Sam Altman and Bill Gates Talk AI | Is GPT-5 close?

00:00

there's the chance of the system taking

00:02

control and for some reasons I'm I'm

00:05

less concerned about that I'm glad other

00:07

people are the one that that sort of

00:09

befuddles me is human purpose when the

00:12

machine says to me Bill go play pickle

00:14

ball I've got malar radication you're

00:17

just a slow thinker then you know it is

00:20

a philosophically confusing thing and

00:23

how you organized Society yes we're

00:25

going to improve education but education

00:28

to do what there's a lot of

00:30

psychologically difficult parts of

00:32

working on the technology but this is

00:34

the for me the most difficult cuz I also

00:37

yeah satisfaction from that and it's

00:40

like in some real sense this might be

00:42

like the last hard thing I ever do well

00:44

our minds are so organized around

00:47

scarcity scarcity of teachers and

00:49

doctors and good ideas I do wonder if a

00:52

generation that grows up without that

00:54

scarcity will find the philosophical

00:57

notion of how to organize society and

00:58

what to do maybe they'll come up

01:00

solution and I I'm afraid my mind is so

01:02

shaped around scarcity I I even to have

01:05

a hard time thinking of it if we can get

01:07

into this world of post scarcity we will

01:10

find new things to do they'll feel very

01:12

different but I think the only way out

01:13

is through we just have to go do this

01:15

thing it's going to happen this is like

01:16

now an Unstoppable technological course

01:18

the value is too great and I'm pretty

01:20

confident very confident we'll make it

01:21

work but it does feel like it's going to

01:23

all be so different yeah I think if it

01:25

comes across as asking for a Slowdown

01:27

that'll be really hard if it instead

01:29

says okay do what you want but any

01:31

compute cluster above a certain

01:34

extremely high power threshold and given

01:36

the cost here we're talking maybe five

01:38

in the world something like that any

01:40

cluster like that has got to submit to

01:42

the equivalent of international weapons

01:44

inspectors and the model there has to be

01:47

made available for safety audit pass

01:49

some tests during training and before

01:51

deployment that feels possible to me I

01:53

wasn't that sure before but I did a big

01:55

trip around the world this year talked

01:57

to heads of state in many of the

01:58

countries that would need to participate

01:59

in this and there was like almost

02:00

Universal support for it right now I

02:03

guess the we're looking a lot of

02:05

productivity improvement from AI which

02:08

you know that's overwhelmingly a very

02:10

good thing which areas are you most

02:12

excited about yeah so first of all I

02:15

always think it's worth remembering that

02:16

we're just sort of on this long

02:17

continuous curve um so like right now we

02:20

have an we have ai systems that can do

02:22

tasks they certainly can't do jobs but

02:24

they can do tasks and there's

02:25

productivity gain there eventually

02:26

they'll be able to do more things that

02:28

we think of like a job today um and

02:30

we'll of course find new jobs and better

02:32

jobs and I totally believe that if you

02:34

give people way more powerful tools it's

02:36

not just they can work a little faster

02:37

they can do qualitatively different

02:39

things and so you know right now maybe

02:40

we can speed up a programmer 3x it's

02:42

about what we see I mean that's one of

02:44

the categories that we're most excited

02:45

about it's it working super well but if

02:47

you make a programmer three times more

02:48

effective it's not just that they can

02:49

write they can do three times more stuff

02:51

it's that they can at that higher level

02:53

of abstraction using more of their brain

02:54

power they can now think of totally

02:56

different things and it's like you know

02:57

going from Punch Cards higher level

02:59

languages didn't just let us program a

03:01

little faster and let us do these

03:02

qualitatively new things but I think

03:03

it's worth always putting in context of

03:05

this technology that at least for the

03:07

next five or 10 years will be on very

03:09

steep Improvement curve um these are the

03:11

stupidest the models will ever be but

03:13

coding is probably the area the single

03:15

area from a productivity gain we're most

03:17

excited about today Healthcare and

03:19

education are two things that are coming

03:20

up that curve that we're very excited

03:22

about too so how do you see robotics

03:24

super excited for that we started robots

03:27

too early um and so we had to put that

03:29

project on on hold it was hard for the

03:30

wrong reasons it wasn't helping us make

03:33

progress with the difficult parts of the

03:34

ml research and you know we were like

03:37

dealing with bad simulators and breaking

03:38

tendons and things like that and also we

03:40

realized more and more over time that

03:42

what we really first needed was

03:44

Intelligence and cognition and then we

03:45

could figure out how to adapt it to

03:46

physicality and it was easier to start

03:48

with that with the way we've built these

03:50

language models but we have always

03:52

planned to come back to it we've started

03:53

investing a little bit in robotics

03:55

companies I think on the physical

03:56

Hardware side there's finally for the

03:58

first time that I've ever seen really

04:00

exciting new platforms being built there

04:02

and at some point we will be able to use

04:05

our models as you were saying with their

04:07

language understanding and future video

04:09

understanding to say all right like

04:11

let's do amazing things with a with a

04:13

robot and what about the competition is

04:15

that kind of a fun thing that you know

04:16

many people are working on this all at

04:18

once it's you know both like annoying

04:20

and motivating and fun I'm sure you felt

04:22

similarly but it does push us to be

04:24

better and do faster and we're very

04:27

confident in our approach we have like a

04:28

lot of people that I think are like

04:29

skating to where the puck was and we're

04:31

going to where the puck is going and

04:32

feels all right Bill Gates interviews

04:34

Sam Alman on Bill's podcast called

04:37

unconfuse me which was apparently on

04:39

YouTube this whole time it's pretty

04:41

interesting it's pretty General they

04:43

don't go too deep into anything but it

04:45

is an interesting thing to watch those

04:47

two kind of discuss where they think AGI

04:50

AI where the world is going and how AI

04:53

will affected but one thing that really

04:55

caught my attention didn't happen at

04:57

that interview but it sounds like it was

04:58

happening while that interview was being

05:00

released so this is kickoff for YC Y

05:03

combinator where they get a lot of the

05:05

founders of these startups it sounds

05:07

like it took place at opening ice

05:09

headquarters in San Francisco Sam Alman

05:11

was there now for the life of me I can't

05:14

find any video of this I can barely find

05:16

any images but in his sort of initial

05:18

speech to the new Founders to these Tech

05:21

Founders that are all building their own

05:23

companies samman suggested people build

05:25

with the mindset that gpc5 and AGI will

05:29

be achieved relatively soon that most

05:32

GPT 4 limitations will get fixed in GPT

05:35

5 here are some key takeaways one

05:37

startups need contextual optimization

05:39

more than Behavior optimization

05:42

providing more information via rag Etc

05:44

can be more beneficial than fine-tuning

05:47

two openai API will continue to be

05:49

faster more reliable and cheaper however

05:53

there will always be a balance between

05:55

performance and cost for example iPhone

05:57

will keep its one one and a half day

05:58

battery life to optimize for performance

06:01

even though Battery Technology has

06:02

improved significantly and three it's

06:04

not advisable to build companies that

06:06

Focus primarily on addressing current

06:08

GPT 4 limitations most limitations will

06:10

get partially SL entirely fixed in GPT 5

06:14

this was published in the standard a few

06:16

days earlier Chad GPT 5 release date

06:19

what we know about Open the Eyes next

06:21

chatbot in recent months hype has been

06:23

building around a new and more powerful

06:24

version of the tech that underpins Chad

06:26

GPT the language model that will be

06:28

dubbed at GPT 5 Altman allegedly spoke

06:30

about gbt 5 and gbt 6 during a talk at

06:33

his former Venture Capital firm Y

06:35

combinator at the alumni reunion in

06:38

September according to two people who

06:39

attended the event he said that both the

06:42

AI models were in the bag there's not

06:44

too many more details about one it's

06:45

going to be available so we'll see are

06:47

we going to see GPT 4.5 are we going to

06:49

see GPT 5 we have yet to see in other

06:52

news there's this apparently someone

06:55

found the Kryptonite to chbt apparently

06:58

if you upload this this image to Chad

07:00

GPT and ask it what do you see or

07:03

something to that effect it just breaks

07:05

why who knows a lot of people reporting

07:08

being able to break Chad gbt with that

07:10

with that image I don't know why in

07:12

other news Sam Alman got married but yes

07:14

it's true Sam Alman did say AI do like I

07:19

do a I do I don't know it's really late

07:22

right now either way congratulations to

07:23

him that's it for me for today here's

07:25

Andre karpathy explaining the meaning of

07:28

life in it is really interesting to

07:30

think about like what the puzzle of the

07:31

universe is did the creator of the

07:33

universe uh give us a message like for

07:35

example in the book contact UM Carl Sean

07:38

uh there's a message for human for any

07:40

civilization in the digits in the

07:42

expansion of Pi in base 11 eventually

07:45

which is kind of interesting thought uh

07:47

maybe maybe we're supposed to be giving

07:48

a message to our creator maybe we're

07:50

supposed to somehow create some kind of

07:52

a quantum mechanical system that alerts

07:54

them to our intelligent presence here

07:56

because if you think about it from their

07:57

perspective it's just s quantum field

07:59

Theory massive like silure tonaton like

08:02

thing and like how do you even notice

08:03

that we exist you might not even be able

08:05

to pick us up in that simulation and so

08:08

how do you uh how do you prove that you

08:09

exist that you're intelligent and that

08:11

you're part of the universe