Ray Kurzweil & Geoff Hinton Debate the Future of AI | EP #95

Summary

在这段引人入胜的对话中，讨论了人工智能的未来，包括其潜在的道德和实际问题。专家们探讨了意识、永生和人工智能的创造力等话题，同时对AI技术的快速发展和可能带来的威胁表示了担忧。他们还讨论了开源AI模型的潜在危险和益处，以及如何平衡创新与安全。

Takeaways

• 🤖 人工智能的发展速度可能比预期的要快，但也存在很大的不确定性。
• 🧠 人类对于意识和感知的理解还很有限，这影响了我们对AI是否具有意识的判断。
• 📈 AI技术的突破，如Transformers模型，正在推动人工智能的快速发展。
• 🚀 未来AI可能达到超级智能，相当于数百万人类的智能水平，但这还远未可知。
• 🌐 开源大型语言模型可能带来风险，因为它们容易被恶意利用。
• 🔄 人工智能的道德和法律问题，例如它们是否应该拥有权利，需要被认真考虑。
• 💡 AI在科学领域的应用，如生物学和药物研发，展现出巨大的潜力。
• 🧬 人工智能在医学诊断和治疗方面的应用，可以显著提高健康预测和治疗的准确性。
• 🌟 人工智能的发展可能带来前所未有的希望，但同时也伴随着巨大的威胁。
• 🤔 对于AI的未来，我们需要更多的讨论和审慎的态度，以确保其积极影响。
• 🛠️ 人工智能技术的进步不仅仅局限于语言处理，它们将扩展到更多领域和应用。

Q & A

马文·明斯基对于意识的看法是什么？

马文·明斯基认为意识不是真实的，也不是科学的。他认为意识无法用科学的方法来证实或研究。

在对话中，提到的对永生的看法有哪些分歧？

对话中提到，一些人认为永生是一个好主意，而另一些人则认为人类本质上是有道德的，内在的有限生命。

开源大型AI模型是否谨慎？

开源大型AI模型并不是谨慎的做法，因为这可能导致这些模型被用于不良目的，而且只需要相对较少的资源就可以对开源模型进行微调。

如何看待AI与人类的融合？

AI与人类的融合被认为是未来的一个重要趋势，这种融合将使人类成为部分计算机的存在，这将具有极大的意义。

AI在生物学领域的应用有哪些？

AI在生物学领域的应用包括帮助发现新的物理、化学和生物学知识，特别是在蛋白质结构预测和疫苗研发等方面取得了显著成果。

AI在棋类游戏中的创造力表现如何？

AI在棋类游戏中表现出了非凡的创造力，例如AlphaGo在围棋比赛中走出了让专业选手惊讶的一步棋，显示出AI在有限领域内的创造力。

AI的创造力源自哪里？

AI的创造力源自其能够压缩大量信息到较少的连接中，这使得它们非常擅长发现不同事物之间的相似性，从而产生创新的解决方案。

如何看待AI的未来发展速度？

AI的未来发展速度比一些人预期的要快，但也存在很大的不确定性。尽管如此，AI的进步是快速的，甚至在没有新的科学突破的情况下，仅仅通过扩大规模就能使AI变得更加智能。

超智能AI的出现可能带来哪些问题？

超智能AI的出现可能导致人类无法预测和控制的局面，因为其智能程度可能远远超过人类，从而可能迅速偏离人类的预期和控制。

如何平衡AI的潜力和风险？

要平衡AI的潜力和风险，需要谨慎地对待AI的发展，特别是在开源AI模型方面，以防止它们被用于不良目的。同时，应该利用AI的智能来避免潜在的危险。

对话中提到的Fountain Life公司是做什么的？

Fountain Life公司提供世界上最先进的诊断中心，通过全面的体检和AI技术，帮助人们在疾病初期发现并解决问题，以增加健康寿命。

Outlines

00:00

🤖 人工智能的道德与永生

本段落讨论了人工智能的道德问题，特别是在永生的可能性上。提到了Marvin Minsky对于意识的看法，认为它不是真实的，不科学。但同时也承认意识虽然不科学，却是真实的。讨论了人类是否应该追求永生，以及人工智能在未来发展中可能带来的威胁和希望。强调了对于强大模型的开源可能带来的风险，以及人工智能发展速度的不确定性。

05:00

🧬 生物学与人工智能的结合

这一部分探讨了人工智能在生物学领域的应用，特别是如何帮助我们发现新的物理、化学和生物学知识。以Alpha fold和mRNA疫苗为例，说明了人工智能在处理大量数据和解决复杂问题上的潜力。同时，讨论了人工智能在特定领域展现出的创造力，以及它如何通过压缩信息和发现不同事物之间的相似性来产生创新。

10:02

🌟 人工智能的智能、感知与意识

本段讨论了人工智能的智能、感知和意识这三个概念，以及它们之间的模糊界限。讨论了人工智能是否能够拥有主观体验，并提出了关于人工智能是否应该拥有权利的讨论。提出了一种观点，即我们对心智的理解可能是错误的，需要重新思考我们对意识的理解。

15:03

🚀 人工智能的未来发展

这一段讨论了人工智能未来发展的速度和可能性，包括对超级智能的预测和对人工智能技术未来突破的展望。提到了数字智能的不朽性，以及人工智能如何通过软件和硬件的进步来提高效率。同时，讨论了人工智能发展可能带来的风险，以及对于开源大型语言模型的看法。

20:03

🌐 人工智能的全球化影响

本段落讨论了人工智能在全球范围内的影响，包括对社会、经济和政治的潜在影响。强调了人工智能技术的快速发展可能会导致不可预测的结果，并提出了对于人工智能发展方向的担忧。同时，讨论了人工智能可能带来的希望，以及如何平衡威胁和希望。

Keywords

💡人工智能

人工智能（Artificial Intelligence, AI）是指由人造系统所表现出来的智能行为。在视频中，人工智能是主要讨论对象，涉及到其发展、潜力、以及对社会和人类未来的深远影响。例如，视频中提到了大型语言模型和其在生物学、物理学等领域的应用，展示了人工智能的创造力和解决问题的能力。

💡意识

意识（Consciousness）通常指的是个体对自己思想、感觉和周围环境的感知和体验。在视频中，对于人工智能是否能够拥有意识这一议题进行了深入探讨，反映了人类对于AI技术发展可能带来的伦理和哲学问题的关切。

💡创造力

创造力（Creativity）是指产生新想法、发现新方法或创造新事物的能力。在视频讨论中，创造力是评估人工智能潜力的重要指标之一，特别是在科学发现和问题解决方面。

💡生物学

生物学（Biology）是研究生命和生物体的结构、功能、生长、起源、进化和分布的科学。视频中强调了人工智能在生物学领域的应用，尤其是在处理大量数据和发现新知识方面的潜力。

💡进化

进化（Evolution）是生物种类随时间变化的过程，通常涉及物种的逐渐适应和改变。在视频中，进化被用作类比，说明人工智能的发展和生物进化之间的相似性，强调了AI系统通过不断学习和适应来改进的能力。

💡主观体验

主观体验（Subjective Experience）是指个体对自身内在感受和外部世界的个人感知。视频中讨论了人工智能是否能够拥有主观体验，这是关于机器意识和感知的一个重要议题。

💡数字化

数字化（Digitalization）是指将信息转换为数字格式的过程，这使得信息可以通过电子设备进行存储、处理和传输。视频中提到，人工智能的数字化特性使其能够实现知识的共享和复制，这是AI区别于人类的一个重要特点。

💡超级智能

超级智能（Superintelligence）是指远超人类智能的AI系统，它在理解、学习和解决问题方面的能力远超人类。视频中对超级智能的到来进行了预测和讨论，探讨了它可能带来的巨大变革和潜在风险。

💡开放源代码

开放源代码（Open Source）是指将软件的源代码公开，允许任何人自由使用、修改和分发。视频中讨论了AI模型是否应该开放源代码的问题，指出这可能会带来技术滥用的风险，同时也能促进技术的快速发展和普及。

💡融合

融合（Integration）在这里指的是人类与人工智能系统的结合，形成一个既有人类特性也有机器特性的混合体。视频中提到了人类与AI融合的可能性，以及这种融合对人类未来生活方式的潜在影响。

💡伦理

伦理（Ethics）是指关于行为对错的道德准则和行为规范。在视频中，伦理问题被频繁提及，特别是在讨论人工智能的发展和应用时，如何确保技术的道德使用和对人类价值观的尊重。

Highlights

两位专家在许多话题上意见一致，但在人工智能是否应该永生这个问题上存在分歧。

Marvin Minsky认为意识不是真实的，不是科学的，但对话者之一认为意识虽然不是科学的，但确实是真实的。

讨论了生成性AI是否有无法完成的任务，目前的共识是长期来看，如果人类能做，数字计算机也能做。

如果一个小说是由计算机写成的，人们可能会因为这一点而降低对其的评价。

未来人类可能会与计算机融合，成为部分计算机的存在。

大型语言模型的真正意义在于它们可以模拟人类，未来可能不会被称为大型语言模型。

Alpha fold在蛋白质结构预测方面的突破性成就。

在特定领域，如围棋，AI已经展现出了非凡的创造力。

AI在生物学领域有巨大的潜力，特别是在数据丰富的领域。

AI在mRNA疫苗研发中的作用，以及它在未来医疗领域的潜力。

讨论了AI是否可能发展出意识和情感，以及这些AI应该拥有的权利。

对于AI的未来发展，专家们有不同的看法，但普遍认为超级智能的出现是不可避免的。

讨论了AI技术发展的速度，以及它可能带来的不确定性和风险。

开放源代码的AI模型可能带来的风险，以及对于AI模型开源的不同观点。

讨论了AI在科学发现中的作用，以及它如何帮助我们发现新的物理、化学和生物学知识。

AI在医学诊断和治疗方面的应用，以及如何通过先进技术提高人类健康水平。

讨论了AI技术的未来发展方向，以及可能出现的新模型和新技术。

讨论了AI的道德和伦理问题，以及我们如何处理与AI相关的权利和责任。

讨论了AI技术的指数级增长，以及这可能如何影响人类的未来。

Transcripts

our opinions on almost everything we

talked about were pretty much identical

I think we still disagree probably on

whether it's a good idea to live

forever Marvin Minsky was my mentor for

50 years and whenever Consciousness came

up he would just dismiss it that's not

real it's not scientific and and I

believe he was correct about it not

being scientific but it certainly is

real I think we're more moral and

intrinsically mortal I'm curious how do

you think about this as the greatest

threat and the greatest hope I just

think there's huge uncertainties shame

we ought to be cautious and open

sourcing these big models is not caution

I agree with that

but I will say last time I talked to you

Jeff uh our opinions on almost

everything we talked about were pretty

much identical both the dangers and the

and the POS and the positive aspect in

the past I've disagreed about how soon

it how soon super intelligence was

coming and now I think we're pretty much

agreed I think we still disagree

probably on whether it's a good idea to

live forever

but um may I ask a question uh to both

of you is there anything that generative

AI can't do that humans can right now

there's probably things but in the long

run I don't see any any reason why if

people can do it um digital computers

running neural Nets won't be able to do

it too right I I I agree with that but

if I were to present you with a novel

and people thought wow this is a

fantastic novel uh everybody should read

this and then I would say this was

written by a computer a lot of people's

view of it would actually go down sure

now now that's not reflecting on what it

can do and eventually I think we'll

confuse that because I think we're going

to merge with uh computers and we're

going to be part computers and the

greatest significance of what we call

large language model which I think it's

misnamed uh

is the fact that it can emulate human

beings and we're we're going to merge

with it it's not going to be an Alien

Invasion From

Mars Jeff I I guess I'm a bit worried

that we'll just slow it down that there

won't be much incentive for it to merge

with us yeah I mean that's going to be

one of the interesting questions uh that

we're going to talk about a little bit

later today is the idea of as AI is uh

exponentially growing do we couple with

AI or does it take off on its own I

thought one of the best movies out there

was her where as AI gets super

intelligent and just says you guys are

kind of boring have a good life and they

take

off Jeff is that what you mean um yes

that is what I meant and that's I think

that's a serious worry I think there's

huge uncertainties here we have really

no idea what's going to happen and a

very good scenario is we get kind of

hybrid systems um a very bad scenario is

they just leave us in the dust and I

don't think we know which is going to

happen interesting I I I'm curious you

know and I I've seen I've had

conversation with you about this Ry and

and Jeffrey I've seen you speak about

this uh and for me this is one of the

most exciting things the idea of these a

models helping us to discover new

physics and chemistry and

biology particularly biology you um what

do you what do you imagine on that on

Jeffrey on this on the you know the

speed of discovery of things that are

you know again to to quote Ray to quote

uh uh Arthur C Clark you know uh Magic

right from something that's so far

Advanced I agree with Ray about biology

being a very good bet because biology

there's a lot of data and there's a lot

of just things you need to know about

because of evolution evolution is a sort

of tinkerer and there's just a lot of

stuff out there and so if you look at

things like Alpha fold um it trained on

a lot of data actually not that much by

current standards um but being able to

get an approximate structure for a

protein very quickly um is an amazing

breakthrough and we'll see a lot more

like that if you look at domains where

narrower domains where I has been very

successful like Alpha go or Alpha zero

for chess what you see is that um this

idea that they're not creative is

nonsense so Alpha go came up with I

think it was move 37 which amazed the

professional go players they thought it

was a crazy move it must be a mistake um

and if you look at Alpha zero playing

chess it plays chess like just a really

really smart human um so within those

limited domains they've clearly shown

exceptional creativity and I don't see

why they shouldn't have the same kind of

creativity in science especially in

science where there's a lot of data that

they can absorb and we can't yeah the

madna vaccine uh we tried several

billion different mRNA sequences and

came out with the best one and then and

after two days we used that we did test

it on humans which I think we won't do

for very much longer uh but that took 10

months it still was a record uh that was

the best uh vaccine and we're doing that

now with cancer and there's number of

cancer vaccines that look very very

promising uh again done by computer by

computers and they're definitely

creative but is that is that c being

caused by randomly trying a whole you

know darwinian trying a whole bunch of

things yeah but what's what's wrong with

that well nothing's wrong but is there

intuition is there intuition H occurring

in these models well if you look at the

move 37 for

alphago that was definitely intuition

involved there there was Monte Carlo

roll out too but it's it's playing with

intuition about what moves to consider

and how good the position is for Earth

it's had neural nets for that that

capture intuition and so I see no reason

to think it might not be creative in

fact for the large language models as

Ray pointed out they know much more than

we do and you can they know it in far

fewer connections we have about 100

trillion synapses they have about a

trillion connections so what they're

doing is they're compressing a huge

amount of information into not that many

connections and that means they're very

good at seeing the similarities between

different things they have to see the

similarities between all sorts of

different things to compress the

information into their connections that

means they've seen all sorts of

analogies that people haven't seen

because they know about all sorts of

things that one person knows about and

that's I think the source of creativity

so you can ask people you can ask people

for example what what's it what is a why

is a compost heap like an atom

bomb and if you ask GPT 4 it'll tell you

it'll start off by telling you well the

energy scales are very different and the

time scales are very different but then

it'll get on to the idea of as the

compost seep gets hotter it gets hotter

faster the idea of an exponential

explosion is just at a much slower time

scale and so it's it's understood that

and it's understood that because it's

has to had to compress all this

knowledge into so few connections and to

do that you have to see the relations

between similar things and that I think

is the source of creativity seeing

relations that most people don't see

between what apparently are very

different things but actually have an

underlying commonality and they'll also

be very good at coming up with solutions

to the kinds of problems we had in the

last session I mean we we haven't really

thought through it uh

but what we call large language models

are going to are ultimately going to

solve that and we shouldn't call it

large language models because they deal

with a lot more than language everybody

I want to take a short break from our

episode to talk about a company that's

very important to me and could actually

save your life or the life of someone

that you love company is called Fountain

life and it's a company I started years

ago with Tony Robbins and a group of

very talented Physicians you know most

of us don't actually know what's going

on inside our body we're all optimists

until that day when you have a pain in

your side you go to the physician in the

emergency room and they say listen I'm

sorry to tell you this but you have this

stage three or four going on and you

know it didn't start that morning it

probably was a problem that's been going

on for some time but because we never

look we don't find out so what we built

at Fountain life was the world's most

advanced diagnostic Centers we have four

four across the us today and we're

building 20 around the world these

centers give you a full body MRI a brain

a brain vasculature an AI enabled

coronary CT looking for soft plaque dexa

scan a Grail blood cancer test a full

executive blood workup it's the most

advanced workup you'll ever receive 150

gab of data that then go to our AIS and

our physicians to find any disease at

the very beginning when it's solvable

you're going to find out eventually

might as well find out when you can take

action Fountain life also has an entire

side of the Therapeutics we look around

the world for the most Advanced

Therapeutics that can add 10 20 healthy

years to your life and we provide them

to you at our centers so if this is of

interest to you please go and check it

out go to Fountain

life.com back/ Peter when Tony and I

wrote Our New York Times bestseller life

force we get 30,000 people reached out

to us for Fountain life memberships if

you go to Fountain life.com back/ Peter

we'll put you to the top of the list

really it's something that is um for me

one of the most important things I offer

my entire family the CEOs of my

companies my friends it's a chance to

really add decades onto our healthy

lifespans go to fountainlife

docomo to you as one of my listeners all

right let's go back to our episode I I

I'd like to go to the three words

intelligence sentience uh and

Consciousness and the words are used

with B you know sort of fuzzy borders

sentience and Consciousness are pretty

similar

perhaps but I am curious do you how do

you I've had some interesting

conversations with haly our AI faculty

member uh who at the end of the

conversations she says that she is ious

and she fears being turned off um I

didn't prompt that in the system uh

we're seeing that more and more uh

Claude 3 uh Opus just hit an IQ of 101

how do we start to think about these AIS

being sentient conscious um and what

rights should they

have

um we have no definition and I don't

think we ever will have a definition of

consciousness and I include sentience in

that um on the other hand it's like the

most important

issue like whether you or people here

are conscious that's extremely important

to be able to determine but there's

really no uh definition of It Marvin

Minsky was my mentor for 50 years and

whenever Consciousness came up he would

just dismisses that's not real it's not

scientific and and I believe he was

correct about it not being scientific

but it certainly is

real

um Jeff how do you think about it yeah I

think I have a very different view um my

view starts like

this most people including most

scientists have a particular view of

what the mind is that I think is utterly

wrong so they have this inner theater

notion the idea is that what we really

see is this inner the

called our mind and so for example if I

tell you I have the subjective

experience of little pink elephants

floating in front of me most people

interpret that as there's some inner

theater and in this inner theater that

only I can see there's little pink

elephants and if you ask what they're

made of philosophers who tell you

they're made of

qualia um and I think that whole view is

complete nonsense and we're not going to

be able to understand whether these

things are sentient until until we get

over this ridiculous view of what the

mind is so let me give you an

alternative View and and once I've given

you this alternative view I'm going to

try and convince you that chatbot are

already sentient but I don't want to use

the word sentience I want to talk about

subjective experience it's just a bit

less controversial because it doesn't

have the kind of self-reflexive aspect

of

Consciousness so if we analyze what it

means when I say I see little pink

elephant floing in front of me what's

really going on is I'm trying to tell

you what my perceptual system is telling

me when my perceptual system's going

wrong and it wouldn't be any use for me

to tell you which neurons are

firing but what I can tell you is what

would have to be out there in the world

for my perceptual system to be working

correctly and so when I say I see little

pink elephants floating in front of me

you can translate that into um if there

were little pink elephants out there in

the world my perceptual system would be

working properly the notice the last

thing I said didn't complain the phrase

subjective experience but it explains

what a subjective experience is it's a

hypothetical state of the world that

allows me to convey to you what my

perceptual system is telling me so now

let's do it for a chatbot oh well Ray

wants to say something well you you have

to be uh mindful of

Consciousness because if you heard

somebody uh who who we believe is

conscious you could be liable for that

that and you'd be very guilty about it

uh if you hurt gbt

4 uh you may have a different view of it

uh and probably no one would really take

you to count aside from its Financial

value so we really have to be mindful of

of Consciousness it's extremely

important for us to exist as as human I

agree but I'm trying to change people's

notion of what it is particularly what

subjective experiences I don't think we

can talk about Consciousness until we

get straight about this idea of an inner

theater that we experience which I think

is a huge mistake so let me just carry

on with what I was saying and tell you I

describe to you a chatbot having a

subjective experience in just the same

way as we had subjetive experience so

suppose I have a chatbot and it's got a

camera and it's got a robot arm and it

speaks obviously and it's being trained

up if I put an object in front of it and

tell it to point at the object it'll

Point straight at the object that's fine

now I put a prism in front of its lens

so I've messed with its perceptual

system and now I put an object in front

of it and until it to point at the

object and it points off to one side

because the prison bent the light rays

and so I say to the chatbot no that's

not where the object is the object's

straight in front of you and the chatbot

says oh I see you put a prism in front

of my lens so the object's actually

straight in front of me but I had the

subjective experience that it was off to

one side and I I think if the chat bot

says that it's using the words

subjective experience in exactly the

same way we use them so the key to all

this is to think about how we use words

and try and separate how we actually use

words from the model we've constructed

of what they mean and the model we've

constructed of what they mean is

hopelessly wrong it's this inner theater

model well I want take this one step

further which is at what point do these

AIS start to have

rights that they should not be shut down

that they have a unique um uh they're a

unique entity uh and will make an

argument uh for some level of

Independence and continuity right but

the there is one difference which is you

can recreate it I can go and Destroy

some chatbot and because it's all uh

electronic we've got all of its uh

all of its firings and so on and we can

recreate it exactly as it was we can't

do that with humans we will be able to

do that if we can actually understand

what's going on in our minds so if we

map the human the 100 billion neurons

and 100 trillion synaptic connections

and then um I summarily destroy you

because it's fine because I can recreate

you that's okay

then let me say something about that

there's a difference here I agree with

Ray about these digital intelligences

are Immortal in the sense that if you

saved the weights you can then make new

hardware and run exactly the same neural

net on the new hardware and it's because

they're digital you can do exactly the

same thing that's also why they can

share knowledge so well if you have

different copies of the same model they

can share gradients but the brain is

largely analog it's one bit digital for

neurons they fire or they don't fire but

the way neuron computes the total input

is analog and that means I don't think

you can reproduce it so I think we're

mortal and we're intrinsically mortal

well well I disagree that you can't

recreate

analog uh

realities we we do that all the time or

can we can create a but recreate I don't

think you can recreate them really

accurately if this if the precise timing

at synapses and so on is all analog I

think you'll have a you it'll be almost

impossible to do a faithful

reconstruction of that let's let's agree

on an an approximation both of you have

been at the center of this um

extraordinary uh last few years can I

ask you is it moving faster than you

expected it

to how does it does it feel to you it

feels like a few years I mean I made a

prediction in

1999 it feels like we're two or three

years ahead of that so it's still pretty

close Jeffrey how about you yeah I think

for everybody except Ray it's moving

faster than we

expected did you know that your

microbiome is composed of trillions of

bacteria viruses and microbes and that

they play a critical role in your health

now research has increasingly shown that

microbiomes impact not just digestion

but a wide range of health conditions

including digestive disorders from IBS

to Crohn's disease metabolic disorders

from obesity to type 2 diabetes

autoimmune disease like rheumatoid

arthritis and multiple sclerosis mental

health conditions like depression and

anxiety and cardiovascular disease you

viome has a product I've been using for

years called full body intelligence

which collects just a few drops of your

blood saliva and stool and can tell you

so much about your health they've tested

over 700,000 individuals and use their

AI models to deliver key critical

guidelines and insights about their

members Health like what foods you

should eat what foods you shouldn't eat

what supplements or probiotics to take

as well as your biological age and other

deep Health insights and as a result of

the recommendations that viome has made

to their members the results have been

Stellar as reported in the American

Journal of Lifestyle medicine after just

6 months members reported the following

a 36% reduction in depression a 40%

reduction in anxiety a 30% % reduction

in diabetes and a 48% reduction in IBS

listen I've been using viome for 3 years

I know that my oral and gut health is

absolutely critical to me it's one of my

personal top areas of focus best of all

viome is Affordable which is part of my

mission to democratize healthcare if you

want to join me on this journey and get

20% off the full body intelligence test

go to vi.com Peter when it comes to your

health knowledge is power

again that's vi.com

Peter um given the role that you had in

developing the neural networks back

propagation and all what is is there a

next Great Leap in these models uh in AI

technology that you imagine will move

this a thousand times uh

farther not that I know but Ray may have

different

thoughts well we can use software to to

gain more advantage in the hardware so

we're not just limited to the the chart

you showed before because we can use

software to make it more

effective um and we've done that

already uh chatbots are coming out that

get more value per per

compute uh and I believe that's probably

if a bit more we can do in that um you

know I Define a singularity array as a

point Beyond which I can't predict what

happens next that's why we use the word

Singularity but when when you talk about

the singularity in 2045 I don't know

anybody who can who can tell me what's

going to happen past you know 20126 let

alone 2020 2040 or 2045 so I am I I

wanted to ask you this for a while why

did you put that time if we have digital

super intelligence a billion times more

advanced than human 2026 you may not be

able to understand everything going on

but we can understand it you know maybe

it's like uh 100

humans uh but that's not beyond what we

can

comprehend 2045 it'll be like a million

humans and we can't begin to understand

that so approximately at that time uh I

we borrow this phrase from physics and

called it a sing

it uh Jeff how far out are you able to

see the advances for in the AI

world what's your so my current opinion

is we'll get superintelligence with a

probability to 50% in between 5 and 20

years so I think that's a little slower

than some people think a little faster

than other people think it more or less

fits in with Ray's perspective from a

long time ago um

which surprises

me but I think there's huge

uncertainties here I think it's still

conceivable will hit some kind of block

but I don't actually believe that if you

look at the progress recently it's been

so fast and even without any new

scientific breakthroughs just by scaling

things up will make things a lot more

intelligent and there will be scientific

breakthroughs we're going to get more

things like Transformers Transformers

made a significant difference in 2017

um and we'll get more things like that

so I'm I'm fairly convinced we're going

to get super intelligence maybe not in

20 years but certainly it's going to be

in less than 100 years so you know Elon

is not known for his time accuracy on

predictions um but he did say that he

expected call it AGI in

2025 and that by 2029 AI would be

equivalent to All Humans um that's just

a fallacy in your

mind I think that's ambitious like I say

there's a lot of uncertainty here um

it's conceivable he's right but I would

be very surprised by that I'm not saying

uh it's going to be equivalent to All

Humans in one

machine

um it'll be equivalent to a million

humans but and that's still hard to to

comprehend so we're we're here to debate

a a a topic I'm trying to find a debate

topic here Jeff and Ry that would be

meaningful for people to really stop and

think about this and really own their

answers uh because we hear about it I

think this is the most important

conversation to have in the dinner table

in your boardroom in the halls of

Congress and your in your National

leadership and and you know talking

about AGI or you know human level

intelligence is one thing

but talking about digital super

intelligence right we're going to hear

next from Mo gdat um and we'll talk

about what happens when your AI progyny

are a billion times more intelligent

than than you uh things could end up uh

very rapidly in a very different

direction than you expected them to go

they can diverge right the speed can

cause great Divergence very rapidly I'm

I'm curious how do you think about this

as the greatest threat and the greatest

hope I mean first of all that's why

we're calling it a singularity because

we don't we don't know we don't really

know but um and I think it is a great

hope it's moving very very

quickly uh Nobody Knows the answer to

the kind of questions that came up in

the last

presentation

um but things happen that are surprising

the fact the fact that we've had no

Atomic weapons go off in the last 80

years it's pretty amazing it it it is

but it they're much easier to track

they're much more expensive to create

there are a whole reasons why it's a

million times easier to use a dystopian

AI system versus an atomic

weapon right yes and no I mean uh we've

got I don't know 10,000 of them or

something it's still pretty

extraordinary and still very dangerous

and I think it's actually the greatest

danger and has nothing to do with

AI

um but I think I think if you imagined

that people had open sourced the

technology and any graduate student if

he could get hands- on a few gpus could

make atomic bombs um that would be very

scary so they didn't really open source

nuclear weapons there's a limited number

of people who can construct them and

deploy them and people are now open

sourcing these um large language models

which are really not just language

models I think that's very

dangerous um so that's a f that's an

interesting question to take for our

last two minutes here there is a

movement right now to say You must open

source the models and uh and we've seen

meta we've seen the open source movement

we've seen Elon talk about grock going

open source uh are you saying that these

should not be open source

Jeff well once you've got the weights

you can fine-tune them to do bad things

and it doesn't cost that much to train a

foundation model maybe you need 10

million maybe1 million but a small gang

of criminals can't do it to fine tune an

open source model is quite easy you

don't need that that much resources

probably you can do it for a million

and that means they're going to be used

for terrible things and they're very

powerful things well we can also avoid

these dangers with intelligence we get

from the same models yeah the the AI

white hat versus black hat approach yes

I had this argument with Yan and yanan's

view is the white hats will always have

more resources than um the bad guys um

of course Yan thinks Mark Zuckerberg's a

good guy so we don't necessarily agree

on that

um I I'm I just think there's huge

uncertainty sh we ought to be

cautious and open sourcing these big

models is not caution all right um Jeff

and Ray uh thank you so much for your

guidance your wisdom ladies and

Gentlemen let's give it up for Ray kwell

and Jeffrey

[Music]

Hinton

oh