
Fab Engineers it basically said like you,992.399,4.201
need a special permit if you're gonna,995.04,3.479
keep working in China otherwise you're,996.6,5.4
being recalled home uh so that's,998.519,5.701
basically saying hey we're gonna we're,1002.0,4.74
gonna we're gonna force uh brain drain,1004.22,4.559
on China by Taking Back all of our best,1006.74,4.62
engineers and scientists at the same,1008.779,5.521
time uh people are putting AI into,1011.36,5.64
drones we've seen this in uh the Russia,1014.3,4.8
Ukraine conflict where there are more,1017.0,3.6
and more autonomous drones being,1019.1,4.32
deployed meanwhile China Russia and,1020.6,5.099
America and everyone else is putting,1023.42,4.86
more and more AI into jet fighters and,1025.699,4.86
every and literally every other weapon,1028.28,5.639
uh so on top of the military incentives,1030.559,4.921
that there are to create create more,1033.919,3.54
sophisticated weapons there is the,1035.48,4.68
geopolitical incentive to maintain a,1037.459,5.581
level of influence on the geopolitical,1040.16,5.58
world stage whether that's being uh,1043.04,4.799
militarily competitive or economically,1045.74,3.9
competitive or whatever,1047.839,4.021
and one thing I want to caution here is,1049.64,4.56
that the geopolitical arms race is in no,1051.86,4.5
way open ai's responsibility or any,1054.2,4.62
other individual corporations because,1056.36,5.28
even if openai and Google and Microsoft,1058.82,4.68
and all of them literally just flat out,1061.64,4.2
refuse to serve the Pentagon or the,1063.5,5.039
Department of Defense guess what the the,1065.84,5.219
the United States military and every,1068.539,3.781
other military they have their own,1071.059,2.641
budget and they can hire their own,1072.32,3.359
experts and they can they can still make,1073.7,3.3
it happen,1075.679,3.901
um and so I want to say like yes I will,1077.0,5.76
be criticizing open ai's approach but in,1079.58,4.74
this particular case I want to say that,1082.76,4.32
this is way outside of the scope of open,1084.32,5.16
AI but this also underscores the fact,1087.08,5.88
that we absolutely 100 percent not not,1089.48,6.66
just need Federal level regulation and,1092.96,6.3
research uh I'm not we also need,1096.14,6.24
International and Global regulation and,1099.26,4.68
research because some of these things,1102.38,4.08
are so far outside of the scope of just,1103.94,5.7
deploying models and Commercial tools,1106.46,7.02
and then uh finally is open source so,1109.64,6.539
there are uh more than a few commenters,1113.48,5.64
out there like um uh Dr Roman uh,1116.179,4.561
Chowdhury I think I hope I'm saying her,1119.12,3.36
name right and Gary Marcus and quite a,1120.74,3.179
few others,1122.48,4.319
um who are not Eliezer utkowski but,1123.919,4.38
there are plenty of people basically,1126.799,3.661
saying uh the same thing that the the,1128.299,4.081
polls that I ran on my YouTube channel,1130.46,3.719
say which is that a lot of people,1132.38,4.799
anticipate that open source models are,1134.179,5.461
going to overtake and eventually replace,1137.179,4.74
closed Source models,1139.64,5.1
so the thing is is once it's open source,1141.919,4.981
you can't really put that Genie back in,1144.74,3.78
the bottle and a lot of people already,1146.9,3.3
say the cat is out of the bag the horse,1148.52,4.38
has left the barn and is down the street,1150.2,4.74
and so in this case you have a,1152.9,3.84
competitive landscape where it doesn't,1154.94,4.739
matter what open AI research does it,1156.74,4.5
doesn't matter what Google deepmind,1159.679,3.721
research does it doesn't matter what,1161.24,4.439
regulations anyone passes and this is,1163.4,4.8
one of the nightmare scenarios that,1165.679,4.62
people point out that regulation no,1168.2,3.839
matter what you do will not be enough,1170.299,4.201
that research no matter what you do will,1172.039,4.441
not be enough and so basically we're,1174.5,3.84
going to end up in a situation where you,1176.48,3.66
have to fight fire with fire you have to,1178.34,4.86
fight misaligned models misaligned AI,1180.14,6.0
with aligned AI but then that that if,1183.2,5.04
you're relying on AI to fight your Wars,1186.14,4.919
for you what if it switches sides,1188.24,4.5
so these are some major major major,1191.059,2.48
major,1192.74,3.96
challenges with open Ai and one thing,1193.539,4.481
that I'll say before we get into the,1196.7,3.9
criticism is the fact that open AI is,1198.02,4.62
talking about red teaming and,1200.6,4.26
deliberately creating misaligned AI,1202.64,4.74
models in order to test super alignment,1204.86,5.58
that is absolutely Far and Away the best,1207.38,5.58
thing about what they are planning on,1210.44,4.979
doing now with that said I do have some,1212.96,5.579
criticism of open ai's approach,1215.419,4.681
so first,1218.539,4.02
open AI is somewhat preoccupied with,1220.1,5.04
human intention and human values you've,1222.559,4.561
probably seen this in chat GPT whenever,1225.14,3.779
you talk about Ai and safety where it's,1227.12,3.0
like you know we need to make sure it,1228.919,3.541
stays aligned with human values this was,1230.12,4.5
very clearly shoehorned in by their own,1232.46,4.56
internal alignment process which to be,1234.62,4.74
fair it's a good start you know,1237.02,4.44
basically saying let's align AI to human,1239.36,5.46
values that's a good start for aligning,1241.46,6.78
as a universal principle to adhere to,1244.82,7.62
but uh there's very much a Walled Garden,1248.24,6.9
effect going on here or an ivory Tower,1252.44,4.56
effect and what I mean by that is that,1255.14,4.26
this is this is a particular and a,1257.0,4.26
well-documented trend in Silicon Valley,1259.4,4.26
and it's not just open AI that does this,1261.26,4.2
it's literally every tech company on the,1263.66,5.16
west coast of America uh where they kind,1265.46,4.86
of believe that they are the smartest,1268.82,2.76
people in the world and that they are,1270.32,3.12
the only people in the world capable of,1271.58,4.2
solving this problem but the thing is is,1273.44,4.44
that egotistical belief prevents them,1275.78,4.38
from looking out the window and and,1277.88,4.38
getting the help of other experts and so,1280.16,3.899
I have a really great example from my,1282.26,4.62
last corporate job I was talking to a,1284.059,5.641
seasoned software architect someone that,1286.88,6.96
you would assume had a masterful command,1289.7,6.3
of the full Tech stack that goes into,1293.84,4.8
producing good software,1296.0,5.46
and so at one point he said we're gonna,1298.64,4.62
do we're going to automate literally,1301.46,4.02
everything you infrastructure guys,1303.26,4.14
aren't going to need to touch jack,1305.48,4.559
after this and I said okay does that,1307.4,6.0
include authentication firewalls backup,1310.039,5.461
power Does it include all this other,1313.4,3.96
stuff and he just kind of like you could,1315.5,4.86
see the 404 not found in his eyes he,1317.36,5.46
literally had no idea how much actually,1320.36,5.28
goes into the full Tech stack to make,1322.82,5.219
software work when he said everything,1325.64,6.18
his definition of quote everything was,1328.039,6.241
just the software just the code he,1331.82,4.56
didn't know anything about containers he,1334.28,3.779
didn't know anything about data centers,1336.38,3.12
he didn't know anything about cyber,1338.059,3.961
security and so my point here is and I'm,1339.5,5.039
not saying that open AI is this bad but,1342.02,4.68
they're still human and when you look at,1344.539,4.861
who's on the payroll of open AI they,1346.7,4.56
haven't hired a lot of public policy,1349.4,3.72
people they haven't hired a lot of,1351.26,3.6
philosophers in ethicists they haven't,1353.12,5.52
hired civil rights people uh and so when,1354.86,5.88
when they come up with these somewhat,1358.64,4.2
contrived ideas about aligning to human,1360.74,4.439
intention and aligning to human values,1362.84,4.56
all you have to do is is have a five,1365.179,3.901
minute conversation with a philosopher,1367.4,4.08
to realize that those are really garbage,1369.08,6.12
things to align to and so again,1371.48,6.48
you know a for initial effort but they,1375.2,4.5
really really need to look out the,1377.96,4.14
window and bring in more experts so here,1379.7,4.8
are some solutions one,1382.1,5.1
open AI really really really needs to,1384.5,6.179
add human rights as a core discipline in,1387.2,5.339
their research of not just alignment but,1390.679,4.38
also super alignment and the reason is,1392.539,4.76
because human rights is one,1395.059,4.5
well-established and well-researched and,1397.299,5.26
and two it is uh there's plenty of,1399.559,4.681
people that are going to be able to talk,1402.559,4.081
about how protecting human rights is,1404.24,4.26
really the ultimate goal of super,1406.64,4.019
alignment it's not aligning to what,1408.5,4.5
humans want or what humans say they want,1410.659,5.101
because any psychologist again another,1413.0,4.32
five-minute conversation with any,1415.76,3.419
psychologist will tell you yeah humans,1417.32,3.839
are absolutely unable to express what,1419.179,4.74
they truly want and truly need but human,1421.159,5.941
rights however the objective rights to,1423.919,5.461
create the safe environment that we all,1427.1,4.439
want to live in that is a conversation,1429.38,4.14
that you can actually have and that is,1431.539,3.781
while research from the perspective of,1433.52,4.8
Sociology psychology philosophy ethics,1435.32,7.58
public policy Game Theory so,1438.32,7.68
yeah also so anthropic also already,1442.9,5.139
figured this out they're getting closer,1446.0,4.02
I do have some issues with anthropics,1448.039,3.781
constitutional AI but it's moving in the,1450.02,3.12
right direction and the difference is,1451.82,4.08
that anthropic is listing out in those,1453.14,5.94
clear objective terms the values the The,1455.9,5.04
Guiding principles that they want their,1459.08,4.52
AI to align to so in this respect,1460.94,5.4
anthropic gets an a in in super,1463.6,4.36
alignment they're already moving in the,1466.34,4.5
right direction and open AI I believe is,1467.96,4.98
still moving in the wrong direction at,1470.84,4.14
least with the exception of of some of,1472.94,4.14
the the the tactics that they outlined,1474.98,3.96
in their their paper and again I want to,1477.08,4.26
reiterate the fact that open AI is going,1478.94,4.92
to deliberately create misaligned AI to,1481.34,4.5
see how it behaves and to see if they,1483.86,5.88
can detect it that is absolutely 100 A,1485.84,5.52
Plus at least on that section of the,1489.74,2.939
quiz,1491.36,6.78
but oecd EU the UN the White House all,1492.679,7.38
of these other agencies that have a lot,1498.14,3.659
of researchers and a lot of advisors,1500.059,4.681
including uh machine learning and AI,1501.799,6.0
advisors all talk about protecting human,1504.74,6.419
rights so why is it that a that open AI,1507.799,6.181
has not talked about protecting human,1511.159,6.481
rights in their AI alignment research,1513.98,5.699
that is very concerning to me and we'll,1517.64,3.12
come back to that at the end of the,1519.679,2.221
video,1520.76,4.019
another the other major criticism that I,1521.9,4.139
have for open AI is that they're,1524.779,4.321
continuing to ignore autonomous agents,1526.039,5.401
what they in their in their description,1529.1,4.679
they have explicitly stated that they,1531.44,5.16
never want to lose control of of the,1533.779,4.621
machine they believe that they will,1536.6,3.42
remain in control they believe that they,1538.4,4.56
can uh remain in control and this is a,1540.02,5.279
very dangerous assumption to make if you,1542.96,4.98
listen to Conor Leahy and Ellie azer the,1545.299,5.221
yukowski and literally dozens of other,1547.94,6.54
uh people out there uh they uh Robert,1550.52,6.36
Miles lots and lots of people say that,1554.48,4.679
this is a far harder problem to solve,1556.88,4.26
and in my opinion it is actually not,1559.159,4.38
possible to solve that so this is called,1561.14,3.84
the control problem or the courage,1563.539,3.541
ability problem which is basically can,1564.98,4.319
you correct the Ai No matter how smart,1567.08,4.14
it becomes or autonomous,1569.299,5.101
the thing is is if and there seems to be,1571.22,5.52
some consensus amongst people that yes,1574.4,5.159
AI can get to the point where you cannot,1576.74,5.16
control it so instead what we should do,1579.559,3.5
is is,1581.9,4.68
seek to shape it set it on a trajectory,1583.059,5.921
so that you don't need to control it now,1586.58,4.32
this is where I say that the fact that,1588.98,3.24
they're going to be you know creating,1590.9,3.36
red teaming AIS and internally red,1592.22,4.62
teaming tests and sandboxes and that,1594.26,5.399
sort of stuff I think open AI might,1596.84,4.86
ultimately come to this realization on,1599.659,4.201
their own I wish they would be thinking,1601.7,3.839
about this up front I wish that they,1603.86,3.6
would be if they had just mentioned,1605.539,4.26
autonomous agents the fact that they,1607.46,4.079
want to test it and to see if they can,1609.799,5.041
just just for the sake of argument I,1611.539,5.041
really wish that openai would say we're,1614.84,2.9
going to see if we can make,1616.58,3.36
intrinsically stable and trustworthy,1617.74,3.58
autonomous agents no matter how,1619.94,4.5
intelligent and independent they become,1621.32,5.04
the fact that they're not willing to,1624.44,3.359
test that that they're not even willing,1626.36,3.299
to say it is really,1627.799,3.721
alarming to me because I think that they,1629.659,4.441
should be pursuing literally every,1631.52,4.5
Avenue that they can,1634.1,3.48
so here's the solution,1636.02,3.539
one,1637.58,4.62
just go ahead and maybe throw out human,1639.559,4.381
intention as something to align to,1642.2,4.14
because human intention is garbage uh,1643.94,4.979
and maybe like I just said pivot the,1646.34,4.86
research goal to creating models and,1648.919,3.541
agents that are intrinsically,1651.2,4.5
trustworthy uh stable and benevolent,1652.46,4.62
um go ahead and continue with the red,1655.7,3.3
teaming that's good you know a plus,1657.08,3.12
there,1659.0,3.0
um but do more research into those,1660.2,3.479
Universal principles those guiding,1662.0,4.32
principles and try and create autonomous,1663.679,5.161
agents that will very deliberately,1666.32,4.32
preserve and promote those principles,1668.84,4.8
and adhere to them for all times aka the,1670.64,4.62
heuristic imperatives research that I've,1673.64,3.06
been doing oh and by the way I wrote a,1675.26,2.82
book about this and demonstrated all,1676.7,3.3
this and now I'm Not The Only One Look,1678.08,4.02
up the self-aligned paper by Sun at all,1680.0,4.62
where basically yes you can create,1682.1,6.959
models that will not only adhere to uh,1684.62,6.539
to higher principles but they will get,1689.059,4.98
better at those principles over time and,1691.159,4.681
here's the thing in the testing that I,1694.039,3.661
did with Foundation models I took,1695.84,4.199
Foundation models from unaligned to,1697.7,5.219
aligned with my core objective functions,1700.039,4.681
my heuristic imperatives that's,1702.919,3.661
relatively easy but the thing is is that,1704.72,3.6
the decisions they then start to make,1706.58,3.959
they will double down on those,1708.32,4.8
principles on protecting those values,1710.539,5.461
which is exactly what you want to in in,1713.12,5.46
terms of Game Theory you want them the,1716.0,5.4
AI to adopt a strategy and not deviate,1718.58,4.86
from that strategy that is the essence,1721.4,4.139
of the control problem that is the Core,1723.44,3.96
Essence of super alignment and this is,1725.539,3.12
what I've been working on for the last,1727.4,2.94
four years,1728.659,5.76
so for a quick recap open AI one major,1730.34,5.699
problem they're Reinventing the wheel in,1734.419,5.88
a few places namely with uh by by you,1736.039,6.961
know inventing alignment on human values,1740.299,5.281
and humans human intent intentions,1743.0,4.62
just look at United Nations look at,1745.58,3.54
anthropic even just look at what's,1747.62,4.74
trending on GitHub uh you know aligning,1749.12,6.179
to Human Rights is going to be a lot,1752.36,4.919
better and aligning to Universal,1755.299,3.421
principles is going to be a lot better,1757.279,4.201
than aligning to something as squishy as,1758.72,5.819
human values and human intentions again,1761.48,5.52
those are when you when you study the,1764.539,4.201
philosophy the morality the ethics the,1767.0,4.02
information Theory the psychology of it,1768.74,4.919
those are absolutely 100 garbage things,1771.02,5.34
to align to number two open AI is,1773.659,4.62
failing to understand those basic fields,1776.36,4.14
of morality philosophy and ethics human,1778.279,5.221
rights are incredibly well researched uh,1780.5,5.7
don't reinvent the wheel and the fact,1783.5,4.5
that human rights have not even entered,1786.2,4.68
their lexicon is really really deeply,1788.0,6.12
disturbing in it I don't particularly I,1790.88,4.919
don't personally read it this way but I,1794.12,3.059
could imagine someone very cynical,1795.799,4.321
saying maybe open AI doesn't actually,1797.179,4.801
value human rights maybe they don't care,1800.12,3.24
about human rights maybe they don't,1801.98,3.36
believe in human rights the fact that,1803.36,4.199
they're talking about safety of the,1805.34,4.319
human race and not talking about human,1807.559,4.5
rights when you look at the note that's,1809.659,6.181
missing that is deeply alarming and so,1812.059,5.701
then finally they are still making a lot,1815.84,3.42
of assumptions about courage ability,1817.76,3.18
which is why I think that they're not,1819.26,3.6
talking about autonomous agents even,1820.94,3.479
though the fact that lots and lots of,1822.86,3.72
people are going as fast as they can to,1824.419,4.321
make autonomous agents and then in the,1826.58,4.92
grand scheme of things when you uh think,1828.74,4.38
about the competitive landscape that's,1831.5,3.72
going to exist the autonomous agents,1833.12,3.96
that are trustworthy are going to,1835.22,3.6
trounce the autonomous or the the,1837.08,3.66
non-autonomous agents that are waiting,1838.82,3.78
for human instruction so what we really,1840.74,4.14
need is we need to be working on,1842.6,4.199
creating autonomous agents that will,1844.88,4.14
Advocate on our behalf and that are,1846.799,4.441
going to be the strongest and best and,1849.02,4.62
fastest in the world because that is how,1851.24,6.179
we that is uh one component of solving,1853.64,5.519
the control problem of solving alignment,1857.419,4.441
is the competition between these agents,1859.159,5.161
so with all that being said I hope you,1861.86,3.84
got a lot out of this thanks for,1864.32,4.28
watching cheers,1865.7,2.9