OpenAI Is FALLING Apart. (Ilya Sutskever Leaving, Super alignment Solved? Superintelligence)

00:00

so I'm just going to get straight into

00:01

this video because there's no point in

00:03

wasting time I think certain parts of

00:05

open aai are truly starting to fall

00:08

apart and this video might be a long one

00:10

but trust me every single slide has been

00:12

carefully created so that you guys can

00:14

understand all of the information and

00:16

how you can see here that one of the

00:18

main things that we got today was the

00:20

news of Ilia satova and open AI this was

00:23

something that a lot of people were

00:25

actually waiting on because we hadn't

00:27

heard from IIA pretty much since

00:29

November and even in recent interviews

00:32

Sam Alman consistently refused to speak

00:34

upon what Ila's status was at opening

00:37

eye but we finally have the news he says

00:39

that Ilia and open ey are going to part

00:41

ways this was very sad to me Ilia is

00:44

easily one of the greatest minds of Our

00:46

Generation a Guiding Light of our field

00:49

and a dear friend his Brilliance and

00:51

vision are well known his warmth and

00:53

compassion are less welln but no less

00:55

important opening ey would not be what

00:57

it is today without him although he has

00:59

some personally meaningful he is going

01:01

to work on I am forever grateful for

01:03

what he did here and committed to

01:05

finishing the mission we started

01:06

together I am happy that for so long I

01:09

got to be close to such a genuinely

01:11

remarkable genius and someone so focused

01:14

on getting the best future for Humanity

01:16

Jacob is going to be our new Chief

01:18

scientist jackob is easily one of the

01:20

greatest minds of Our Generation and I'm

01:22

thrilled he's taking the Baton here he

01:24

has run many of our most important

01:26

projects and I'm very confident he will

01:28

lead us to make rapid and safe progress

01:29

press towards our mission of ensuring

01:32

that AGI benefits everyone so that

01:34

statement there clearly shows that Elia

01:36

satova is no longer working at open aai

01:39

and is going to be working on something

01:41

else and this is something that at least

01:44

we can say now we have some kind of

01:45

closure on where one of the greatest

01:47

mines in AI where they are going to be

01:50

now one of the things I did want to know

01:52

is I wanted to know that who is going to

01:54

be replacing Elia satova at open Ai and

01:57

that is of course jackob now essentially

02:00

if we look at who jackob is jaob is now

02:03

the chief scientist at open aai where he

02:05

has led transformative research

02:07

initiative since 2017 he has previously

02:10

served as director of research

02:12

spearheading the development of GPT 4

02:14

and open A5 and fundamental research in

02:17

large scale reinforcement learning and

02:19

deep learning optimization he has been

02:21

instrumental in refocusing the company's

02:23

Vision towards scaling deep Learning

02:25

System and jackob holds a PhD in

02:27

theoretical computer science from

02:28

Carnegie melan University so clearly

02:31

this is someone with a very very

02:34

impressive resume and clearly has all of

02:36

the necessary skills to take on the AI

02:39

Niche now as for Elia satova something

02:42

that I know many people have been

02:44

wondering is what did he say and he

02:46

posted this tweet after quite some time

02:49

he said after almost a decade I have

02:51

made the decision to leave open AI the

02:54

company's trajectory has been nothing

02:55

short of miraculous and I'm confident

02:58

that open AI will build AGI that is both

03:00

safe and beneficial under the leadership

03:02

of Sam mman Greg Brockman miror moratti

03:05

and now under the excellent research

03:07

leadership of jackob it was an honor and

03:10

a privilege to have worked together and

03:12

I will miss everyone dearly so long and

03:14

thanks for everything I am excited for

03:16

what comes next a project that is very

03:18

personally meaningful to me about which

03:19

I shall share details in due time so

03:23

clearly Ilia seems to have left on good

03:25

terms despite the entire tumultuous

03:29

period that was the firing of Sam Alman

03:32

but I think one of the most interesting

03:34

things that most people are looking

03:36

forward to now is of course what comes

03:38

next so it says that he is excited for

03:41

what comes next and it's a project

03:43

that's very personally meaningful to me

03:45

about which I will share details in due

03:47

time so whatever Ilia satova is going to

03:49

do next I'm guessing that we will

03:51

receive an update I guess maybe a few

03:53

weeks could be a few months I have no

03:54

idea but it seems that we're going to be

03:56

getting an update sometime in the near

03:59

future now essentially there was also

04:01

this Peach whilst things on the surface

04:04

might look like opening eyes completely

04:05

fine the following slides that I'm about

04:07

to show you do showcase an entirely

04:09

different picture because open AI has

04:12

been losing key members of their most

04:14

important team in regards to AI safety

04:17

and I'm about to break this all down for

04:18

you because when I started to do the

04:20

research on this I was like wow things

04:22

are starting to look a little bit

04:25

pessimistic in terms of AI safety now

04:28

Ilia also did tweet this pict picture

04:30

with the rest of the open AI team but I

04:32

do remember that during December there

04:34

were some tweets by IIA satova where

04:36

there were some tweets that were vaguely

04:38

worded in a way that kind of implied

04:41

that openai was in a toxic work

04:43

environment he said that I learned many

04:44

lessons this past month one such lesson

04:47

is that the phrase the beatings will

04:48

continue until morale improves applies

04:50

more than than it has any right to and

04:53

this of course could be just dubious

04:55

speculation with regards to what he was

04:56

talking about but during the time that

04:58

this was tweeted it could be argued that

05:00

it was only related to one key event and

05:03

that of course was open AI at that time

05:06

but there was not really many statements

05:08

in addition to this because of the

05:10

secrecy of why Sam one was fired now

05:13

here's where things get really really

05:15

crazy most people don't know about super

05:17

alignment because it's something that is

05:19

in the future so if you don't know what

05:21

super alignment is this is opening I's

05:24

plan to solve super intelligence and

05:26

that is basically a system that is much

05:29

better than AGI so it says our goal is

05:31

to solve the core technical challenges

05:33

of super intelligence alignment in 4

05:35

years and basically what they did was

05:38

they decided to build a specific team to

05:41

solve this specific problem because they

05:43

knew that in the future they're going to

05:44

have AGI and after AGI comes ASI which

05:47

is artificial super intelligence and

05:49

super intelligence is quite hard to

05:52

explain but just think of it like this

05:54

if AGI can do a better task than any

05:56

human at pretty much everything and it's

05:58

going to be everywhere artificial super

06:00

intelligence is going to be able to do

06:03

things that you can't even fathom it's

06:04

going to be able to create new knowledge

06:06

do new research discover cures for

06:09

certain incurable diseases at the moment

06:11

it's pretty much going to feel like a

06:13

magical time if we manage to get super

06:15

intelligence right now essentially with

06:18

super intelligence you've got the

06:19

problem because you're Building A system

06:20

that is that smart it could go Rogue and

06:22

if a super intelligence goes Rogue we

06:24

literally don't stand a chance because

06:26

if a system is super intelligent it's

06:28

going to be able to outsmart us and

06:30

we're not going to be able to understand

06:31

what its goals are or even what it's

06:33

doing and you can see here that they

06:35

said that this might not even work it

06:36

says whil this is an incredibly

06:38

ambitious goal we're not guaranteed to

06:40

succeed we optimistic that a focused

06:42

concentrated effort can help solve this

06:44

problem there are many ideas that have

06:46

shown promise in preliminary experiments

06:49

and we have increasingly useful metrics

06:51

for progress and we can use today's

06:52

models to study many of these problems

06:54

empirically so here's where things start

06:56

to fall apart for open AI okay here's

06:59

where things really start to you know

07:01

like the alarm bills really start to go

07:02

so elas atava okay has made this core

07:05

research his focus and will be

07:07

co-leading this team with Jan like the

07:09

head of alignment okay and it says

07:11

joining the team are researchers and

07:13

Engineers from previous alignment teams

07:15

as well as other researchers across the

07:16

company now the thing is okay this super

07:19

alignment team that was meant to figure

07:21

out how to solve the alignment problem

07:23

for artificial super intelligence okay

07:25

elas satava is now gone and Jan like

07:28

today actually quit quit Okay so two key

07:31

members of the super alignment team are

07:33

now gone you can see that earlier today

07:35

Jan like literally tweeted I resigned

07:37

okay now the thing that I think was the

07:40

most interesting about this was the fact

07:42

that Janik didn't say anything other

07:44

than I resigned remember when Elia

07:46

satova resigned he said this heartfelt

07:48

meage that you know shows that he kind

07:51

of cares about opening eye and where it

07:52

goes but Jan like stating that you know

07:54

I resigned with you know just no further

07:57

context I mean it leaves it open to

08:00

complete speculation with as to why he

08:02

did resign now like I said before this

08:04

is a problem because these were the

08:06

people that were trying to solve super

08:08

alignment but you have to understand

08:10

that this does actually get a lot worse

08:12

because they're not the only people that

08:14

resigned from Super alignment in terms

08:16

of the team now one of the things that

08:18

you need to know about super alignment

08:19

and a lot of people are starting to

08:21

speculate is that because the head of

08:24

alignment at Super alignment resigned

08:26

today maybe this means that they solved

08:29

the alignment problem okay in terms of

08:31

aligning a super intelligence so if you

08:33

don't believe this let's take a look at

08:35

what Jan like said in a recent interview

08:37

now I've listened to this it's over two

08:39

hours but this is the main bit that you

08:41

need to pay attention to so he says if

08:43

you're thinking about how you align the

08:45

super intelligence how do you align a

08:46

system that's vastly smarter than humans

08:49

I don't know I don't have an answer I

08:50

don't think anyone really has an answer

08:52

but that's also not the problem that we

08:54

fundamentally need to solve maybe this

08:56

problem isn't even solvable by humans

08:59

who live today but there is this easier

09:01

problem how do you align the system that

09:03

is in the Next Generation how do you

09:05

align GPT n+ one and that is a

09:08

substantially easier problem so it's

09:10

basically saying how do you align

09:11

currently GPT 5 then when you have GPT 5

09:14

how do you align gp6 when you have GPT 6

09:17

how do you align gpt7 so he's basically

09:20

stating that that is a much easier

09:22

problem and he goes into more detail

09:23

here and he says okay and this is

09:25

basically the highlighted part and so if

09:27

you get a virtual system to be aligned

09:29

it can then solve the alignment problem

09:31

for GPT n plus1 and then you can itly

09:33

bootstrap yourself until you're at Super

09:35

intelligence level and you figured out

09:37

how to align that so basically he's

09:39

stating that even if humans can solve

09:41

the problem of GPT n plus1 basically

09:44

humans shouldn't be doing that but a

09:45

virtual system as smart as human should

09:47

be doing that and then if you get that

09:49

system to be aligned it can then solve

09:51

the alignment problem for GPT n plus1

09:54

and then you can itely bootstrap

09:55

yourself until you actually at Super

09:57

intelligence if you're thinking about

09:59

like how do you actually align a super

10:00

intelligence how do you align the system

10:02

that's vastly smarter than humans I

10:04

don't know I don't have an answer I

10:06

don't think anyone really has an answer

10:08

but it's also not the problem that we

10:10

fundamentally need to solve right

10:12

because like maybe this problem isn't

10:14

even solvable by like humans who live

10:16

today but there's this like easier

10:19

problem which is like how do you align

10:21

the system that is the next Generation

10:24

how do you align GPD n plus1 and that is

10:27

a substantially easier problem

10:29

and then even more if humans can solve

10:32

that problem then so should a virtual

10:36

system that is as smart as the humans

10:39

working on the problem and so if you get

10:42

that virtual system to be aligned it can

10:44

then solve you know the alignment

10:47

problem for GPT n plus one and then you

10:49

can iteratively bootstrap yourself until

10:52

you you know actually you're like at

10:55

Super intelligence level and you figured

10:57

out how to align that and of course

10:59

what's important when you're doing this

11:01

is like at Ed each step you have to make

11:04

enough progress on the problem that

11:05

you're confident that gbd n plus one is

11:08

aligned enough that you can use it for

11:10

alignment research and he says of course

11:13

what's important when you're doing this

11:15

is that at each step you have to make

11:16

enough progress on the problem that

11:18

you're confident that GPT n plus1 the

11:20

next model whatever it is is aligned

11:22

enough so that you can actually use it

11:24

for alignment research so basically what

11:26

he's saying is that every time we go an

11:27

increase from GPT 4 to GPT or whatever

11:30

next Frontier Model is we have to make

11:32

sure that that model is so aligned that

11:33

we can then use that current model for

11:36

alignment research but this is why a lot

11:38

of people now have thought okay that if

11:41

two of the key members of the founding

11:43

members of super alignment have now left

11:45

okay remember these guys left they

11:47

didn't get fired these guys left that

11:49

means and remember he said I resigned

11:51

okay with nothing no further context

11:53

that means that maybe they actually

11:54

managed to solve this alignment problem

11:56

okay and if you think that it's not that

11:58

bad take a look at this this person

12:00

Leopold who used to work at Super

12:03

lineman at open ey he also no longer

12:05

works at open ey now he was actually

12:08

fired okay it says open ey has fired two

12:11

researchers for allegedly leaking

12:13

information according to a person with

12:15

knowledge of the situation okay Leopold

12:18

and someone else okay who was also an

12:21

ally of Ilia satova who participated in

12:23

a failed effort to force Sam mman last

12:25

fall and you can also see here it says

12:27

opening ey staffers had actually

12:29

disagreed on whether the company was

12:30

developing AI safely enough now what's

12:33

crazy about this okay is that the two

12:35

people that were fired recently at open

12:38

they were actually members of the super

12:39

alignment team so when you look at the

12:41

picture now you can see that in the

12:44

recent paper where openai were talking

12:45

about weak to strong generalization

12:47

eliciting strong capabilities with weak

12:49

supervision this is basically a paper

12:52

where they're trying to solve super

12:53

alignment and they showing how it can be

12:55

done I've crossed out four names here

12:57

because these four people no long work

12:59

at open ey we can see pavl Leopold Jan

13:02

like and Elia Sova none of these people

13:05

work at open ey anymore this is

13:06

basically a paper where they're trying

13:08

to solve super alignment and they're

13:09

showing how it can be done I've crossed

13:11

out four names here because these four

13:13

people no longer work at open ey we can

13:16

see pavl Leopold Jan like and elas Sova

13:19

none of these people work at open ey

13:21

anymore pavl who was a you know

13:23

researcher okay who was on the previous

13:25

paper and super AI alignment open AI he

13:28

actually now works at Elon musk's AI

13:30

company so it will be interesting to see

13:32

how his career also develops now what's

13:34

also crazy is that you might be thinking

13:36

okay just a few people from Super

13:38

alignment left that might not be that

13:39

crazy well that's not the truth as well

13:42

as I dug into more things I realized

13:44

that more people from open ey also left

13:46

as well so you can see right here that

13:48

it says open ey researchers Daniel and

13:51

William recently left the company behind

13:54

C chat GPT it says Daniel said on a

13:56

forum he doesn't think that open ey will

13:58

behave responsibly around the time of

14:00

AGI so this is a key reason and I think

14:04

it's important to understand that when

14:06

people leave we have to look at why they

14:07

leave okay if someone's leaving for you

14:09

know family reasons or personal reasons

14:11

that is completely different but if

14:13

someone's stating that they're leaving

14:14

this company because he doesn't believe

14:16

that they'll behave responsibly around

14:17

the time of AGI that is a key key key

14:21

indicator that maybe just maybe

14:23

something is wrong you can see here it

14:25

says Daniel was on the governance team

14:28

and Saunders were worked on super

14:29

alignment team at openi so this is five

14:33

people from the original Super alignment

14:35

team have now completely left open eye

14:38

and that is remarkable considering the

14:40

fact that since the super alignment team

14:42

was formed it was meant to solve super

14:44

alignment within 4 years but literally

14:46

five of the founding members have gone

14:48

and four of them that were on this

14:50

recent paper are no longer there and you

14:52

have to understand that super alignment

14:53

is one of the key things that we need to

14:55

solve if we're going to increase air

14:57

capabilities which is why many is

14:59

speculating that the GPT n+1 alignment

15:02

problem has been solved since it was a

15:04

substantially easier problem to solve

15:06

now one thing that did really concern me

15:09

was Daniel's actual post about magical

15:13

capabilities and other futuristic things

15:16

that artificial super intelligence can

15:18

do and this is something that I did

15:21

Cover in another video which I'm going

15:23

to include here now but the point is as

15:25

I'm about to explain to you how crazy

15:27

this is about to get guessing that if we

15:30

take a look at what we've seen here

15:32

which I'm about to explain and the fact

15:34

that members of the super lineman team

15:36

are just gone like literally five six

15:37

people from the team are now gone some

15:39

people are speculating whether or not

15:41

this problem has solved and now it's

15:43

literally just a compute problem and

15:44

we're just building data centers because

15:46

we already know how to get to AGI but

15:49

anyways if you want to take a look at

15:51

what Daniel said because he said he

15:53

doesn't believe that open a are going to

15:55

uh you know behave responsibly around

15:57

the time of AGI because it's a very

15:59

powerful tool that will be allowed to do

16:00

pretty much anything in terms of the

16:02

fact that it's going to grant them

16:03

immense power and will literally shift

16:05

Dynamics in certain countries I think

16:07

you need to take a look at how crazy

16:10

this document is what Daniel said before

16:12

he left open a eye because it was truly

16:14

eye opening to see what he thinks the

16:16

future is going to be like intense so

16:18

we're going to go through this list

16:19

because there actually is quite a lot of

16:21

things to talk about and a lot of things

16:23

that you do need to be aware of because

16:25

a lot of people that saw this list

16:26

thought about a few things but didn't

16:28

think about how the industry is going to

16:30

evolve as a whole so let me tell you why

16:32

this was actually genuinely a shocking

16:34

statement because there was one of them

16:36

that I saw and I was like okay that is a

16:39

super big deal and that just completely

16:40

changes my timeline so one of the things

16:42

that he did state is essentially

16:45

probably there will be AGI coming soon

16:47

and that's any year now and this is

16:50

something that unfortunately doesn't

16:51

surprise me if you've been paying

16:53

attention to this base you'll know that

16:54

we've had many different instances and

16:57

inklings of AGI and essentially many of

17:00

us do kind of feel like we're on the

17:02

edge of our seat because we know that

17:04

AGI is just around the corner now that's

17:06

for a variety of factors like the fact

17:08

that openi has been you know in a really

17:10

really strange position with it you know

17:12

delaying releases on multiple products

17:14

with it having the Sam Alman firing um

17:17

and other companies are also having

17:19

major breakthroughs and working on AGI

17:21

as well so um literally any year now we

17:24

could definitely be getting AGI not only

17:26

just because of the way that these

17:27

companies are but the significant

17:29

investment that they're also getting to

17:31

now he also uh spoke here and you can

17:33

see that he responded to someone's

17:35

question in terms of the percentage of

17:37

AGI so he said why do you have a 15%

17:40

chance for 2024 and only an additional

17:42

15 for 2025 now I do think we get AGI by

17:45

the end of 2025 or at least you know

17:48

some kind of lab makes an insane

17:50

breakthrough and we have AGI by the end

17:52

of 2025 that's just what I kind of

17:53

believe but he says do you really think

17:55

that there's a 15% chance of AGI this

17:57

year and he says yes I really do I'm

17:59

afraid I can't talk about all of the

18:00

reasons for this you know I work at open

18:02

AI but mostly it should be figureoutable

18:04

from the publicly available information

18:06

which we've discussed several times on

18:07

this channel now my timelines were

18:09

already fairly short 2029 is the median

18:11

which is essentially the most common and

18:13

when I joined openi in early 2022 and

18:15

the things have gone mostly as I've

18:17

expected I've learned a bunch of stuff

18:19

some of which updated me towards and

18:21

some of which updated me downwards as

18:23

for the 15 15% thing I don't feel

18:26

confident that those are the right

18:27

numbers rather those numbers express my

18:29

current state of uncertainty I could see

18:31

the case for making 2024 number higher

18:34

than the 2025 of course because

18:36

exponential distribution Vibes if it

18:38

doesn't work now then there evidence it

18:39

won't works next year and I could also

18:41

see the case for making 2025 year higher

18:44

of course projects take twice as long as

18:46

one expects due to the planning fallacy

18:47

so essentially he's stating that you

18:49

know 15% this year 30% chance next year

18:51

but of course he's saying that you know

18:52

he could be completely wrong now with

18:54

AGI a predictions of course it's

18:56

anyone's guess but um this prediction by

18:58

AR invest essentially is a good visual

19:01

media to kind of look at when you're

19:03

looking at how AGI is going to progress

19:05

in the next couple of years now

19:06

something I do want to say about this is

19:08

the exponential nature of things because

19:10

they also do you know take that into

19:11

account with the fact that essentially

19:13

it was predicted for the end of uh you

19:15

know the 2029 or 2030 which is where

19:18

many people have predicted it I'm not

19:20

going to get into the next point in a

19:21

moment which is really what just kind of

19:23

shocked me um but essentially you can

19:24

see here that every time you know as as

19:27

time is going down you can see that

19:28

we're going down like this way as the

19:30

for forecast era continues it seems that

19:32

by 2027 but of course we can see that

19:34

there are these huge drop offs where

19:35

technology kind of just keeps on

19:37

dropping off so of course it's like

19:39

exponential it's kind of like s-curve

19:40

growth where you kind of go up and then

19:42

you kind of plateau for a little bit and

19:43

then you kind of get that next boom once

19:45

there is that next thing we're kind of

19:46

seeing like an inverted S curve um on

19:48

that graph as well and I know I showed

19:50

this in a previous video just wanted to

19:51

show it again so you guys can visualize

19:52

where things are going so the median the

19:54

most common prediction is 2029 some

19:56

people are predicting next year and

19:57

there are a few small reasons why but I

19:59

definitely do believe that if anyone is

20:00

going to get to it it will be open AI

20:02

obviously because of um the kind of

20:04

talent that they have the kind of you

20:05

know researches that they have it's a

20:06

unique team and open AI researchers are

20:09

some of the most sought after Talent

20:11

like you know um essentially it's so

20:13

crazy that you know I think there was

20:14

someone recently that was hired from uh

20:16

Google that went to a different company

20:18

and then Google started paying the

20:19

person four times more just because AI

20:20

researchers are so in demand right now

20:22

because it's such a competitive space

20:24

that um there is one tweet that I do I'm

20:26

going to come back to again that I think

20:27

you guys need to understand that if

20:29

someone develops AGI I think you guys

20:31

have to understand that it's going to be

20:32

a Winner Takes all scenario because it's

20:34

only a race until one company reach

20:36

reaches AI once once that happens the

20:38

distance to their competitors is

20:40

potentially infinite and it will be

20:41

noticeable for example raising a quar

20:43

trillion of the US GP now of course some

20:46

people are stating that this is where

20:47

you know open AI has already achieved

20:48

AGI they're just trying to raise raise

20:50

compute because they realized that we're

20:52

going to need a lot more compute for

20:53

this kind of system but Others May

20:55

disagree and I kind of do agree that you

20:57

know potentially um they probably have

20:59

internal AGI but just need more compute

21:01

to actually bring the system to reality

21:03

because certain things they just simply

21:05

can't test because they're trying to run

21:06

GPT 4 they're also trying to run some

21:08

other systems like Sora they're also

21:09

trying to to to give some of their

21:11

compute to Super alignment um so that is

21:14

a thing as well now this is the

21:15

statement that really did Shock Me Okay

21:18

um and this is why uh I my timelines got

21:21

updated because it changes everything

21:23

okay so it said he says uh probably

21:26

whoever controls AGI will be you will be

21:28

able to use it to get to artificial

21:30

super intelligence shortly thereafter

21:32

maybe in another year give it or take a

21:34

year now you have to understand that AGI

21:37

is essentially a robot that is

21:38

apparently as good as all

21:41

humans pretty much any task okay so

21:43

pretty much any task you can think of um

21:45

that can be done in a non-physical realm

21:47

and AGI is going to be able to be better

21:49

than 99% of humans okay according to um

21:52

deep mined levels of AGI paper and

21:54

essentially right now we have gbt 4

21:56

which is a lowlevel AGI so when we do

21:58

get that AGI system it's going to

22:00

accelerate everything because it means

22:01

that you know we can just duplicate

22:03

researchers we can duplicate

22:04

mathematicians we can duplicate people

22:06

um doing a whole bunch of stuff okay and

22:08

essentially this is crazy because

22:10

artificial super intelligence is

22:12

completely Next Level artificial

22:14

superintelligence is an intelligence

22:16

that is so smart that it will be able to

22:17

just make consistent breakthroughs and

22:19

it's going to fundamentally change our

22:21

understanding of everything we know

22:22

because because it's going to be that

22:23

smart okay and that is of course a

22:26

problem because of course there's

22:27

alignment problems and stuff like that

22:28

but the problem is is that artificial

22:29

super intelligence is something that

22:31

people didn't really even talk about

22:32

because it's so it's seemingly so far

22:34

away but they're sating that whoever

22:37

controls AGI will be able to use it to

22:38

get to ASI shortly after so if it's a

22:40

true AGI like a really good one um

22:43

getting to ASI won't take that long and

22:45

that is a true statement and something

22:46

that I didn't think about that much but

22:48

it's crazy because super intelligence

22:49

open I have openly stated that um super

22:52

intelligence will be the most impactful

22:53

technology Humanity's ever invented and

22:55

could help us solve many of the world's

22:57

most important problems but the vast

22:59

power of superintelligence could also be

23:01

very dangerous and it could lead to

23:03

disempowerment of humanity or even human

23:05

extension and it states that while super

23:07

intelligence seems far off now we

23:09

believe it could arrive this decade and

23:11

that's why this is kind of shocking

23:13

because open eye are saying that you

23:14

know okay some people think that AGI is

23:16

going to be by 2029 but they're stating

23:18

that not AGI by 29 2029 we state that

23:20

super intelligence could be here by the

23:22

end of this decade so super intelligence

23:24

could be here which means that you know

23:26

if we take a look and we kind of like

23:28

look at the actual you know data and we

23:30

think okay what's actually going on here

23:32

we could get AGI realistically by 2026

23:35

then we could get AGI by 2029 that's

23:38

something that could happen due to the

23:39

nature of exponential growth and these

23:41

timelines and open AI stated that

23:43

themselves so that's why they're also

23:44

actually working on that kind of

23:46

alignment because they know that it is

23:48

very very soon now in addition if you

23:50

want to talk predictions you have to

23:51

call on Rayo as well essentially he's a

23:53

futurist and he has made a lot of

23:56

predictions 147 and there's an 86% win

23:59

ratio I guess whatever you want to call

24:00

it now of course some people have you

24:03

know debated whether or not this ratio

24:05

is as high as he claims but um I would

24:07

say that his predictions have come true

24:09

a decent amount now essentially his

24:11

prediction on AGI is that artificial

24:14

intelligence will achieve human level by

24:16

2029 which is once again still going to

24:18

be pretty crazy even if it does happen

24:19

at 2029 because if we take a look at it

24:21

because I remember Elon Musk stated that

24:23

by 2027 everyone's timelines is getting

24:25

shorter and shorter by the day and we do

24:27

know that if we take a look at what's

24:29

actually going on right now um if we had

24:31

AGI within 2 years it's something that

24:32

genely wouldn't surprise everyone

24:34

especially with what we saw with s and

24:36

now another thing about um you know Ray

24:38

kwell that he actually stated that was

24:39

actually quite shocking and this is why

24:42

um I don't think you guys understand the

24:43

kind of world that we could be living in

24:45

if we actually do get AGI and then ASI

24:47

is because um he's stating that you know

24:49

there's a possibility that we might

24:50

achieve immortality by the year 2030 and

24:52

that's because of course we are like

24:54

doing well in terms of you know

24:56

longevity research and that kind of

24:57

stuff but if we do have artificial super

24:59

intelligence it's going to allow us to

25:00

do a lot of things like a lot of

25:02

breakthroughs that are just going to

25:03

completely change everything and that's

25:05

why this is so shocking because I didn't

25:07

realize that it could only take a year I

25:10

I don't know I mean I think that maybe

25:12

people aren't thinking about things such

25:13

as you know the actual compute the

25:15

actual you know laws in place that might

25:17

try to regulate this kind of stuff into

25:19

the ground the kind of uh maybe there's

25:21

going to be some kind of I guess you

25:22

could say Financial crashes or

25:24

essentially other things that could

25:25

potentially stop this but provided

25:27

everything is smooth like like there's

25:28

no you know Black Swan event there's no

25:30

like Bubonic plague the world doesn't

25:31

need to go into a shutdown and AGI

25:33

research isn't kind of delayed um ASI by

25:36

the end of the decade is a pretty scary

25:38

thing to think about okay and that is

25:39

why I stated that this genuinely did

25:41

shock me and one of the craziest things

25:43

as well like I said I was going to come

25:44

back to this okay how on Earth do other

25:46

companies catch up like I think think

25:48

about this okay so let's say your

25:50

opening eye okay you are working on

25:51

artificial general intelligence you do

25:53

it one day you wake up your researchers

25:55

you know your whole team is like look

25:56

we've done it we've achieved AGI we've

25:58

benchmarked it on all of this it's 99%

26:00

on this on that and that and that um

26:02

we've done it we've achieved AGI boom

26:04

okay how on Earth do other companies

26:06

catch up because the moment you get AGI

26:09

you can use it to I guess you could say

26:10

get towards ASI and you know you

26:13

immediately get like your company your

26:15

company just scales like 10x overnight

26:16

or even 100x overnight because all you

26:18

need to do is get the AGI to be able to

26:20

do certain things and you know it's

26:22

going to be relatively cheap to you in

26:24

terms of hiring another person that

26:26

you'd have to pay like a million a year

26:27

with open AI you could essentially have

26:29

these super powerful researchers doing

26:31

tons and tons of alignment research you

26:33

know ASI research and your company could

26:35

get an additional 100 employees every

26:36

day as long as you're scaling with

26:37

compute how on Earth do other AI

26:40

companies catch up to a company that's

26:42

basically achieved escape velocity and I

26:45

don't think they will like I genuinely

26:46

don't think that other companies will

26:48

catch up unless they quickly unless you

26:50

know somehow it leaks and the agit Tech

26:52

is you know widely distributed and then

26:54

of course you know when I say AGI Tech

26:56

I'm actually talking about the fact that

26:57

the agite tech is going to be the

26:59

research papers and the research behind

27:01

it not you know opening eye giving you

27:03

restrained access like they do with gbt

27:05

4 because the verion that we even get a

27:07

very very nerfed down model to what the

27:09

raw capabilities of the models offer so

27:11

essentially um you know this is some of

27:14

anthropics pitch deck when they wanted

27:16

to raise money in 2023 and they

27:19

basically said that we believe that

27:20

companies that train the best 2025 to

27:22

2026 models will be too far ahead for

27:25

anyone to catch up in subsequent Cycles

27:27

so um and if you don't know who

27:29

anthropic are they're a big AI company

27:31

that is kind of competing with openi

27:32

some of the open AI researchers did

27:34

leave to create anthropic because they

27:36

wanted to focus on safety but the point

27:38

here is that I don't think they catch up

27:39

and it does make sense if you have a

27:41

company that has AGI they have arguably

27:43

the best technology in in the last 20

27:45

years and with that they can grow their

27:47

company exponentially so I don't think

27:49

people catch up I think it's just you

27:51

know they're going to be so far gone

27:52

that it's going to be pretty crazy to

27:54

see what happens and I think the reason

27:56

that this is a thing is because this is

27:58

why people are stating that openi have

28:00

achieved Ai and they're currently using

28:01

it to develop things like Sora and stuff

28:03

and if that is true it kind of does make

28:05

sense cuz Sora definitely blew my hat

28:07

off like it's just like whoa like even

28:09

as someone who looks at AI all the time

28:11

when I saw that I was like whoa okay I

28:12

didn't think we were that close but um

28:14

yeah it's it's definitely pretty crazy

28:16

and um it it goes on okay so here it

28:18

states Godlike Powers okay so it says

28:20

probably whoever controls ASI listen to

28:22

this this is the craziest bit that I was

28:23

reading this and I was like is this even

28:25

real am I even living in a reality right

28:26

now it says probably who whoever

28:28

controls artificial superintelligence

28:30

will have access to spread to a spread

28:32

of powerful skills and abilities that

28:34

will be able to build and wield

28:36

technologies that seem like magic to us

28:38

just as modern tech would seem like

28:40

magic to medievals this will probably

28:42

give them Godlike Powers over whoever

28:44

doesn't control ASI so that brings an

28:46

important question do you think open AI

28:48

let's say they have ASI they have it

28:50

aligned do you think open AI are going

28:52

to distribute ASI or are they just going

28:53

to you know patent all the Technologies

28:55

as a kind of subsidiary of open as of

28:57

open cuz if they have ASI and nobody

29:00

else has it that's going to be the most

29:01

valuable thing on the planet and if

29:03

they're able to distribute cure if

29:05

they're able to distribute you know new

29:06

technology I mean that's going to make

29:08

the company super super super valuable

29:10

because like it states here they're

29:12

probably going to give them Godlike

29:13

Powers over anyone who doesn't control

29:14

ASI because that level of smartness is

29:17

unfathomable like it's very hard to

29:19

conceptualize how smart it is because

29:21

according to several reports and you

29:22

know researchers and stuff like that

29:23

it's basically like trying to explain

29:26

economics to essentially a b like you

29:29

know a b the thing that buzzes around

29:30

try to explain economics to that it it

29:33

it's it's very hard to conceptualize how

29:35

you would even begin to explain that to

29:37

a be I mean first you'd have to teach it

29:39

English then you'd have to teach it so

29:40

many other different concepts and um

29:42

that is going to be something that is

29:44

pretty pretty crazy I mean I mean trying

29:46

to even teach it abstract context so um

29:49

whilst this does seem good and whilst

29:50

you know Godlike powers and stuff like

29:52

that and you know which is why all these

29:53

companies are racing to achieve AGI

29:55

because they know once that is there

29:57

it's like you gain an instant 100x speed

29:59

boost in this kind of race the problem

30:01

is is that there is the blackbox problem

30:04

and a lot of people are starting to

30:06

forget about this problem as we Edge

30:08

closer and closer towards the edge of

30:09

this huge Cliff that we could be on um

30:12

is the fact that it states in general

30:13

there's a lot we don't understand about

30:15

modern deep learning modern AIS are

30:17

trained not built SL programmed we can't

30:20

theorize for example that they are

30:22

generally robustly helpful and in and

30:24

honest instead of just biing their time

30:26

we can't check so the problem here is

30:28

that um we don't know how these AR

30:30

models work we actually don't know

30:32

what's inside them we don't know how

30:35

everything is going together it's not

30:36

like you write a code and you understand

30:38

exactly how the code works this is not

30:40

how these AI models are going to be and

30:42

in the future um it's going to be a

30:43

bigger problem because if we're you know

30:45

growing an AI which is what some

30:46

researchers have claimed which is

30:48

essentially that would be a more

30:49

accurate description if we're doing that

30:51

how on Earth are we then going to

30:53

understand really um these even super

30:55

intelligent systems if we don't really

30:57

understand the ones we have now so um

30:59

it's pretty crazy it's it's generally

31:01

like like I'm trying hard to put it into

31:04

words but it is a very very giant

31:05

problem that people are trying to solve

31:07

and of course um here's we have this

31:09

okay so the alignment problem further

31:11

currently no one knows how to control

31:12

artificial super intelligence which is

31:14

true and they are working on it this is

31:15

what open ey is currently working on and

31:17

it says if one of our training runs

31:19

turns out to work way better than we

31:21

expect we'd have a rogue artificial

31:23

super Intelligence on our hands and

31:25

hopefully it would have internalized

31:27

enough human effects that things would

31:29

be okay and that's a crazy statement I

31:30

don't care what you say that is insane

31:32

because he's basically saying that look

31:34

if our training runs work out to be

31:36

better than we expect unfortunately

31:37

we're going to have a rogue ASI on our

31:39

hands because we don't know how to we

31:41

don't know how to align it they're

31:42

basically saying that look if we train

31:43

the next model and it's super smart or

31:45

artificially super intelligent which I

31:46

don't think it will be I do think that

31:48

you need a ton of compute just like how

31:49

you skilled things up with before it

31:51

says hopefully we're just we're just

31:52

basically just hoping that it's not

31:54

crazy okay um and that is quite scary

31:56

that you know hope is mentioned here so

31:58

it says there are some reasons to be

32:00

hopeful about that but there are also

32:01

some reasons to be pessimistic and the

32:02

literature on the topic is small and

32:04

preag dtic which is of course true then

32:07

of course we have um Sam Alman which is

32:09

a great clip which you guys should take

32:10

a look at because he actually talking

32:12

about um the alignment problem is like

32:13

we're going to make this incredibly

32:15

powerful system and be really bad if it

32:17

doesn't do what we want or or if it sort

32:20

of has you know goals that are uh either

32:23

in conflict with ours um many Sci-Fi

32:25

movies about what happens there or goals

32:27

where it just like doesn't care about us

32:28

that much and so the alignment problem

32:30

is how do we build AI that that does

32:34

what is in the best interest of humanity

32:36

how do we make sure that Humanity gets

32:38

to determine the you know the future of

32:40

humanity um and how do we avoid both

32:43

like accidental misuse um like where

32:45

something goes wrong that we didn't

32:46

intend intentional misuse where like a

32:48

bad person is like using an AGI for

32:50

great harm even if it that's what the

32:52

person wants and then the kind of like

32:54

you know inner alignment problems where

32:55

like what if this thing just becomes a

32:57

creature that views this as a threat the

32:59

the way that I think the self-improving

33:00

systems help us is not necessarily by

33:03

the nature of self-improving but like we

33:05

have some ideas about how to solve the

33:07

alignment problem at small scale um and

33:09

we've you know been able to align open