OpenAI's Altman and Makanju on Global Implications of AI

Summary

Sam Altman and Anna Patterson of OpenAI discuss AI safety, regulation, energy use, and preparing for artificial general intelligence at a Bloomberg roundtable interview during the 2024 Davos conference. They talk about OpenAI's content moderation efforts, relationship with publishers and artists, the pace of AI progress, and the need for vastly more energy to power advanced AI while combatting climate change.

Takeaways

• OpenAI aims to restrict harmful uses of AI like misinformation while enabling beneficial ones.
• Demand for AI compute power will drive breakthroughs in fusion, solar and energy storage.
• AI will augment and enhance human productivity more than replace jobs.
• OpenAI seeks partnerships with news publishers to properly attribute content.
• Advanced AI may discover new scientific knowledge and even do AI research.
• Regulating AI risks stifling entrepreneurial innovation.
• AI progress will likely continue at an exponential, continuous pace.
• AI safety standards and best practices still need development.
• Well-designed AI can make technology feel more human-compatible.
• Preparing for transformative AI requires humility about the future.

Q & A

What content moderation efforts is OpenAI making regarding AI?

OpenAI is banning use of ChatGPT for political campaigns, adding cryptographic watermarks to AI images, and partnering with secretaries of state to surface authoritative voting information.

How does OpenAI view AI's impact on jobs and employment?

OpenAI sees AI as more likely to augment human productivity than replace jobs, though concedes AI may still significantly alter many occupations.

What is OpenAI's perspective on training AI systems?

OpenAI believes future AI systems will need vastly less training data than currently speculated, using smaller high-quality datasets rather than mass quantities of data.

How does OpenAI characterize the pace of AI progress?

OpenAI expects AI capabilities to improve at an exponential, continuous pace with each new system delivering impressive advances.

What is the greatest uncertainty regarding advanced AI?

What societal impacts will emerge when everyone has access to highly capable AI assistants and collaborators.

Why does OpenAI iteratively deploy AI systems?

To give people time to gradually adapt to the technology while mistakes can occur at low stakes, allowing co-evolution of technology and society.

What role can AI have regarding scientific knowledge?

Advanced AI systems may discover entirely new scientific knowledge that benefits humanity.

How can AI research best progress safely?

Through collaboration on safety standards and best practices among organizations like the new Frontier Model Forum.

What makes for good AI system design?

Systems that feel natural, intuitive and human-compatible in how they are interacted with.

What mindset is most constructive regarding transformative AI?

Cautious optimism along with humility about the difficulty inherent in predicting the future societal impacts of advanced AI.

Outlines

00:00

Discussion on AI Guidelines and Elections

Paragraph 1 covers a discussion between the speakers on OpenAI's newly announced guidelines restricting the use of AI tools like ChatGPT in political campaigns. They talk about how OpenAI plans to enforce these guidelines at scale using safety systems and partnerships.

05:00

Bipartisan Nature of AI Regulation

Paragraph 2 continues the discussion on AI regulation, with the speakers highlighting the bipartisan support and agreement so far on the need to regulate AI technology.

10:01

AI Applications for National Security

Paragraph 3 has the speakers talk about OpenAI's evolving policies around use of their AI for military and national security applications. They are open to cybersecurity and veteran wellbeing focused collaborations but not developing weapons.

15:04

Engaging Responsibly with Artists

Paragraph 4 covers the issues around use of AI generative models like DALL-E in art. The speakers talk about respecting artist preferences on use of their style and data while finding ways for them to benefit from and collaborate on the technology.

20:04

Learning from Past Controversies

In Paragraph 5, Anna talks about leveraging learnings from past industry controversies in government partnerships and getting policymakers to better understand AI technology from early on.

25:05

Preparing for Societal Impacts

Paragraph 6 has Sam highlight the potential for rapid advancement in AI capabilities and the resulting responsibility leaders have for thoughtful governance and policies to shape positive societal impacts.

30:06

Developing Safely with Humility

Sam continues on responsible development of AI in paragraph 7, emphasizing the need for humility and safety-focused iterative deployment to give society time to gradually adapt and government policies to evolve.

35:09

Possibilities of AI-Powered Devices

In the final paragraph, when asked about reports of collaborations with Elon Musk's Neuralink, Sam leaves open the possibility of building new kinds of AI empowered devices, though not as replacements for smartphones.

Keywords

💡Artificial General Intelligence (AGI)

AGI refers to AI systems that can perform any intellectual task that a human being can. It is considered the holy grail of AI research. Sam talks about AGI as the future capability he expects AI systems like GPT to achieve, where they can discover new scientific knowledge, do AI research, and integrate deeply into the economy. This would profoundly change society in ways we can't fully anticipate.

💡Compute

Sam mentions compute as one of the two key 'currencies' of the future along with intelligence. He believes massive amounts of computing power will be needed to develop AGI. This poses an energy and climate challenge, forcing breakthroughs in fusion power and renewable energy to supply AI systems sustainably.

💡Regulation

With the rapid development of AI, governments are starting to explore regulating it. Sam believes different countries will take different approaches initially. Anna discusses specific regulatory efforts in the EU and US. They agree regulation needs an iterative, experimental approach as AI and society co-evolve.

💡Governance

The governance of AI development companies is discussed, including OpenAI's nonprofit/for-profit structure. Sam admits OpenAI's governance needs reviewing after recent events, but wants to focus first on reforming the board of directors.

💡Safety

Ensuring AI is developed safely and responsibly is a core aim. OpenAI uses techniques like red teaming to anticipate and mitigate risks. But Anna notes there is no consensus yet on what AI safety means and entails.

💡Alignment

Sam stresses the importance of AI being human-compatible and human-focused - aligned with human values and creativity rather than autonomous. He is relieved GPT has so far been more of a beneficial tool than a disruptive agent.

💡Climate change

Sam acknowledges AI's voracious demand for compute power could drive climate change, unless matched by clean energy breakthroughs. He thinks AI's needs will force greater investment in fusion, renewables and storage.

💡Exponential growth

Sam argues people underestimate the implications of the exponential improvements in AI capabilities. Each new version represents a huge jump, so the pace of change will likely accelerate rapidly even if deployments remain gradual.

💡Uncertainty

A recurring theme is how little we can predict the societal impacts of more advanced AI. Sam emphasizes remaining humble and cautions against false precision in forecasts, since the effects are beyond human intuition.

💡Co-evolution

Sam and Anna both argue AI deployment should be iterative so that technology and society can co-evolve. This allows time for societal adaptation and policy experimentation to ensure AI benefits humanity.

Highlights

OpenAI introduced new guidelines banning ChatGPT use in political campaigns and adding cryptographic watermarks to AI-generated images.

Sam Altman doesn't believe AI will displace jobs as much as previously thought. He sees it more as a productivity-enhancing tool that lets people do more.

Anna Chen says governments are becoming more interested in understanding and potentially regulating AI, though most are not yet ready to incorporate it operationally.

Sam believes developing AGI will require a massive increase in energy production and breakthroughs like fusion or much cheaper solar and storage.

Sam thinks the world will change more slowly then more quickly as AI capabilities advance exponentially, but says no one knows exactly what happens next.

Anna sees government AI regulation accelerating in 2024 as more comprehensive policies like the EU AI Act and US executive order are implemented.

Sam wants leaders at Davos to understand 2023 woke people up to AI's potential, but what's coming with models like GPT-5 will be far more impressive and transformative.

Sam says open AI considers its systems AI assistants rather than standalone products or entities.

Sam expects open AI's next model after GPT-4 to be very impressive, doing new things not possible before and improving on GPT-4's capabilities.

Sam thinks new specialized AI devices could be created to augment humanity, but they won't replace general-purpose tools like smartphones.

Anna wants leaders at Davos to balance realistic concerns about AI with messaging that lets people engage with and benefit from its potential.

Sam believes the way AI products are designed matters hugely for making the technology feel accessible rather than scary or mystical.

Sam thinks no one has strong intuitions for what happens when AI becomes thousands of times cheaper and more capable.

Sam argues that gradual, iterative AI deployment builds public familiarity and helps society and policy co-evolve alongside rapid technical advances.

Anna thinks clearer industry standards are needed for concepts like AI safety that lack common definitions and approaches.

Transcripts

you guys made some news today announcing

some new guidelines around the use of AI

in elections I'm sure it's all uh stuff

that the the the davo set loved to hear

uh you banned the use of chat GPT in

political campaigns you introduced

cryptographic watermarks or images

created by do to to create kind of

Providence and transparency around the

use of AI generated images I read it and

I thought you know this is great some of

these principles are shared by much

larger platforms like Facebook and Tik

Tok and YouTube and they have struggled

to enforce it how do

you make it

real I mean these a lot of these are

things that we've been doing for a long

time and we have a really strong safety

systems team that um not only sort of

has monitoring but we're actually able

to leverage our own tools in order to

scale our enforcement which gives us I

think a significant Advantage um but uh

so there this there also some some

really important Partnerships like with

the National Association with the

secretaries of state so we can Surface

authoritative voting information so we

have quite a few ways that we are able

to enforce this mean Sam are you does

this put your mind at ease that we don't

that that open AI doesn't move the

needle in in some 77 upcoming critical

Democratic elections in 2024 no we're

quite focused on it uh and I think it's

good that our mind is not at EAS I think

it's good that we have a lot of anxiety

and are going to do everything we can to

get it as right as we can um I think our

role is very different than the role of

a distribution platform but still

important we'll have to work with them

too uh it'll you know it's like you

generate here and distribute here uh and

there needs to be a good conversation

between them but we also have the

benefit of having watched what's

happened in previous Cycles with

previous uh you know Technologies and I

don't think this will be the same as

before I I think it's always a mistake

to try to fight the last war but we do

get to take away some learnings from

that

and so I I wouldn't you know I I think

it'd be terrible if I said oh yeah I'm

not worried I feel great like we're

going to have to watch this incredibly

closely this year super tight monitoring

super tight feedback loop Anna you you

were at Facebook for open

AI so so I almost apologize for asking

it this in this way uh probably a

trigger phrase but do you worry about

another Cambridge analytical analytica

moment I think as Sam alluded to there

are a lot of learnings that we can

leverage but also open the eye from its

Inception has been a company that thinks

about these issues that it was one of

the reasons that it was founded so I

think I'm a lot less concerned because

these are issues that our teams have

been thinking about from the beginning

of uh our building of these tools Sam

Donald Trump just won the Iowa caucus

yesterday uh we are now sort of

confronted with the reality of this

upcoming election what do you think is

at

stake in the in the US election for for

Tech and for the safe stewardship of AI

do you feel like that's a a critical

issue that voters should and will have

to consider in this election I think the

now confronted as part of the problem uh

I actually think most people who come to

D say that again I didn't quite get that

I think part of the problem is we're

saying we're now confronted you know it

never occurred to us that what Trump is

saying might be resonating with a lot of

people and now all of a sudden after

this performance in Iowa oh man um it's

a very like Davos Centric you know um

I've been here for two days I guess

just uh so I I would love if we had a

lot more reflection and if we started it

a lot sooner um about and we didn't feel

now confronted but uh I think there's a

lot at stake at this election I think

elections are you know huge deals I

believe that America is going to be fine

no matter what happens in this election

I believe that AI is going to be fine no

matter what happens in this election and

we will have to work very hard to make

it so um but this is not you know no one

wants to sit up here and like hear me

rant about politics I'm going to stop

after this um but I think there has been

a real failure to sort of learn lessons

about what what's kind of like working

for the citizens of America and what's

not Anna I want to ask you the same

question uh um you know taking your

political background into account what

do you feel like for Silicon Valley for

AI is at stake in the US election I

think what has struck me and has been

really remarkable is that the

conversation around AI has remained very

bipartisan and so you know I think that

the one concern I have is that somehow

both parties hate

it no but you know this is like an area

where um

you Republicans tend to of course have a

an approach where they are not as in

favor of Regulation but on this I think

there's agreement on both parties that

they are consider they believe that

something is needed on this technology

you know Senator Schumer has this

bipartisan effort that he is running

with his Republican counterparts again

uh when we speak to people in DC on both

sides of the aisle for now it seems like

they're on the same page and do you feel

like all the existing campaigns are

equally articulate about the about the

issues relating to Ai No know that AI

has really been a campaign issue to date

so it will be interesting to see how

that if we're right about what's going

to happen here this is like bigger than

just a technological re ution in some

sense I mean sort of like all

technological revolutions or societal

revolutions but this one feels like it

can be much more of that than usual and

so it it is going to become uh a social

issue a political issue um it already

has in some ways but I think it is

strange to both of us that it's not more

of that already but with what we expect

to happen this year not with the

election but just with the the increase

in the capabilities of the products uh

and as people really

catch up with what's going to happen

what is happening what's already

happened uh there's like a lot of a Nur

always in society well I mean there are

political figures in the US and around

the world like Donald Trump who have

successfully tapped into a feeling of

yeah

dislocation uh anger of the working

class the feeling of you know

exacerbating inequality or technology

leaving people behind is there the

danger that uh you know AI furthers

those Trends yes for sure I think that's

something to think about but one of the

things that surprised us very pleasantly

on the upside uh cuz you know when you

start building a technology you start

doing research you you kind of say well

we'll follow where the science leads us

and when you put a product you'll say

this is going to co-evolve with society

and we'll follow where users lead us but

it's not you get you get to steer it but

only somewhat there's some which is just

like this is what the technology can do

this is how people want to use it and

this is what it's capable of and this

has been much more of a tool than I

think we expected it is not yet and

again in the future it'll it'll get

better but it's not yet like replacing

jobs in the way to the degree that

people thought it was going to it is

this incredible tool for productivity

and you can see people magnifying what

they can do um by a factor of two or

five or in some way that doesn't even

talk to makes sense to talk about a

number because they just couldn't do the

things at all before and that is I think

quite exciting this this new vision of

the future that we didn't really see

when we started we kind of didn't know

how it was going to go and very thankful

the technology did go in this direction

but where this is a tool that magnifies

what humans do lets people do their jobs

better lets the AI do parts of jobs and

of course jobs will change and of course

some jobs will totally go away but the

human drives are so strong and the sort

of way that Society works is so strong

that I think and I can't believe I'm

saying this because it would have

sounded like an ungrammatical sentence

to me at some point but I think AGI will

get developed in the reasonably

close-ish future and it'll change the

world much less than we all think it'll

change jobs much less than we all think

and again that sounds I may be wrong

again now but that wouldn't have even

compiled for me as a sentence at some

point given my conception then of how

AGI was going to go as you've watched

the technology develop have you both

changed your views on how significant

the job dislocation and disruption will

be as AGI comes into Focus so this is

actually an area that we know we have a

policy research team that studies this

and they've seen pretty significant

impact in terms of changing the way

people do jobs rather than job

dislocation and I think that's actually

going to accelerate and that it's going

to change more people's jobs um but as

Sam said so far it hasn't been the

significant a replacement of jobs you

know you hear a coder say okay I'm like

two times more productive three times

more productive whatever than they used

to be and I like can never code again

without this tool you mostly hear that

from the younger ones but

um it turns out and I think this will be

true for a lot of Industries the world

just needs a lot more code than we have

people to write right now and so it's

not like we run out of demand it's that

people can just do more expectations go

up but ability goes up

too goes up I want to ask you about

another news report today that suggested

that open AI was relaxing its

restrictions around the use of AI in

military projects and developing weapons

can you say more about that and you what

work are you doing with the US

Department of Defense and other military

agencies so a lot of these policies were

written um before we even knew what

these people would use our tools for so

what this was not actually just the

adjustment of the military use case

policies but across the board to make it

more clear so that people understand

what is possible what is not possible

but specifically on this um area we

actually still prohibit the development

of weapons um the destruction of

property harm to individuals but for

example we've been doing work with the

Department of Defense on um cyber

security tools for uh open source

software that secures critical

infrastructure we've been exploring

whether it can assist with veteran

suicide and because we previously had a

what essentially was a blanket

prohibition on Military many people felt

like that would have prohibited any of

these use cases which we think are very

much aligned with what we want to see in

the world has the US government asked

you to restrict the level of cooperation

with uh militaries in other

countries um they haven't asked us but

we certainly are not you know right for

now actually our discussion are focused

on um United States national security

agencies and um you know I think we have

always believed that democracies need to

be in the lead on this technology uh Sam

changing topics uh give us an update on

the GPT store and are you seeing maybe

probably explain it briefly and are you

seeing the same kind of explosion of

creativity we saw in the early days of

the mobile app stores yeah the same

level of creativity and the same level

of crap but it I mean that happens in

the early days as people like feel out a

technology there's some incredible stuff

in there too um give us an example the

gpts should I say what gpts are first

yeah sure um so gpts are a way to do a

very lightweight customization of chat

GPT and if you want it to behave in a

particular way to use particular data to

be able to call out to an external

service um you can make this thing and

you can do all sorts of like uh great

stuff with it um and then we just

recently launched a store where you can

see what other people have built and you

can share it and um I mean personally

one that I have loved is Al Trails I

have this like every other weekend I

would like to like go for a long hike

and there's always like the version of

Netflix that other people have where

it's like takes an hour to figure out

what to watch it takes me like two hours

to figure out what hike to do and the

all Trails thing to like say I want this

I want that you know I've already done

this one and like here's a great hike

it's been I it's sounds silly but I love

that one have you added any gpts of your

own have I made any yeah um I have not

put any in the store maybe I will great

um can you give us an update on the

volume or or the pace at which you're

seeing new gpts um the number I know is

that there had been 3 million created

before we launched the store I have been

in the middle of this trip around the

world that has been quite hectic and I

have not been doing my normal daily

metrics tracking so I don't know how

it's gone since launch but I'll tell you

by the slowness of chat GPT it's

probably doing really

well um I want to ask you about open

ai's copyright issues uh how important

are publisher relations to open ai's

business considering for example the

lawsuit last month file against open AI

by the New York Times They are important

but not for the reason people think um

there is this belief held by some people

that man you need all of my training

data and my training data is so valuable

and actually uh that is generally not

the case we do not want to train on the

New York Times data for example um and

all more generally we're getting to a

world where it's been like data data

data you just need more you need more

you need more you're going to run out of

that at some point anyway so a lot of

our research has been how can we learn

more from smaller amounts of very high

quality data and I think the world is

going to figure that out what we want to

do with Publishers if they want is when

one of our users says what happened to

Davos today be able to say here's an

article from blueberg here's an article

from New York Times and here you know

here's like a little snippet or probably

not a snippet there's probably some

cooler thing that we can do with the

technology and you know some people want

to partner with us some people don't

we've been striking a lot of great

Partnerships and we have a lot more

coming um and then you know some people

don't want want to uh we'd rather they

just say we don't want to do that rather

than Sue us but like we'll defend

ourselves that's fine too I just heard

you say you don't want to train on the

New York Times does that mean given the

the legal exposure you would have done

things differently as you trained your

model here's a tricky thing about that

um people the web is a big thing and

there are people who like copy from The

New York Times and put an article

without attribution up on some website

and you don't know that's a New York

Times article if the New York Times

wants to give us a database of all their

articles or someone else does and say

hey don't put anything out that's like a

match for this we can probably do a

pretty good job and um solve we don't

want to regurgitate someone else's

content um but the problem is not as

easy as it sounds in a vacuum I think we

can get that number down and down and

down have it be quite low and that seems

like a super reasonable thing to

evaluate us on you know if you have

copyrighted content whether or not it

got put into someone else's thing

without our knowledge and you're willing

to show us what it is and say don't

don't put this stuff as a direct

response we should be able to do that

um again it won't like thousand you know

monkeys thousand typewriters whatever it

is once in a while the model will just

generate something very close but on the

whole we should be able to do a great

job with this um so there's like there's

all the negatives of this people like ah

you know don't don't do this but the

positives are I think there's going to

be great new ways to consume and

monetize news and other published

content and for every one New York Times

situation we have we have many more

Super productive things about people

that are excited to to build the future

and not do their theatrics and and what

and what about DOI I mean there have

been artists who have been upset with

Dolly 2 Dolly 3 what what has that

taught you and how will you do things

differently we engage with the artist

Community a lot and uh you know we we

try to like do the requests so one is

don't don't generate in my style um even

if you're not training on my data super

reasonable so we you know Implement

things like that

um you know let me opt out of training

even if my images are all over the

Internet and you don't know what they

are what I'm and so there's a lot of

other things too what I'm really excited

to do and the technology isn't here yet

but get to a point where rather than the

artist say I don't want this thing for

these reasons be able to deliver

something where an artist can make a

great version of Dolly in their style

sell access to that if they want don't

if they don't want just use it for

themselves uh or get some sort of

economic benefit or otherwise when

someone does use their stuff um and it's

not just training on their images it

really is like you know it really is

about style uh and and that's that's the

thing that at least in the artist

conversations I've had that people are

super interested in so for now it's like

all right let's know what people don't

want make sure that we respect that um

of course you can't make everybody happy

but try to like make the community feel

like we're being a good partner um but

what what I what I think will be better

and more exciting is when we can do

things that artists are like that's

awesome Anna you are open AI ambassador

to Washington other capitals around the

world I I'm curious what you've taken

from your experience in Facebook what

you've taken from the tense relations

between a lot of tech companies and

governments and Regulators over the past

few decades and how you're putting that

to use now in open open AI I mean so I

think one thing that I really learned

working in government and of course I

worked in the White House during the

2016 Russia election interference and

people think that that was the first

time we'd ever heard of it but it was

something that we had actually been

working on for years and thinking you

know we know that this happens what do

we do about it and one thing I never did

during that period is go out and talk to

the companies because it's not actually

typical thing you do in government and

was much more rare back then especially

with you know these emerging tools and I

thought about that a lot as I entered

the tech space that I regretted that and

that I wanted governments to be able to

really understand the technology and how

the decisions are made by these

companies and also just honestly when I

first joined openi no one of course had

heard of openi in government for the

most part

and I thought every time I used it I

thought my God if IID had this for8

years I was in the administration I

could have gotten 10 times more done so

for me it was really how do I get my

colleagues to use it um especially with

open eyes mission to make sure these

tools benefit everyone I don't think

that'll ever happen unless governments

are incorporating it to serve citizens

more efficiently and faster and so this

is actually one of the things I've been

most excited about is to just really get

governments to use it for everyone's

benefit I mean I'm hearing like a lot of

sincerity in that pitch are Regulators

receptive to it it feels like a lot are

coming to the conversation probably with

a good deal of skepticism because of

past interactions with Silicon Valley I

think I mostly don't even really get to

talk about it because for the most part

people are interested in governance and

Regulation and I think that they know um

theoretically that there is a lot of

benefit the government many governments

are not quite ready to incorporate I

mean there are exceptions obviously

people who are really at the Forefront

so it's not you know I think often I

just don't even really get to that

conversation

so I want to ask you both about the

dramatic turn of events in uh November

Sam one day the window on these

questions will close um that is not you

think they

will I think at some point they probably

will but it hasn't happened yet so it

doesn't doesn't matter um I guess my

question is is you know have you

addressed the Govern the governance

issues the very unique uh corporate

structure at open AI with the nonprofit

board and the cap profit arm that led to

your ouer we're going to focus first on

putting a great full board in place um I

expect us to make a lot of progress on

that in the coming months uh and then

after that the new board uh will take a

look at the governance structure but I

think we debated both what does that

mean is it should open AI be a

traditional Silicon Valley for-profit

company we'll never be a traditional

company but the structure I I think we

should take a look at the structure

maybe the answer we have now is right

but I think we should be willing to

consider other things but I think this

is not the time for it and the focus on

the board first and then we'll go look

at it from all angles I mean presumably

you have investors including Microsoft

including uh your Venture Capital

supporters um your employees who uh over

the long term are seeking a return on

their investment um I think one of the

things that's difficult to express about

open aai is the degree to which our team

and the people around us investors

Microsoft whatever are committed to this

Mission um in the middle of that crazy

few days uh at one point I think like 97

something like that 98% of the company

signed uh a letter saying you know we're

all going to resign and go to something

else and that would have torched

everyone's equity and for a lot of our

employees like this is all or the great

majority of their wealth and people

being willing to go do that I think is

quite unusual our investors who also

were about to like watch their Stakes go

to zero which just like how can we

support you and whatever is best for for

the mission Microsoft too um I feel very

very fortunate about that uh of course

also would like to make all of our

shareholders a bunch of money but it was

very clear to me what people's

priorities were and uh that meant a lot

I I I sort of smiled because you came to

the Bloomberg Tech Conference in last

June and Emily Chang asked uh it was

something along along the lines of why

should we trust you and you very

candidly says you shouldn't and you said

the board should be able to fire me if

if they want and of course then they did

and you quite uh adeptly orchestrated

your return actually let me tell you

something um I the board did that I was

like I think this is wild super confused

super caught off guard but this is the

structure and I immediately just went to

go thinking about what I was going to do

next it was not until some board members

called me the next morning that I even

thought about really coming back um when

they asked you don't want you want to

come back uh you want to talk about that

but like the board did have all of the

Power there now you know what I'm not

going to say that next thing but I I I

think you should continue I think I no I

would I would also just say that I think

that there's a lot of narratives out

there it's like oh well this was

orchestrated by all these other forces

it's not accurate I mean it was the

employees of open AI that wanted this

and that thought that it was the right

thing for Sam to be back the you know

like yeah I thing I'll will say is uh I

think it's important that I have an

entity that like can fire this but that

entity has got to have some

accountability too and that is a clear

issue with what happened right Anna you

wrote a remarkable letter to employees

during The Saga and one of the many

reasons I was excited to to have you on

stage today was ju to just ask you what

were those five days like for you and

why did you step up and write that uh

Anna can clearly answer this if she

wants to but like is really what you

want to spend our time on like the soap

opera rather than like what AI is going

to do I mean I'm wrapping it up but but

um I mean go I think people are

interested okay well we can leave it

here if you want no no yeah let's let's

answer that question and we'll we'll we

can move on I would just say uh for

color that it happened the day before

the entire company was supposed to take

a week off so we were all on Friday uh

preparing to you know have a restful

week after an insane year so then you

know many of us slept on the floor of

the office for a week right there's a

question here that I think is a a really

good one we are at Davos climate change

is on the agenda um the question is does

do well I'm going to give it a different

spin considering the compute costs and

the the need for chips does the

development of AI and the path to AGI

threaten to take us in the opposite

direction on the climate

um we do need way more energy in the

world than I think we thought we needed

before my my whole model of the world is

that the two important currencies of the

future are compute SL intelligence and

energy um you know the ideas that we

want and the ability to make stuff

happen and uh the ability to like run

the compute and I think we still don't

appreciate the energy needs of this

technology um the good news to the

degree there's good news is there's no

way to get there without a breakthrough

we need Fusion or we need like radically

cheaper solar Plus Storage or something

at massive scale like a scale that no

one is really planning for um so

we it's totally fair to say that AI is

going to need a lot of energy but it

will force us I think to invest more in

the technologies that can deliver this

none of which are the ones that are

burning the carbon like that'll be those

all those unbelievable number of fuel

trucks and by the way you back one or

more nuclear yeah I I personally think

that

is either the most likely or the second

most likely approach feel like the world

is more receptive to that technology now

certainly historically not in the US um

I think the world is still

unfortunately pretty negative on fishing

super positive on Fusion it's a much

easier story um but I wish the world

would Embrace fishing much more I look I

I may be too optimistic about this but I

think

I I think we have paths now to

massive a massive energy transition away

from burning carbon it'll take a while

those cars are going to keep driving

there you know there's all the transport

stuff it'll be a while till there's like

a fusion reactor in every cargo ship um

but if if we can drop the cost of energy

as dramatically as I hope we can then

the math on carbon captur just so

changes uh I still expect unfortunately

the world is on a path where we're going

to have to do something dramatic with

climate look like geoengineering as a as

a as a Band-Aid as a stop Gap but I

think we do now see a path to the

long-term solution so I I want to just

go back to my question in terms of

moving in the opposite direction it

sounds like the answer is potentially

yes on the demand side unless we take

drastic action on the supply side but

there there is no I I see no way to

supply this with to to manage the supply

side without

a really big breakthrough right which is

this is does this frighten you guys

because um you know the world hasn't

been that versatile when it comes to

supply but AI as you know you have

pointed out is not going to take its

time until we start generating enough

power it motivates us to go invest more

in fusion and invest more in nor new

storage and and not only the technology

but what it's going to take to deliver

this at the scale that AI needs and that

the whole globe needs so I think it

would be not helpful for us to just sit

there and be nervous um we're just like

hey we see what's coming with very high

conviction it's coming how can we use

our

abilities uh our Capital our whatever

else to do this and in the process of

that hopefully deliver a solution for

the rest of the world not just AI

training workloads or inference

workloads Anna it felt like in 2023 we

had the beginning of a almost

hypothetical conversation about

regulating AI what what should we expect

in 2024 and you know does it do do do

governments act does it does it become

real and what is what is AI safety look

like so I think we it is becoming real

you know the EU is uh on the cusp of

actually finalizing this regulation

which is going to be quite extensive and

the Biden Administration uh wrote the

longest executive order I think in the

history of executive orders uh covering

this technology and is being implemented

in 2024 because they gave agencies you

know a bunch of homework for how to

implement this and govern this

technology and and it's happening so I

think it is really moving forward um but

what exactly safety looks like of what

it even is I think this is still a

conversation we haven't bottomed out on

you know we founded this Frontier Model

Forum in part yeah maybe explain what

that is so this is um for now this is um

Microsoft openai anthropic and um Google

but it will I think expand to other

Frontier Labs but really right now all

of us are working on safety we all red

teamr models um we all do a lot of this

work but we really don't have even a

common vocabulary um or a standardized

approach and to the extent that people

think like well this is just industry

but uh this is in part in response to

many governments that have asked us for

this very thing so like what is it

across industry that you think are

viable best practices is there a risk

that regulation starts to discourage

entrepreneurial activity in in AI I mean

I think people are terrified of this um

this is why I think Germany and France

and Italy in interjected into the EU um

AI act discussion because they are

really concerned about their own

domestic Industries being sort of

undercut before they've even had a

chance to develop were you satisfied

with your old boss's executive order and

was was there anything in there that uh

you had lobbied against no and in fact

you know I think it's it was really good

in that it wasn't just these are the

restrictions it's like and then also

please go and think about how your

agency will actually leverage this to do

your work better so I was really

encouraged that they actually did have a

balanced

approach um Sam first time at Davos

first time okay is um uh you mentioned

that uh You' prefer to spend more of our

time here on stage talking about AGI

what is the message you're bringing to

political leaders and other Business

Leaders here if you could distill it

thank you

um

so I think 2023 was a year where the

world woke up to the possibility of

these systems becoming increasingly

capable and increasingly General but GPT

4 I think is best understood as a

preview and it was more Over the Bar

than we expected of utility for more

people in more ways but you know it's

easy to point out the limitations and

again we're thrilled that people love it

and use it as much as they do but this

is progress here is not linear and this

is the thing that I think is really

tricky humans have horrible intuition

for exponentials at least speaking for

myself but it seems like a common part

of the human condition um what does it

mean if GPT 5 is as much better than gp4

is four was to three and six is to five

and what does it mean if we're just on

this trajectory now um what you know on

the question of Regulation I think it's

great that different countries are going

to try different things some countries

will probably ban AI some countries will

probably say no guard rails at all both

of those I think will turn out to be

suboptimal and we'll we'll get to see

different things work but as these

systems become more powerful um as they

as they become more deeply integrated

into the economy as they become

something we all used to do our work and

then as things beyond that happen as

they become capable of discovering new

scientific knowledge for

Humanity even as they become capable of

doing AI research at some point um the

world is going

to change more slowly and then more

quickly than than we might imagine but

the world is going to change um this is

you know a thing I I always say to

people is no one knows what happens next

and I really believe that and I think

keeping the humility about that is

really important you can see a few steps

in front of you but not too many

um but when cognition the when the cost

of cognition Falls by a factor of a

thousand or a million when the

capability of it becomes uh it augments

Us in ways we can't even imagine you

know uh like one example I I try to give

to people is what if everybody in the

world had a really competent company of

10,000 great virtual employees experts

in every area they never fought with

each other they didn't need to rest they

got really smart they got smarter at

this rapid Pace what would we be able to

create for each other what would that do

to the world that we experience and the

answer is none of us know of course and

none of us have strong intuitions for

that I can imagine it sort of but it's

not like a clear picture um and this is

going to happen uh it doesn't mean we

don't get to steer it it doesn't mean we

don't get to work really hard to make it

safe and to do it in a responsible way

but we are going to go to the Future and

I think the best way to get there in a

way that works

is the level of Engagement we now have

part of the reason a big part of the

reason we believe in iterative

deployment of our technology is that

people need time to gradually get used

to it to understand it we need time to

make mistakes while the stakes are low

governments need time to make some

policy mistakes and also technology and

Society have to co-evolve in a case like

this uh so technology is going to change

with each iteration but so is the way

Society works and that's got to be this

interactive iterative process um and we

need to embrace it but have caution

without fear and how long do we have for

this iterative process to play I I think

it's surprisingly continuous I don't

like if I try to think about

discontinuities I can sort of see one

when AI can do really good AI research

um and I can see a few others too but

that's like an evocative example um but

on the whole I don't think it's about

like Crossing this one line I think it's

about this continuous exponential curve

we climb together and so how long do we

have like no time at all in

infinite I saw GPT five trending on X

earlier this week and I clicked and I

you know couldn't I it sounded uh you

know probably misinformed but what what

can you tell us about gbt 5 and is it an

exponential uh you know improvement over

what we've seen look I don't know what

we're going to call our next model um I

don't know when are you going to get

creative with the uh the naming process

uh I don't want to be like shipping

iPhone

27 um so you know it's not my style

quite uh but I I think the next model we

release uh I expect it to be very

impressive to do new things that were

not possible with gp4 to do a lot of

things better and I expect us to like

take our time and make sure we can

launch something that we feel good about

and responsible about within open AI

some employees consider themselves to be

quote building God is that I haven't

heard that okay is um I mean I've heard

like people say that factiously but uh I

think almost all employees would say

they're building a tool more so than

they thought they were going to be which

they're thrilled about you know this

confusion in the industry of Are We

building a creature are we building a

tool um I think we're much more building

a tool and that's much

better uh to transition to something

yeah goad no no no no you finish your

thought oh I was just going to say like

the

the we think of ourselves as tool

Builders um AI is much more of a tool

than a product and much much more of a

tool than this like entity and uh one of

the most wonderful things about last

year was seeing just how much people

around the world could do with that tool

and they astonished us and I think we'll

just see more and more and human

creativity uh and ability to like do

more with better tools is remarkable and

and before we have to start wrapping up

you know there was a report that you

were working with Johnny I on an AI

powered device either within open AI

perhaps as a separate company you know I

bring it up because CES was earlier this

month and AI powered devices were the

the talk of of the conference you know

can you give us an update on that and

are we approach does AI bring us to the

beginning of the end of the smartphone

era smartphones are fantastic I don't

think smartphones are going anywhere uh

I think what they do they do really

really well and they're very general if

if there is a new thing to make uh I

don't think it replaces a smartphone in

the way that I don't think smartphones

replace computers but if there's a new

thing to make that helps us do more

better you know in a in a new way given

that we have this unbelievable change

like I don't think we quite I don't

spend enough time I think like marveling

at the fact that we can now talk to

computers and they understand us and do

stuff for us like it is a new affordance

a new way to use a computer and if we

can do something great there uh a new

kind of computer we should do that and

if it turns out that the smartphone's

really good and this is all software

then fine but I bet there is something

great to be done and um the partnership

with Johnny is that an open AI effort is

that another company I have not heard

anything official about a partnership

with

Johnny okay um Anna I'm going to give

you the last word as you and Sam meet

with business and world leaders here at

Davos what's the message you want to

leave them

with um I think the that there is an a

trend where people feel more fear than

excitement about this technology and I

understand that we have to work very

hard to make sure that the best version

of this technology is realized but I do

think that many people are engaging with

this via the leaders here and that they

really have a responsibility to make

sure that um they are sending a balanced

message so that um people can really

actually engage with it and realize the

benefit of this technology can I have 20

seconds absolutely one one of the things

that I think open ey has not always done

right in the field hasn't either is find

a way to build these tools in a way uh

and also talk about them that don't

don't get that kind of response I think

chat gbt one of the best things it did

is it shifted the conversation to the

positive not because we said trust us

it'll be great but because people used

it and are like oh I get this I use this

in a very natural way the smartphone was

cool cuz I didn't even have to use a

keyboard and phone I could use it more

naturally talking is even more natural

um speaking of Johnny Johnny is a genius

and one of the things that I think he

has done again and again about computers

is figuring out a way to make them very

human compatible and I think that's

super important with this technology

making this feel like uh you know not

this mystical thing from sci-fi not this

scary thing from sci-fi but this this

new way to use a computer that you love

and that really feels like I still

remember the first iMac I got and what

that felt like to me

relative it was heavy but the fact that

it had that handle even though it is

like a kid it was very heavy to carry um

it did mean that I was like I had a

different relationship with it because

of that handle and because of the way it

looked I was like oh I can move this

thing around I could unplug it and throw

it out the window if it tried to like

wake up and take over that's nice um and

I think the way we design our technology

and our products really does matter