How AI Will Shape Society Over The Next 20 Years
Summary
TLDR在这段视频脚本中,讨论了人工智能(AI)的未来及其对世界的影响。讨论涉及AI在医疗、教育、交通和资源管理等领域的潜力,以及AI如何帮助提高人类能力,实现个性化教育,并可能在20年内使道路上25%的车辆实现某种形式的自主性。同时,讨论了AI带来的挑战,包括就业变化、数据隐私和去中心化问题。专家们对AI的长期影响持乐观态度,认为尽管存在风险和误解,AI技术的发展将为社会带来积极的变化,并强调了在AI发展中考虑激励机制和经济学的重要性。
Takeaways
- 🤖 **AI发展速度**:人工智能的发展速度比以前的技术革命快得多,例如CRISPR技术经过多年才获得FDA批准,而AI技术的普及速度则更快。
- 🌐 **AI的普及**:预计在未来20年内,大约20-25%的车辆将具有某种形式的自主性,但不太可能超过这个比例。
- 🚀 **AI在教育中的应用**:AI将使教育更加个性化,允许人类教师快速了解每个学习者的状态,并且可以远程进行,提高教育的可及性。
- 🔍 **AI的风险与治理**:尽管AI带来了许多积极的可能性,但也存在风险,如因AI导致的死亡事件,社会需要找到方法来管理和适应这些风险。
- 🧠 **AI与人类智能**:AI将帮助我们更好地学习,提高我们的认知能力,使人类在某些领域如绘画或歌曲创作方面变得更好。
- 💼 **AI在资源管理中的应用**:AI在资源管理方面,如大数据或分布式资源分配,已被证明是有效的,并将看到更广泛的应用。
- 🌱 **AI在可持续发展目标中的作用**:AI的生产力提升可能会改变某些领域的经济效益,如水资源和农业,吸引更多的顶尖人才。
- 📈 **AI经济学**:AI将改变许多职业的经济效益,使得以前昂贵的服务如法律咨询变得更加便宜和可及。
- 🌐 **去中心化的AI**:AI的未来可能在于去中心化,包括数据、计算和治理的去中心化,以及通过联邦学习和专业化模型实现。
- 🚗 **自动驾驶车辆**:自动驾驶车辆的发展将是一个渐进的过程,预计不会完全自动化,而是会与人类驾驶员共存。
- 🧪 **AI在复杂系统中的应用**:AI在处理如微生物组这样复杂的系统时,可以提供帮助,通过分析大量样本来理解整体功能。
Q & A
Ramesh Raskar是如何看待AI在未来20年内对人类生活的影响的?
-Ramesh Raskar是一个技术乐观主义者,他预计AI将在未来20年内改善我们的健康,特别是在通过植入物提高记忆力方面。他还预测,大约20%到25%的路面车辆将实现某种形式的自主性。他认为AI将帮助我们更好地学习,并改善我们的学习方式。
Hari Balakrishnan最近启动了什么项目,它在AI领域有什么作用?
-Hari Balakrishnan最近启动了CMI,这是世界上最大的远程信息技术服务提供商,旨在帮助在道路上保持安全。
在AI领域,我们如何从以往的技术革命中学习并找到平衡点?
-以往技术革命,如互联网、CRISPR基因编辑技术、核能等,都引发了大规模的变化讨论,包括风险和治理。我们从这些技术中学到的是,最终我们能够找到一种平衡,使得益处明显超过风险,并且风险可以得到管理。AI的发展速度比以往技术快,我们需要从这些历史中学习,以期达到类似的平衡。
AI在教育领域的潜力是什么,它将如何改变教育?
-AI在教育领域的潜力在于提供个性化学习体验。它将允许人类教师快速了解特定学习者的状态,从而提供定制化的教学。此外,AI还可以使远程教育变得更加有效,通过提高教师的教学效率,让教育规模化的同时更加个性化。
AI在交通领域的应用将如何发展,特别是在车辆自主性方面?
-在未来20年内,我们可以预见到20%到25%的路面车辆将实现某种形式的自主性。但这种自主性可能不会超过这个比例,因为实现完全的自动驾驶还面临着许多技术和伦理挑战。
AI如何帮助解决复杂的全球性问题,比如水资源短缺?
-AI可以通过模拟和优化来帮助解决复杂的全球性问题。例如,AI可以模拟水流和分配,帮助我们更好地理解和管理水资源。此外,AI还可以帮助设计更高效的灌溉系统和水处理技术,从而提高水资源的利用效率。
在AI领域,为什么需要考虑去中心化的重要性?
-去中心化在AI领域很重要,因为它可以防止数据和计算能力集中在少数大公司手中。去中心化可以促进创新,让更多的参与者进入市场,并且可以更好地保护个人隐私和数据安全。此外,去中心化还可以提高系统的鲁棒性和抗风险能力。
AI在模拟环境中的作用是什么,它如何帮助人类提高技能?
-AI在模拟环境中可以进行大量的训练和学习,比如在棋类游戏中,AI可以快速学习并掌握游戏规则,达到超越人类专家的水平。这种能力可以帮助人类通过与AI对弈来提高自己的技能,同时也可以让AI帮助人类探索新的策略和创意。
AI在资源管理方面有哪些应用,它如何提高效率?
-AI在资源管理方面的应用包括在大数据中心里分配资源、优化分布式计算资源的使用等。AI可以帮助预测资源需求,自动调整资源分配,从而提高资源利用效率,降低运营成本。
AI在现实世界中的应用面临的最大挑战是什么?
-AI在现实世界中的应用面临的最大挑战是如何将AI技术与现实世界的复杂性结合起来。现实世界的不确定性和多变性要求AI系统不仅要有高度的智能,还要有良好的适应性和鲁棒性。此外,还需要解决伦理、法律和社会接受度等问题。
AI技术发展的速度是否超过了我们的治理能力?
-是的,AI技术的发展速度非常快,这在一定程度上超过了我们的治理能力。这意味着我们需要加快制定相关的法律法规,建立伦理标准,并提高公众对AI技术的理解,以确保AI技术的健康发展。
AI技术在短期内是否会导致大规模的失业?
-AI技术在短期内可能会导致某些工作岗位的消失,但历史上的技术变革也证明了,新技术的出现往往也会创造新的工作岗位和职业机会。关键在于社会如何适应这种变化,以及如何通过教育和培训帮助劳动力转型。
Outlines
🤖 AI的未来与社会适应
第一段主要讨论了AI的未来发展及其对社会的影响。提到了rames rcar和Hari bar Christen两位专家,分别在媒体实验室和计算机科学领域有所成就。讨论了AI可能带来的变革,包括对教育、工作和日常生活的影响。强调了AI发展速度之快,以及如何从以往的技术革命中学习,找到风险与收益之间的平衡。同时,表达了对AI在医疗、交通和教育领域应用的乐观态度,但也指出了AI在现实世界中应用的挑战。
🚀 AI的积极展望与潜在风险
第二段中,讨论者表达了对AI未来发展的乐观态度,认为AI不会带来人类的灭绝。他们预见了AI在医疗、交通和教育领域的积极作用,尤其是AI在个性化教育中的应用。同时,也提到了AI可能带来的风险,如因AI导致的死亡和对工作的影响。讨论了AI在资源管理中的应用,以及它如何可能改变人类的工作方式,包括创造新的工作机会和提高人类的创造力。
🌐 AI的去中心化趋势
第三段探讨了AI的去中心化趋势,以及这一趋势如何影响全球经济和各个行业。讨论了苏联模式的失败和AI在资源分配中的作用,以及AI如何在不吸引大投资的情况下帮助解决全球性问题。强调了AI在提高生产力和改变经济模式方面的潜力,以及它如何可能带来新的商业机会和挑战现有的经济结构。
🧠 AI与人类智能的结合
第四段讨论了AI与人类智能结合的可能性,以及这种结合如何帮助解决复杂的社会问题。提到了AI作为顾问在提升中低技能人群能力方面的潜力,以及AI在法律等专业领域的应用。讨论了AI在解决水危机等全球性问题中的潜在作用,以及AI如何帮助科学家和工程师在复杂系统中找到解决方案。
🌿 AI的民主化与资源限制
第五段强调了AI民主化的重要性,以及如何通过民主化来缩小技术拥有者和非拥有者之间的差距。讨论了AI硬件的发展,特别是为特定任务设计的高效硬件,以及AI在解决实际问题中的应用。同时,也提到了数据和资源限制对AI发展的影响,以及如何通过联邦学习和分布式系统来克服这些限制。
📈 AI的投资周期与市场潜力
第六段讨论了AI技术的投资周期,以及当前对某些AI领域的过度投资和对其他领域的投资不足。提到了互联网时代投资周期的类比,以及在基础设施建设之后,基于这些技术的应用层公司如何取得更大的成功。强调了中国和印度等国家在AI领域的快速发展,以及这些地区可能成为未来AI技术和应用的重要来源。
Mindmap
Keywords
Please replace the link and try again.
Highlights
Ramesh Raskar 是媒体实验室的同事,也是 SIGGRAPH 奖的获得者,拥有约 100 项专利,最近创立了 C10 AI Ventures,这是一个风险工作室,他担任首席科学家。
Hari Balakrishnan 是计算机科学和 AI 领域的知名人物,最近创立了 CMI,这是世界上最大的远程信息服务平台,致力于在道路上保障安全。
讨论了 AI 的未来方向,以及世界如何适应 AI,特别是从 MIT 的专家角度来看。
提到了关于机器人接管世界的普遍担忧,以及这些担忧如何与 AI 发展相关联。
讨论了 AI 在未来 20 年可能对世界产生的影响,包括教育、工作角色和社会角色的变化。
强调了以往技术革命(如互联网、基因编辑、核能)带来的变革,以及我们如何找到合理的平衡点。
Ramesh Raskar 表达了对 AI 未来 20 年的乐观态度,认为 AI 将帮助改善健康和提高生活质量。
预计在未来 20 年内,大约 20-25% 的车辆将实现某种形式的自主性。
讨论了 AI 在教育领域的潜力,以及它如何使大规模教育变得更加个性化。
提到了 AI 在未来可能导致人类死亡的风险,以及社会将如何应对这些风险。
讨论了 AI 对就业的影响,以及人们如何适应技术变革带来的职业变化。
提出了 AI 将如何帮助人类学习新事物,并提高我们的智能。
讨论了 AI 在模拟环境中的表现,以及它如何使人类在某些领域(如国际象棋)中变得更好。
强调了 AI 在资源管理方面的潜力,以及它在大数据和分布式资源分配中的应用。
讨论了 AI 在现实世界中的挑战,包括机器人技术、医学和工厂自动化等领域。
预测了 AI 将如何通过提高生产力来改变联合国可持续发展目标相关领域的经济模式。
讨论了 AI 在法律等专业领域的潜力,以及它如何使法律服务变得更加经济实惠和易于获取。
强调了 AI 在解决复杂问题(如微生物组研究)方面的潜力,以及它如何帮助我们理解整体功能的复杂性。
讨论了 AI 在大型系统中的应用,以及它如何帮助我们处理复杂的交互和多维数据。
Ramesh Raskar 表达了对 AI 民主化和分散化的看法,以及它如何影响硬件发展和专用芯片的设计。
讨论了 AI 生态系统中的经济力量和激励机制,以及它们如何影响 AI 的发展和应用。
预测了 AI 技术在接下来的 3 到 5 年内将在目前投资不足的领域取得显著进展。
提醒听众,西方公司并不是地球上唯一的参与者,中国和印度等国家的 AI 投资周期和数据系统增长非常迅速。
Transcripts
you know so we're going to talk a little
bit about that so we have rames rcar who
is a colleague of mine here in the media
lab uh he's the winner of the llon prize
sigraph prize you have what a 100
patents now something like that oh crazy
uh and uh he just started C10 uh AI
Ventures which is a venture studio uh
where you're the chief scientist so we
hope to hear less about that and Hari
bar Christen who's over in computer
science and AI uh and is very famous but
more recently he started something
called CMI which is the world's largest
telematic service provider helps keep
you safe on the roads Fair okay good so
what I thought we would do here is have
a conversation about where is AI going
how is the world going to adapt to it uh
at from the perspective of the people
that you see here right so this is like
the
MIT these people are are the best that
we have here right so good okay so I
thought I'd start with uh something that
was the motivation for this panel to
begin with back at the beginning which
is so the robot overlords are coming
right the robot overlords are going to
kill us all or take over we're going to
work for them and then they're going to
decide that we aren't any good and
they're going to get rid of us that's
the sort of thing you hear a lot lot of
people say uh and all of this is due to
AI somehow and I wonder if you guys have
sort of an idea what the world is going
to look like in say 20 years because of
all this AI stuff I mean I don't think
either of you are in the extinction Camp
I don't think so I don't think so two of
my both daughters are here sitting in
the audience you know they're under 15
both of them and when I think about
first of all what should they learn in
school uh to all the way to what will
our role be all three of us are
professors but also entrepreneurs um
what it would be and also what's the
role for society I think it's a very
interesting time uh to think about that
I think without nobody knows uh what's
going to happen but if you think about
the previous revolutions that have
projected massive changes whether it's
internet whether it's crisper uh whether
it's you know nuclear energy uh nuclear
power
uh there have been a lot of discussions
uh about this risks uh and their
governance and eventually we have
figured out a reasonable equilibrium uh
where the
benefits uh will definitely outweigh the
risks and the risks can be managed I
think what we can learn in AI is is is
going to be very interesting
unfortunately it's moving much faster
than the previous two technologies I
mean crisper only now we have an FDA
approval approved approved uh treatment
so it took years after you know
chrisopher came on the scene uh with AI
That's not the case so can we learn from
previous risky but exciting Innovations
or not is is an open question oh har
yeah um so I'm I'm a techno optimist
about about these things and I think
that there's a few things that will work
in the next 20 years and a few things
that I think we are crazy to imagine
that it'll work so I think on the crazy
side I don't think this is going to lead
to Extinction or any of those um types
of uh possibilities um I think that on
the positive side I I'm looking forward
to a time when AI can help improve our
health um you know with implants with
not only for physical stuff but also for
uh you know to improve our memory and to
make that uh so that it's almost like
you have that as an assistant um I think
those things are going to become real um
I suspect that in 20 years we will
probably see 20 to 25% % of the vehicles
on the road worldwide have a significant
form of autonomy um it's probably not
going to be more than that worldwide um
and I suspect that the real challenge uh
that we will continue to tackle is
operationalizing AI um you know I think
there are going to be some use cases
where it completely replaces humans and
it's completely automated and you know
everything is actuated and done with AI
but I suspect that in a lot of other
cases figuring out how to operationalize
that in the rest of the business work
flow um would be a significant Challenge
and I think just specifically with
education I think you know it we've
always wanted technology to improve
education and largely speaking very few
technologies have I suspect AI will
actually make a huge difference um but
it's an example where um it's not going
to replace uh human teachers I think
that it actually will allow uh education
at scale to become far more personalized
by allowing the human teacher to very
quickly get a summary of where that
particular learner stands and you could
make this completely remote and have
great telepresence so you know they
could be in a different you completely
different part of the world so I think
there's a lot of that that will happen
and I think 20 years is a good time
frame for that I think on the negative
side we will see people die due to Ai
and I think there will be a reaction to
that but we'll eventually cope with it
because I think it's a technology that
can't be stopped I think one of the
things that um I was had my attention
drawn to recently is the first wave of
AI which was in the 50s one part of it
was about logic but the other thing was
optimal resource uh allocation which is
essentially linear solving under
constraints and this was the thing that
was going to save the Soviet economy it
was the planned economy and it didn't
work out real well but on the other hand
uh every spreadsheet on the planet has
this sort of thing built into it it's
the most common con reputation out there
so it may be that all the things about
you know robot overlords or infinite uh
uh you know advanced intelligence and
stuff isn't going to happen any better
than the Soviet Union planned economy
but we're going to see it
everywhere and and I think s pick up on
the thing that you just said which is
that so what happens to the people so
one of the sets of Visions people have
about this is that all the people will
be out of work what are we going to do
we got to have Universal basic income
things like that the other thing that is
the more uh historical thing is is we
find new things to do and that new sort
of productivity helps what what what
I'll start go the other direction since
you brought it up so I mean the short
answer is we are all guessing as to
what's going to happen to the uh to the
employment or to to the occupation or or
Hobbies of of people I mean it's clear
that some jobs will vanish but that's
always been true with any technological
change that's happened things things
just kind of vanish I think that what um
if we're sort of addressing people who
are probably younger than 25 or 30 um I
think that no longer already it's not
the case that what you learn before the
age of 20 or 25 gives you the ability to
have a 40 to 45 year career that's just
gone actually it's tough for us we're
learning new things from our students
every day more so than they learning
from us and it's it's it's a you're
supposed to tell people that yeah I know
but but I think AI will help us learn
better it'll help us learn how to learn
new things it'll help us improve it and
I do believe this idea of us becoming a
lot smarter um you know if we choose to
become smarter uh would be a positive as
for what people will do you know the
story was you know the AI will do all of
the hard stuff The Drudge work and you
know we can write poetry it looks like
the opposite is happening so um I feel
that at the end of the day there are
three classes of AI very very broad
generalization I'm going to a lot of
stuff there's AI that works really well
in simulation you know this is it'll
beat everybody playing any game like
chess or go or whatever but actually the
Advent of AI and algorithms in chess has
made people much more excited about
human chess because humans are becoming
better at chess due to those machines
and I think that we will see a lot of
that happen in other walks of life we'll
start to see how humans can become
better painters or better songwriters of
what have you using those tools so I
feel like that will happen and AI will
absolutely uh change those fields the
second is in things like Resource
Management like in Big Data Centers or
allocating large scale distributed
decentralized resources and I think
there AI research has already shown that
it's going to work I think we'll just
see more and more of that happening the
biggest challenge for AI and I think
where humans will continue to be in the
loop is AI in the real world whether it
be robotics whether it be medicine
whether it be in Factory floors and so
on and I think that it is going to be a
new form of machine human symbiosis um
where
um one possibility is that humans act in
a more supervisory role the other
possibility is the AI acts in a more
supervisory role and be interesting to
see how it how it turns out what yeah
I'm really glad you brought this point
of um does optimization take us into new
regimes because if you think about you
know un's sustainable development goals
which are about water and health and
poverty and Agriculture and so on um the
the challenge has has been that many of
those fields just don't have the unit
economics to become you know multi-
trillion dollar businesses so there's no
trillion dollar company in the sectors I
just mentioned and and the reason for
that is the margins are so low that
doesn't attract the top talent um but
what could happen over time just the way
it took a long time for solar to become
comparable to fossil fuels and only now
the unit economics works I think the
productivity gains because of AI will
similarly bring unit economics so it'll
be exciting to work in water and
agriculture because the productivity is
so high that right now you your margins
may be 10% or minus 10% but very soon
the margins might be 100% to 200% so I
think that's an exciting time that as as
starts coming into the real world uh
you'll start seeing this exciting
opportunity and even for C10 lab which
is a venture Studio we launched here
that's the single hypothesis which is
productivity gains in are areas that are
unexplored will give you unreal you know
you know tremendous gains they'll become
very very lucrative the question is the
world GDP is1 trillion how can we figure
out what are the sectors that'll impact
in the next two years versus be impacted
in the next 10 years and if you can do
the right matchmaking between
opportunist in Ai and opport in the real
world sectors I think we can go very far
and then one quick point to add to your
question about the Soviet model is
unfortunately that's how it's working
right now actually shantu bachara who's
a scientist in our group gave me the
same analogy for Soviet he said you know
when we're thinking about C
supercomputers or IBM mainframes in the
80s we thought that's the way computer
is going to play out you know right here
on one Route 128 but very quickly we had
the PC and the mobile and the iot and
things became highly decentralized so
very kind of a Soviet mindset of
centralize everything the way we are
doing it with big companies right now
centralized data centralized compute
centralized
governance and we just use their apis
you know gp4 and um cloud and so on um I
think that model is going to shatter
very soon uh and we'll see this
decentralization of a come in so I think
the intersection of productivity gains
in the real world and ability to do this
in a very decentralized way are two
major trends that are going to intersect
very soon can I push back on something
yes please okay so I sort of agree with
the decentralization but I think that's
because of the highly data Centric
nature of it and maybe we'll cover that
later with the fact that more and more
organizations and entities don't want
all of the data to be shipped to some
centralized place I totally agree with
you but I think with AI it's not clear
to me that this long tale of um
businesses or u a lot opportunities for
society where there's not enough of a
market to justify big investment uh AI
is really well suited because in the
short term I actually feel from some
experience we've had at my company and
other startups I talk to that the
short-term costs of everything from
large scale data acquisition to training
to everything else actually you're far
better off hiring a bunch of very smart
human beings more economical to get
going like that so I wonder about
whether this hypothesis that AI is going
to help us scale the heavy tail which
will never get that big is really true I
mean if you look at all the big
companies right now forget the oil era
of Exxon and Exon Mobile and so on but
nearly all top 10 fortune f companies
have all made their their their revenue
on the long tail you know Facebook um uh
meta sorry Facebook that's not long T
that's 3 billion four billion people
using I mean all of us are buying
$500,000 mobile phones as opposed to
companies that are selling large servers
uh same thing if you look at the catalog
of of Amazon there's a huge number of
long the long is where the money treat
those uniformly is what makes it
profitable yeah yeah I think your
question is also about centralization of
talent because I was talking about
centralization of you know compute data
and and and governance centralization of
talent is a very important thing but
like you said AI is going to change
education so the smart kids in Tanzania
will be 80 90% of the talent level as
the best people out there so I think
we're going to see a lot of
decentralization of those opportunities
as well um and it remains to be seen
whether you know there's always going to
be a Delta between you know a talent
available in Tanzania versus Talent
available somewhere else so and
decentralization on the other hand
because of Regulation because of Trade
Secrets you know if I think about the
health system in Tanzania just to pick
an example you know the insurance
players there the hospitals there the
government systems are not going to wake
up one day say like let me send all my
data to Sam and Sam is going to build
this really nice model for us and we'll
just use Sam's apis and and you know run
the healthare system what I see in all
the companies that I work with is
they're all building little AI models on
their data they use some like something
like llama to begin with but then they
specialize it with their process data
right and and that what they hope to do
is get something that's better than just
a linear constraint solver or an expert
system and they do it's not that hard
and because a graph is going to be much
a graph network is going to be much more
stable we're talking about risks and
opportunities in this panel you know a
graph is much more stable than you know
a a a hierarchical tree um so I mean you
you mentioned the the different ways of
AI you know Marvin miny called it the
Society of the mind he didn't call it
the mind like one Mega AI he said a
society of multiple small AIS and we are
all by the way are multiple small AIS
like even a large organization has an HR
department and a sales department and
marketing department and tech department
are all small AIS human AIS uh a CEO
doesn't run all those departments and uh
I think the same thing is true for you
know future so one of the the most
hopeful things is that most of the
experiments people have done on making
AI as an advisor helps middle skill
people or lower skill people more than
High skll people and and that's actually
really interesting because that allows
you to onboard people it allows you to
to make up some of the skills Gap there
was an interesting illustration I ran
into yesterday which is someone was
talking about AI for law so AI is very
effective in law but currently law is so
expensive that you can't afford to hire
a lawyer you can't get representative if
you could get something that was pretty
good yeah cheaper that would help lots
of people again economics it's changing
the yeah it's changing the economics of
the whole thing yeah I think it's true
in many professions I just don't know
about some of these like solving the
Water Crisis for example whether AI
magically tackles those types of the
problems yeah I mean if you have I think
that's a good question so like you know
no like it's not considered a glamorous
job to go figure out water but if a
starts behaving like a scientist and you
know you have this you know amazing
tools not AI as a as a assistant a as an
engineer but a as a scientist uh then
they can create all the solutions on the
Fly you know the geometric design the
mechanical design you know the cfds and
so on and a lot of these issues can be
solved same thing with on the
incentivization of how do you not just
invent but how do you have a diffusion
of those inventions in the society and
again you know unfortunately the whole
wave of decentralized Finance didn't
work out as we wanted but the
incentivization mechanisms if they if
they are done right and S is going to do
it for us um you know both the invention
as well as dissemination of those
invention I think one of the worst
things that happened is all of the
crypto stuff where you don't have
identity you don't have all sort of
normal sort of guard rails on it and as
a consequence theft and fraud in in
crypto has ended up uh just destroying a
lot of things bad reputation for a lot
of this stuff come back funded 50% of
the North Korean uh nuclear missile
program I mean come on this is this is
pretty bad so there are some real
downsides to this stuff right um one of
the more interesting things that we've
run into is uh looking at really
complicated things like the microbiome
and it turns out that you know we don't
it's millions of organisms and and
hundreds of thousands of of little RNA
fragments in there but it turns out that
if you actually get enough samples of
that you can begin understanding how the
thing as a whole functions and we've
been able to do it for you know uh CO2
reduction and and are looking at it for
human health so some there's a lot of
things where the reductionist approach
just sort of doesn't work right because
it's all connected and and maybe this is
some way where we can begin to get a
handle on really complex phenomena
you're not in your head so I'm going to
yeah I agree yeah I mean we see this in
uh large scale systems right of any kind
where uh you really have to you know
it's really they're all decomposed into
individual components but the
interactions are so complicated that AI
as a tool uh has been tremendously
useful and I'm not even talking about
generative AI just things like
enforcement learning work really
well yeah I mean I think this is
something that people don't appreciate
is that a lot of the reason the tools we
have today don't work is because they're
sort of abstractions of the situation
and the real situation is a lot more
complicated and one of the strengths of
this sort of thing is is it has lots of
Dimensions so you can begin to actually
model some of that complexity raes what
what have you seen in this sort of space
um I think I think the the
I mean I'm a I'm a kind of a
decentralization maximalist here and I
think the the equilibrium is going to
come when we have democratization of AI
so that's enough players and you don't
have this big gap between the halves and
Have Nots because the moment that
happens when you only have one big chip
company like it's happening with Nvidia
if you have only one big co-pilot
company like Microsoft versus when there
are a lot of players in this space uh
and they're distributed geographically
as well I think we're going to see this
you know interesting back and fourth you
do need the leaders to bring the
Innovation first you know Nvidia has to
spend billions of dollars to bring us
you know highly compact chips that have
a very small energy footprint and so on
so I just worry about how long does it
take to go from the SC super computers
to iot uh and so on so that's one of the
biggest things that that bothers me
about how long is it going to take us to
go this is one of the things that
everybody talks about is it's going to
take a trillion dollars to scale this
stuff and you know but wait I don't
really want an AI that speaks Romanian
it doesn't do me any good special I want
narrow things that solve real problems
and those are lots easier and that's
sort of this beginning path to like AI
everywhere as opposed to the over Uber
things and then I see a lot of things in
Hardware too uh as6 being built special
for this sort of stuff that are much
more efficient than the current gpus
what what do you think is going to
happen in this ecology how how are we
going to get to sort of AI everywhere
where AI is maybe with small letters
rather than big letters yeah I think
it's inevitable I I don't know if it's
going to be completely decentralized the
way you're talking but I think there are
two big trends people don't in many
cases don't want all of the data shipped
to some other entity there are
regulations against it uh people feel
uncomfortable about sharing their own
like you know you take your home audio
Alexa type system um but I think that
the way this will come about is
um specialized models and highly
Federated learning and uh the way I like
to think about it is there's sort of
this big model but then it's a big
distributed system and you end up being
able to partition that amongst different
pieces and different entities which each
has some well- defined set of things
that it provides into the higher layers
so in some cases you have completely
independent autonomous decision making
in some other cases they do need to
collaborate and you don't have to do
this with uh just you know some
generative AI you could use graphical
neural networks there are many many
different techniques that we've
developed I mean distributed
reinforcement learning and so I think we
should be thinking about it very broadly
but it's inevitable it's going to happen
and there are many cases where there are
fundamental resource constraints that
prevent you from delivering all of the
data all the time to a central entity I
mean if you imagine the 1.4 billion
vehicles in the world all delivering 60
frames a second video all the time from
you know continuously there's not really
it's going to kill the networks and I
don't think anyone actually even wants
to do that and you would end up with a
high degree of uh dis I mean if if you
think about human you know how the human
society behaves uh and I'm sure
anthropologists are going to play a big
role in how kind of governance of you
know AI are going to work and and so you
know as a human society we have figured
out how each of us have expertise and we
help each other out and you know we
can't imagine you know a super brilliant
person governing all of us at least for
most of the time most of the time uh
most of the time and and so I think this
this uh this the challenge of
decentralization is key but that alone
is not going to solve the problem
because you also want the economic
forces so you need the incentivization
of why should somebody should some why
should somebody participate in the AI
ecosystem are they going to get paid in
dollars or this fuzzy tokens from
cryptocurrencies or they're going to get
paid in compute or they get ahead of the
queue because so I think there's going
to be a lot of incentive mechanisms that
you have to design as well so some of
the research we do in our group here uh
in decentralized AI is actually thinking
about data markets model markets and
incentive mechanisms um and it's one of
those areas that seems adjacent to uh
progress in AI but I think they'll
converge very soon instead of behavior
economics we'll call it AI economics
okay so we're almost out of time here
one last question where are we in the
hype
cycle well with generative we're near
the peak or maybe we're about so we're
going to crash next year well I think
that what people will find is a
tremendous amount of misinformation and
I don't know if it'll crash next year
but uh 18 months 18 months okay there
you are get your investments in get out
in 18 months well then go back up well
no it's always goes back up but that
could be 10 years right yeah um I I
would say there's a absolutely a lot of
overinvestment in certain areas but
surprisingly underinvestment in many
other areas and uh if you can I would
say there's underinvestment in most of
the areas so as we said earlier many
real world sectors are just not getting
enough attention that's right and
they're going to they're going to just
Bloom they're just going to Boom uh in
the next 3 to five years not right away
because it takes a long time for new
technology to be absorbed you know to
see the productivity gains for rest of
the ecosystem to have right protocols to
work with each other and so that'll take
some time but over the next 3 to 5 years
you'll see these highly underinvested
areas really take off and you can use
analogy from the internet era you know
in the beginning everybody bought stock
in you know chip companies and Os
companies right and Cisco and IBM and so
on but very soon what was built on top
of that which is the applied internet
you know eBay Yahoo Google Facebook and
so on they're so much bigger than
anybody who's selling I mean not right
now with Nvidia but but in the internet
era anybody who was selling chips or you
know internet infrastructure okay
infrastructure play which is highly over
invested right now is uh I don't know 18
months two years okay and let me just
remind people that Western companies are
not the only ones on the planet and that
the investment cycle in China is very
different and they have a lot of
Engineers and they're are 100% on it and
they don't worry too much about who owns
the data uh and the fastest growing area
in the world in terms of data systems is
India and the surrounding countries
there they've gone from essentially zero
to a couple billion on some of their
systems in the last two three
years guess what that's where a lot of
it's going to come from not from the
people in this room so keep your radar
out I just kind I just mention um so the
decentralized AI team working on that is
downstairs and they're showing demos on
how you can use decentral AI for the
Indian stack so please go check them out
I think they're right on the bottom
floor sounds good that's the backend
people yeah backend people yeah okay
good okay thank you know who's supposed
to be up next
浏览更多相关视频
In conversation with the Godfather of AI
AI and Quantum Computing: Glimpsing the Near Future
Google CEO SHOCKS Everyone " We MUST Prepare NOW!" NEW AI Prediction
2 Ex-AI CEOs Debate the Future of AI w/ Emad Mostaque & Nat Friedman | EP #98
AGI: solved already?
《與楊立昆的對話:人工智能是生命線還是地雷?》- World Governments Summit
5.0 / 5 (0 votes)