Elon Musk: Tesla Autopilot
AI 与机器学习技术与编程心理与人性哲学与宗教历史与文明
🤖
AI 智能总结
马斯克谈特斯拉自动驾驶、AI与人类未来
这是 Lex Fridman 与埃隆·马斯克的首次对话，录制于 MIT 关于特斯拉 Autopilot 驾驶员行为研究发布后不久。对话聚焦于自动驾驶技术的愿景、AI 安全、人类意识的本质，以及马斯克对文明未来的深层思考。
自动驾驶人工智能特斯拉AI安全意识与模拟宇宙
埃隆·马斯克（Elon Musk）是特斯拉、SpaceX、Neuralink 等公司的 CEO 和联合创始人，被认为是当代最具影响力的科技企业家之一，致力于电动汽车、太空探索和脑机接口等领域的变革。
📌 核心观点
- 马斯克认为特斯拉 Autopilot 的核心目标不是监控驾驶员，而是让自动驾驶系统本身足够安全，使人类行为因素变得无关紧要——这与 Lex 的摄像头监控方案形成直接分歧。
- 他预测完全自动驾驶（FSD）将在数年内实现，核心挑战是处理长尾边缘案例，而非基础感知能力，神经网络已经能够处理绝大多数常见场景。
- 马斯克对 AI 存在性风险持严肃态度，认为超级智能 AI 是人类面临的最大生存威胁之一，这也是他创立 OpenAI（后来离开）和 Neuralink 的动机之一。
- 关于意识和模拟宇宙，马斯克认为我们很可能生活在模拟中，概率接近十亿分之一，因为基础现实只有一个，而模拟可以无限多。
- 他分享了对人类文明延续的使命感：成为多星球物种是防止文明灭绝的保险，SpaceX 的火星计划不是逃离地球，而是为人类文明购买一份备份。
✨ 金句摘录
马斯克：自动驾驶的统计安全收益将超越任何对人类行为和心理的担忧。
马斯克：我们很可能生活在模拟中。基础现实只有一个，但模拟可以有无数个。
马斯克：成为多星球物种是人类文明的生命保险，不是逃跑，是备份。
📋 章节目录
暂无章节信息
🔑 关键词
autopilotcarteslafulldrivingdatacarscomputerdriverhumanelevatormonitoringsafetydrivedonselfpersonsimulationconversationvision
💬 精彩语录
暂无语录
🎙️ 完整对话（503 条）
Lex Fridman (00:00.000)
The following is a conversation with Elon Musk.
以下是与埃隆·马斯克的对话。
Lex Fridman (00:03.000)
He's the CEO of Tesla, SpaceX, Neuralink, and a cofounder of several other companies.
他是 Tesla、SpaceX、Neuralink 的首席执行官，也是其他几家公司的联合创始人。
Lex Fridman (00:09.200)
This conversation is part of the Artificial Intelligence podcast.
这段对话是人工智能播客的一部分。
Lex Fridman (00:13.180)
The series includes leading researchers in academia and industry, including CEOs
该系列包括学术界和工业界的领先研究人员，其中包括首席执行官
Lex Fridman (00:18.540)
and CTOs of automotive, robotics, AI, and technology companies.
汽车、机器人、人工智能和科技公司的首席技术官。
Elon Musk (00:24.060)
This conversation happened after the release of the paper from our group at MIT
这次谈话发生在我们麻省理工学院小组发表论文之后
Lex Fridman (00:28.260)
on Driver Functional Vigilance, during use of Tesla's Autopilot.
在使用特斯拉自动驾驶仪期间，关于驾驶员功能警惕性。
Elon Musk (00:32.540)
The Tesla team reached out to me offering a podcast conversation with Mr.
特斯拉团队联系了我，提供了与特斯拉先生的播客对话。
Lex Fridman (00:36.140)
Musk.
马斯克。
Elon Musk (00:37.140)
I accepted, with full control of questions I could ask and the choice
我接受了，并完全控制我可以提出的问题和选择
Lex Fridman (00:41.280)
of what is released publicly.
公开发布的内容。
Elon Musk (00:43.220)
I ended up editing out nothing of substance.
我最终没有编辑任何实质性内容。
Lex Fridman (00:46.480)
I've never spoken with Elon before this conversation, publicly or privately.
在这次谈话之前，我从未与埃隆公开或私下交谈过。
Elon Musk (00:51.380)
Neither he nor his companies have any influence on my opinion, nor on the rigor
他和他的公司对我的观点和严谨性都没有任何影响
Lex Fridman (00:56.220)
and integrity of the scientific method that I practice in my position at MIT.
以及我在麻省理工学院所实践的科学方法的完整性。
Elon Musk (01:01.360)
Tesla has never financially supported my research, and I've never owned a Tesla
特斯拉从未在经济上支持我的研究，我也从未拥有过特斯拉
Lex Fridman (01:05.940)
vehicle, and I've never owned Tesla stock.
车，而且我从未拥有过特斯拉股票。
Elon Musk (01:09.700)
This podcast is not a scientific paper.
该播客不是科学论文。
Lex Fridman (01:12.340)
It is a conversation.
这是一次对话。
Elon Musk (01:13.880)
I respect Elon as I do all other leaders and engineers I've spoken with.
我尊重埃隆，就像我尊重所有与我交谈过的其他领导者和工程师一样。
Lex Fridman (01:18.220)
We agree on some things and disagree on others.
Elon Musk (01:20.980)
My goal is always with these conversations is to understand the way
Lex Fridman (01:24.560)
the guest sees the world.
Elon Musk (01:26.860)
One particular point of disagreement in this conversation was the extent to
Lex Fridman (01:30.700)
which camera based driver monitoring will improve outcomes and for how long
Elon Musk (01:36.040)
it will remain relevant for AI assisted driving.
Lex Fridman (01:39.900)
As someone who works on and is fascinated by human centered artificial
Elon Musk (01:44.080)
intelligence, I believe that if implemented and integrated effectively,
Lex Fridman (01:48.640)
camera based driver monitoring is likely to be of benefit in both the short
Elon Musk (01:52.880)
term and the long term.
Lex Fridman (01:55.580)
In contrast, Elon and Tesla's focus is on the improvement of autopilot such
Elon Musk (02:01.520)
that it's statistical safety benefits override any concern of human behavior
Lex Fridman (02:06.960)
and psychology.
Elon Musk (02:09.000)
Elon and I may not agree on everything, but I deeply respect the engineering
Lex Fridman (02:13.880)
and innovation behind the efforts that he leads.
Elon Musk (02:16.800)
My goal here is to catalyze a rigorous nuanced and objective discussion in
Lex Fridman (02:21.920)
industry and academia on AI assisted driving.
Elon Musk (02:26.180)
One that ultimately makes for a safer and better world.
Lex Fridman (02:30.820)
And now here's my conversation with Elon Musk.
Lex Fridman (02:35.560)
What was the vision, the dream of autopilot when, in the beginning, the
Lex Fridman (02:40.140)
big picture system level, when it was first conceived and started being
Lex Fridman (02:44.440)
installed in 2014, the hardware and the cars, what was the vision, the dream?
Lex Fridman (02:48.900)
I wouldn't characterize the vision or dream, simply that there are obviously
Elon Musk (02:52.200)
two massive revolutions in, in the automobile industry.
Lex Fridman (02:59.280)
One is the transition to electrification and then the other is autonomy.
Lex Fridman (03:06.920)
And it became obvious to me that in the future, any car that does not have
Lex Fridman (03:14.420)
autonomy would be about as useful as a horse, which is not to say that
Elon Musk (03:19.360)
there's no use, it's just rare and somewhat idiosyncratic if somebody
Lex Fridman (03:23.540)
has a horse at this point.
Elon Musk (03:24.900)
It's just obvious that cars will drive themselves completely.
Lex Fridman (03:27.480)
It's just a question of time.
Lex Fridman (03:29.040)
And if we did not participate in the autonomy revolution, then our cars
Lex Fridman (03:37.340)
would not be useful to people relative to cars that are autonomous.
Elon Musk (03:42.260)
I mean, an autonomous car is arguably worth five to 10 times more than
Lex Fridman (03:49.100)
a car which is not autonomous.
Elon Musk (03:52.480)
In the long term.
Lex Fridman (03:53.780)
Turns out what you mean by long term, but let's say at least for the
Elon Musk (03:57.140)
next five years, perhaps 10 years.
Lex Fridman (04:00.240)
So there are a lot of very interesting design choices with autopilot early on.
Elon Musk (04:04.520)
First is showing on the instrument cluster or in the Model 3 on the
Lex Fridman (04:10.120)
center stack display, what the combined sensor suite sees, what was the
Lex Fridman (04:15.200)
thinking behind that choice?
Lex Fridman (04:16.600)
Was there a debate?
Lex Fridman (04:17.600)
What was the process?
Lex Fridman (04:19.160)
The whole point of the display is to provide a health check on the
Elon Musk (04:25.480)
vehicle's perception of reality.
Lex Fridman (04:26.800)
So the vehicle's taking information from a bunch of sensors, primarily
Elon Musk (04:30.440)
cameras, but also radar and ultrasonics, GPS, and so forth.
Lex Fridman (04:34.540)
And then that, that information is then rendered into vector space and that,
Elon Musk (04:41.720)
you know, with a bunch of objects with, with properties like lane lines and
Lex Fridman (04:46.480)
traffic lights and other cars.
Lex Fridman (04:48.460)
And then in vector space that is rerendered onto a display.
Lex Fridman (04:53.400)
So you can confirm whether the car knows what's going on or not
Elon Musk (04:58.060)
by looking out the window.
Lex Fridman (04:59.860)
Right.
Elon Musk (05:00.200)
I think that's an extremely powerful thing for people to get an understanding.
Lex Fridman (05:04.940)
So it become one with the system and understanding what
Elon Musk (05:06.960)
the system is capable of.
Lex Fridman (05:08.840)
Now, have you considered showing more?
Lex Fridman (05:11.840)
So if we look at the computer vision, you know, like road segmentation,
Lex Fridman (05:16.040)
lane detection, vehicle detection, object detection, underlying the system,
Elon Musk (05:19.920)
there is at the edges, some uncertainty.
Lex Fridman (05:22.680)
Have you considered revealing the parts that the vehicle is
Elon Musk (05:28.280)
in, the parts that the, the uncertainty in the system, the sort of probabilities
Lex Fridman (05:34.260)
associated with, with say image recognition or something like that?
Elon Musk (05:36.760)
Yeah.
Lex Fridman (05:37.160)
So right now it shows like the vehicles in the vicinity, a very clean, crisp image.
Lex Fridman (05:41.840)
And people do confirm that there's a car in front of me and the system
Lex Fridman (05:45.140)
sees there's a car in front of me, but to help people build an intuition
Elon Musk (05:49.000)
of what computer vision is by showing some of the uncertainty.
Lex Fridman (05:53.040)
Well, I think it's, in my car, I always look at the sort of the debug view.
Lex Fridman (05:57.440)
And there's, there's two debug views.
Lex Fridman (05:59.440)
Uh, one is augmented vision, uh, where, which I'm sure you've seen where it's
Elon Musk (06:04.620)
basically, we draw boxes and labels around objects that are recognized.
Lex Fridman (06:10.760)
And then there's a work called the visualizer, which is basically vector
Elon Musk (06:15.960)
space representation, summing up the input from all sensors that doesn't,
Lex Fridman (06:22.340)
that doesn't, does not show any pictures, but it shows, uh, all of the, it's
Elon Musk (06:28.180)
basically shows the car's view of, of, of the world in vector space.
Lex Fridman (06:32.480)
Um, but I think this is very difficult for people to know, normal people to
Elon Musk (06:36.320)
understand, they would not know what they're looking at.
Lex Fridman (06:39.460)
So it's almost an HMI challenge to the current things that are being
Elon Musk (06:42.660)
displayed is optimized for the general public understanding of
Lex Fridman (06:47.200)
what the system is capable of.
Elon Musk (06:48.720)
It's like, if you have no idea what, how computer vision works or anything,
Lex Fridman (06:51.600)
you can sort of look at the screen and see if the car knows what's going on.
Lex Fridman (06:55.740)
And then if you're, you know, if you're a development engineer or if you're,
Lex Fridman (06:59.320)
you know, if you're, if you have the development build like I do, then you
Elon Musk (07:02.720)
can see, uh, you know, all the debug information, but those would just be
Lex Fridman (07:07.560)
like total diverse to most people.
Elon Musk (07:11.200)
What's your view on how to best distribute effort.
Lex Fridman (07:14.200)
So there's three, I would say technical aspects of autopilot
Elon Musk (07:17.560)
that are really important.
Lex Fridman (07:18.800)
So it's the underlying algorithms, like the neural network architecture,
Elon Musk (07:22.000)
there's the data, so that the strain on, and then there's a hardware development.
Lex Fridman (07:26.000)
There may be others, but so look, algorithm, data, hardware, you don't, you
Elon Musk (07:32.400)
only have so much money, only have so much time, what do you think is the most
Lex Fridman (07:35.960)
important thing to, to, uh, allocate resources to, or do you see it as pretty
Lex Fridman (07:40.800)
evenly distributed between those three?
Lex Fridman (07:43.440)
We automatically get a fast amounts of data because all of our cars have eight
Elon Musk (07:51.040)
external facing cameras and radar, and usually 12 ultrasonic sensors, uh, GPS,
Lex Fridman (07:58.560)
obviously, um, and, uh, IMU.
Lex Fridman (08:02.920)
And so we basically have a fleet that has, uh, and we've got about 400,000
Lex Fridman (08:10.400)
cars on the road that have that level of data, I think you keep quite
Elon Musk (08:13.880)
close track of it actually.
Lex Fridman (08:14.840)
Yes.
Elon Musk (08:15.520)
Yeah.
Lex Fridman (08:15.800)
So we're, we're approaching half a million cars on the road that have the full sensor
Elon Musk (08:20.720)
suite.
Elon Musk (08:21.520)
Um, so this is, I'm, I'm not sure how many other cars on the road have the sensor
Elon Musk (08:27.720)
suite, but I would be surprised if it's more than 5,000, which means that we
Lex Fridman (08:32.400)
have 99% of all the data.
Lex Fridman (08:35.200)
So there's this huge inflow of data.
Lex Fridman (08:37.400)
Absolutely.
Elon Musk (08:37.920)
Massive inflow of data, and then we, it's, it's taken us about three years, but now
Lex Fridman (08:43.800)
we've finally developed our full self driving computer, which can process, uh,
Lex Fridman (08:51.160)
and in order of magnitude as much as the Nvidia system that we currently have in
Elon Musk (08:54.720)
the, in the cars, and it's really just a, to use it, you've unplugged the Nvidia
Elon Musk (08:59.000)
computer and plug the Tesla computer in and that's it.
Lex Fridman (09:01.600)
And it's, it's, uh, in fact, we're not even, we're still exploring the boundaries
Elon Musk (09:06.400)
of capabilities, uh, but we're able to run the cameras at full frame rate, full
Elon Musk (09:10.080)
resolution, uh, not even crop the images and it's still got headroom even on one
Elon Musk (09:16.600)
of the systems, the harder full self driving computer is really two computers,
Lex Fridman (09:21.240)
two systems on a chip that are fully redundant.
Lex Fridman (09:23.840)
So you could put a bolt through basically any part of that system and it still
Lex Fridman (09:27.320)
works.
Elon Musk (09:27.820)
The redundancy, are they perfect copies of each other or also it's purely for
Elon Musk (09:33.100)
redundancy as opposed to an argue machine kind of architecture where they're both
Elon Musk (09:37.140)
making decisions.
Lex Fridman (09:37.780)
This is purely for redundancy.
Elon Musk (09:39.540)
I think it would more like it's, if you have a twin engine aircraft, uh, commercial
Lex Fridman (09:43.620)
aircraft, the system will operate best if both systems are operating, but it's,
Elon Musk (09:51.180)
it's capable of operating safely on one.
Elon Musk (09:53.140)
So, but as it is right now, we can just run, we're, we haven't even hit the, the,
Elon Musk (09:59.020)
the edge of performance.
Lex Fridman (10:01.020)
So there's no need to actually distribute functionality across both SOCs.
Elon Musk (10:10.020)
We can actually just run a full duplicate on, on, on each one.
Lex Fridman (10:13.660)
Do you haven't really explored or hit the limit of this?
Elon Musk (10:17.100)
Not yet at the limiter.
Lex Fridman (10:18.220)
So the magic of deep learning is that it gets better with data.
Elon Musk (10:22.900)
You said there's a huge inflow of data, but the thing about driving the really
Lex Fridman (10:28.340)
valuable data to learn from is the edge cases.
Lex Fridman (10:32.260)
So how do you, I mean, I've, I've heard you talk somewhere about, uh, autopilot
Lex Fridman (10:39.100)
disengagements being an important moment of time to use.
Elon Musk (10:42.460)
Is there other edge cases where you can, you know, you can, you can, you can
Lex Fridman (10:46.660)
drive, is there other edge cases or perhaps can you speak to those edge cases?
Lex Fridman (10:53.060)
What aspects of that might be valuable or if you have other ideas, how to
Lex Fridman (10:56.900)
discover more and more and more edge cases in driving?
Elon Musk (11:00.780)
Well, there's a lot of things that are learned.
Lex Fridman (11:02.860)
There are certainly edge cases where I say somebody is on autopilot and they,
Elon Musk (11:06.980)
they take over and then, okay, that, that, that, that's a trigger that goes to our
Lex Fridman (11:12.580)
system that says, okay, did they take over for convenience or do they take
Elon Musk (11:16.220)
over because the autopilot wasn't working properly.
Elon Musk (11:19.380)
There's also like, let's say we're, we're trying to figure out what is the optimal
Elon Musk (11:23.700)
spline for traversing an intersection.
Lex Fridman (11:27.420)
Um, then then the ones where there are no interventions and are the right ones.
Lex Fridman (11:33.660)
So you then say, okay, when it looks like this, do the following.
Lex Fridman (11:38.300)
And then, and then you get the optimal spline for a complex, uh,
Elon Musk (11:42.060)
navigating a complex, uh, intersection.
Lex Fridman (11:44.660)
So that's for this.
Lex Fridman (11:46.260)
So there's kind of the common case you're trying to, uh, capture a huge amount of
Lex Fridman (11:50.780)
samples of a particular intersection, how, when things went right, and then
Elon Musk (11:54.500)
there's the edge case where, uh, as you said, not for convenience, but
Lex Fridman (11:59.420)
something didn't go exactly right.
Elon Musk (12:01.140)
Somebody took over, somebody asserted manual control from autopilot.
Lex Fridman (12:05.020)
And really like the way to look at this as view all input is error.
Elon Musk (12:08.900)
If the user had to do input, it does something all input is error.
Lex Fridman (12:12.980)
That's a powerful line.
Elon Musk (12:13.940)
That's a powerful line to think of it that way, because they may very well be
Lex Fridman (12:17.460)
error, but if you want to exit the highway, or if you want to, uh, it's
Elon Musk (12:21.380)
a navigation decision that all autopilot is not currently designed to do.
Lex Fridman (12:25.380)
Then the driver takes over.
Lex Fridman (12:27.540)
How do you know the difference?
Lex Fridman (12:28.380)
That's going to change with navigate an autopilot, which we were just
Elon Musk (12:31.180)
released and without still confirm.
Lex Fridman (12:33.820)
So the navigation, like lane change based, like a certain control in
Elon Musk (12:38.340)
order to change, do a lane change or exit a freeway or, or doing a highway
Lex Fridman (12:42.780)
under change, the vast majority of that will go away with, um, the
Elon Musk (12:47.580)
release that just went out.
Lex Fridman (12:48.900)
Yeah.
Lex Fridman (12:49.140)
So that, that I don't think people quite understand how big of a step that is.
Lex Fridman (12:54.580)
Yeah, they don't.
Lex Fridman (12:55.900)
So if you drive the car, then you do.
Lex Fridman (12:58.260)
So you still have to keep your hands on the steering wheel currently when
Elon Musk (13:00.980)
it does the automatic lane change.
Lex Fridman (13:03.420)
What are, so there's, there's these big leaps through the development of
Lex Fridman (13:07.780)
autopilot through its history and what stands out to you as the big leaps?
Lex Fridman (13:13.580)
I would say this one, navigate an autopilot without, uh, confirm
Elon Musk (13:18.580)
without having to confirm is a huge leap.
Lex Fridman (13:21.100)
It is a huge leap.
Elon Musk (13:22.500)
It also automatically overtakes low cars.
Lex Fridman (13:24.900)
So it's, it's both navigation, um, and seeking the fastest lane.
Lex Fridman (13:31.060)
So it'll, it'll, it'll, you know, overtake a slow cause, um, and exit the
Lex Fridman (13:36.420)
freeway and take highway interchanges.
Elon Musk (13:38.540)
And, and then, uh, we have, uh, traffic lights, uh, recognition, which
Lex Fridman (13:47.380)
introduced initially as a, as a warning.
Elon Musk (13:50.220)
I mean, on the development version that I'm driving, the car fully, fully
Lex Fridman (13:53.460)
stops and goes at traffic lights.
Lex Fridman (13:56.900)
So those are the steps, right?
Lex Fridman (13:58.500)
You've just mentioned something sort of inkling a step towards full autonomy.
Lex Fridman (14:02.220)
What would you say are the biggest technological roadblocks
Lex Fridman (14:06.860)
to full self driving?
Elon Musk (14:08.900)
Actually, I don't think, I think we just, the full self driving computer that we
Lex Fridman (14:11.860)
just, uh, that the Tesla, what we call the FSD computer, uh, that that's now in
Elon Musk (14:17.660)
production.
Elon Musk (14:20.540)
Uh, so if you order, uh, any model SRX or any model three that has the full self
Elon Musk (14:26.300)
driving package, you'll get the FSD computer.
Lex Fridman (14:29.700)
That, that was, that's important to have enough, uh, base computation, uh, then
Elon Musk (14:34.980)
refining the neural net and the control software, uh, which, but all of that can
Lex Fridman (14:39.820)
just be provided as an over there update.
Elon Musk (14:42.660)
The thing that's really profound and where I'll be emphasizing at the, uh, sort
Lex Fridman (14:47.460)
of what that investor day that we're having focused on autonomy is that the
Elon Musk (14:51.100)
cars currently being produced with the hardware currently being produced is
Elon Musk (14:55.820)
capable of full self driving, but capable is an interesting word because, um, like
Elon Musk (15:01.940)
the hardware is, and as we refine the software, the capabilities will increase
Elon Musk (15:07.980)
dramatically, um, and then the reliability will increase dramatically, and then it
Elon Musk (15:11.740)
will receive regulatory approval.
Lex Fridman (15:13.420)
So essentially buying a car today is an investment in the future.
Elon Musk (15:16.420)
You're essentially buying a car, you're buying the, I think the most profound
Elon Musk (15:21.580)
thing is that if you buy a Tesla today, I believe you are buying an appreciating
Elon Musk (15:26.860)
asset, not a depreciating asset.
Lex Fridman (15:30.140)
So that's a really important statement there because if hardware is capable
Elon Musk (15:33.940)
enough, that's the hard thing to upgrade usually.
Lex Fridman (15:37.260)
Exactly.
Lex Fridman (15:37.700)
So then the rest is a software problem.
Lex Fridman (15:40.820)
Yes.
Elon Musk (15:41.500)
Software has no marginal cost really.
Lex Fridman (15:44.940)
But what's your intuition on the software side?
Lex Fridman (15:48.420)
How hard are the remaining steps to, to get it to where, um, you know, uh, the,
Lex Fridman (15:57.500)
the experience, uh, not just the safety, but the full experience is something
Elon Musk (16:03.260)
that people would, uh, enjoy.
Lex Fridman (16:06.220)
Well, I think people enjoy it very much so on, on, on the highways.
Elon Musk (16:09.500)
It's, it's a total game changer for quality of life for using, you know,
Lex Fridman (16:15.180)
Tesla autopilot on the highways, uh, so it's really just extending that
Elon Musk (16:19.340)
functionality to city streets, adding in the traffic light recognition, uh,
Lex Fridman (16:26.220)
navigating complex intersections and, um, and then, uh, being able to navigate
Elon Musk (16:32.540)
complicated parking lots so the car can, uh, exit a parking space and come and
Lex Fridman (16:37.860)
find you, even if it's in a complete maze of a parking lot, um, and, and, and,
Lex Fridman (16:43.740)
and then if, and then you can just, it can just drop you off and find a
Lex Fridman (16:46.300)
parking spot by itself.
Elon Musk (16:48.940)
Yeah.
Lex Fridman (16:49.140)
In terms of enjoyability and something that people would, uh, would actually
Elon Musk (16:53.860)
find a lot of use from the parking lot is a, is a really, you know, it's, it's
Lex Fridman (16:58.300)
rich of annoyance when you have to do it manually.
Lex Fridman (17:00.700)
So there's a lot of benefit to be gained from automation there.
Lex Fridman (17:04.580)
So let me start injecting the human into this discussion a little bit.
Elon Musk (17:08.380)
Uh, so let's talk about, uh, the, the, the, the, the, the, the, the, the, the,
Lex Fridman (17:13.620)
about full autonomy.
Elon Musk (17:15.660)
If you look at the current level four vehicles being tested on
Lex Fridman (17:18.060)
road, like Waymo and so on, they're only technically autonomous.
Elon Musk (17:23.380)
They're really level two systems with just the different design philosophy,
Lex Fridman (17:28.860)
because there's always a safety driver in almost all cases and
Elon Musk (17:31.660)
they're monitoring the system.
Lex Fridman (17:33.060)
Right.
Lex Fridman (17:33.340)
Do you see Tesla's full self driving as still for a time to come requiring
Lex Fridman (17:42.580)
supervision of the human being.
Lex Fridman (17:44.780)
So it's capabilities are powerful enough to drive, but nevertheless requires
Lex Fridman (17:48.580)
the human to still be supervising, just like a safety driver is in a
Elon Musk (17:54.820)
other fully autonomous vehicles.
Lex Fridman (17:57.340)
I think it will require detecting hands on wheel for at least, uh, six months
Elon Musk (18:05.900)
or something like that from here.
Elon Musk (18:07.660)
It really is a question of like, from a regulatory standpoint, uh, what, how much
Elon Musk (18:15.060)
safer than a person does autopilot need to be for it to be okay to not monitor
Lex Fridman (18:20.900)
the car, you know, and, and this is a debate that one can have it.
Lex Fridman (18:25.460)
And then if you, but you need, you know, a large sample, a large amount of data.
Elon Musk (18:30.980)
Um, so you can prove with high confidence, statistically speaking, that the car is
Elon Musk (18:36.340)
dramatically safer than a person, um, and that adding in the person monitoring
Lex Fridman (18:40.940)
does not materially affect the safety.
Lex Fridman (18:44.060)
So it might need to be like two or 300% safer than a person.
Lex Fridman (18:48.300)
And how do you prove that incidents per mile incidents per mile crashes and
Elon Musk (18:53.460)
fatalities, fatalities would be a factor, but there, there are just not enough
Lex Fridman (18:58.100)
fatalities to be statistically significant at scale, but there are enough.
Elon Musk (19:03.060)
Crashes, you know, there are far more crashes than there are fatalities.
Lex Fridman (19:08.180)
So you can assess what is the probability of a crash that then there's another step
Elon Musk (19:14.460)
which probability of injury and probability of permanent injury, the
Lex Fridman (19:19.140)
probability of death, and all of those need to be a much better than a person,
Elon Musk (19:24.660)
uh, by at least perhaps 200%.
Lex Fridman (19:28.900)
And you think there's, uh, the ability to have a healthy discourse with the
Lex Fridman (19:33.500)
regulatory bodies on this topic?
Elon Musk (19:36.020)
I mean, there's no question that, um, but, um, regulators pay just disproportionate
Elon Musk (19:41.140)
amount of attention to that, which generates press.
Lex Fridman (19:44.700)
This is just an objective fact.
Elon Musk (19:46.420)
Um, and Tesla generates a lot of press.
Lex Fridman (19:49.260)
So the, you know, in the United States, this, I think almost, you know,
Elon Musk (19:55.660)
uh, in the United States, this, I think almost 40,000 automotive deaths per year.
Lex Fridman (20:01.820)
Uh, but if there are four in Tesla, they'll probably receive a thousand
Elon Musk (20:06.100)
times more press than anyone else.
Lex Fridman (20:08.780)
So the, the psychology of that is actually fascinating.
Elon Musk (20:11.460)
I don't think we'll have enough time to talk about that, but I have to talk to
Lex Fridman (20:14.820)
you about the human side of things.
Lex Fridman (20:16.980)
So myself and our team at MIT recently released the paper on functional
Lex Fridman (20:21.860)
vigilance of drivers while using autopilot.
Elon Musk (20:23.980)
This is work we've been doing since autopilot was first released publicly
Lex Fridman (20:28.580)
over three years ago, collecting video of driver faces and driver body.
Lex Fridman (20:34.020)
So I saw that you tweeted a quote from the abstract, so I can at least, uh,
Lex Fridman (20:40.980)
guess that you've glanced at it.
Elon Musk (20:42.820)
Yeah, I read it.
Lex Fridman (20:43.940)
Can I talk you through what we found?
Elon Musk (20:45.740)
Sure.
Lex Fridman (20:46.140)
Okay.
Lex Fridman (20:46.420)
So it appears that in the data that we've collected, that drivers are maintaining
Lex Fridman (20:53.620)
functional vigilance such that we're looking at 18,000 disengagement from
Elon Musk (20:57.260)
autopilot, 18,900 and annotating, were they able to take over control in a timely
Lex Fridman (21:04.420)
manner?
Lex Fridman (21:05.100)
So they were there present looking at the road, uh, to take over control.
Lex Fridman (21:09.500)
Okay.
Lex Fridman (21:09.860)
So this, uh, goes against what, what many would predict from the body of literature
Lex Fridman (21:15.500)
on vigilance with automation.
Elon Musk (21:18.060)
Now, the question is, do you think these results hold across the broader
Lex Fridman (21:22.260)
population?
Lex Fridman (21:23.300)
So ours is just a small subset.
Lex Fridman (21:25.780)
Do you think, uh, one of the criticism is that, you know, there's a small
Elon Musk (21:30.700)
minority of drivers that may be highly responsible where their vigilance
Lex Fridman (21:35.420)
decrement would increase with autopilot use?
Elon Musk (21:38.180)
I think this is all really going to be swept.
Lex Fridman (21:40.260)
I mean, the system's improving so much, so fast that this is going to be a mood
Elon Musk (21:46.660)
point very soon where vigilance is like, if something's many times safer than a
Lex Fridman (21:55.860)
person, then adding a person, uh, does the, the, the effect on safety is, is
Elon Musk (22:01.620)
limited.
Lex Fridman (22:02.100)
Um, and in fact, uh, it could be negative.
Elon Musk (22:09.580)
That's really interesting.
Lex Fridman (22:10.420)
So the, uh, the, so the fact that a human may, some percent of the population may,
Elon Musk (22:16.660)
uh, exhibit a vigilance decrement will not affect overall statistics numbers of
Lex Fridman (22:20.980)
safety.
Elon Musk (22:21.380)
No, in fact, I think it will become, uh, very, very quickly, maybe even towards
Lex Fridman (22:27.460)
the end of this year, but I'd say I'd be shocked if it's not next year.
Elon Musk (22:30.860)
At the latest, that, um, having the person, having a human intervene will
Lex Fridman (22:35.300)
decrease safety decrease.
Elon Musk (22:38.980)
It's like, imagine if you're an elevator and it used to be that there were
Lex Fridman (22:42.220)
elevator operators, um, and, and you couldn't go on an elevator by yourself
Lex Fridman (22:46.780)
and work the lever to move between floors.
Lex Fridman (22:49.940)
Um, and now, uh, nobody wants it an elevator operator because the automated
Elon Musk (22:56.900)
elevator that stops the floors is much safer than the elevator operator.
Lex Fridman (23:01.940)
And in fact, it would be quite dangerous to have someone with a lever that can
Elon Musk (23:05.420)
move the elevator between floors.
Lex Fridman (23:07.740)
So that's a, that's a really powerful statement and really interesting one.
Lex Fridman (23:12.500)
But I also have to ask from a user experience and from a safety perspective,
Lex Fridman (23:16.620)
one of the passions for me algorithmically is a camera based detection of, uh,
Elon Musk (23:22.580)
of just sensing the human, but detecting what the driver is looking at, cognitive
Lex Fridman (23:26.380)
load, body pose on the computer vision side, that's a fascinating problem.
Lex Fridman (23:30.140)
But do you, and there's many in industry believe you have to have
Lex Fridman (23:33.620)
camera based driver monitoring.
Lex Fridman (23:35.540)
Do you think there could be benefit gained from driver monitoring?
Lex Fridman (23:39.700)
If you have a system that's, that's at, that's at or below a human level
Elon Musk (23:44.660)
reliability, then driver monitoring makes sense.
Lex Fridman (23:48.220)
But if your system is dramatically better, more likely to be
Elon Musk (23:51.540)
better, more liable than, than a human, then drive monitoring monitoring
Lex Fridman (23:55.780)
is not just not help much.
Elon Musk (23:59.420)
And, uh, like I said, you, you, just like, as an, you wouldn't want someone
Lex Fridman (24:03.500)
into like, you wouldn't want someone in the elevator, if you're in an elevator,
Lex Fridman (24:06.580)
do you really want someone with a big lever, some, some random person
Lex Fridman (24:09.780)
operating the elevator between floors?
Elon Musk (24:12.940)
I wouldn't trust that or rather have the buttons.
Lex Fridman (24:17.420)
Okay.
Elon Musk (24:17.860)
You're optimistic about the pace of improvement of the system that from
Lex Fridman (24:21.900)
what you've seen with the full self driving car computer, the rate
Elon Musk (24:25.780)
of improvement is exponential.
Lex Fridman (24:28.300)
So one of the other very interesting design choices early on that connects
Elon Musk (24:32.900)
to this is the operational design domain of autopilot.
Lex Fridman (24:38.020)
So where autopilot is able to be turned on the, so contrast another vehicle
Elon Musk (24:44.820)
system that we're studying is the Cadillac SuperCrew system.
Lex Fridman (24:48.860)
That's in terms of ODD, very constrained to particular kinds of highways, well
Elon Musk (24:53.620)
mapped, tested, but it's much narrower than the ODD of Tesla vehicles.
Lex Fridman (24:58.940)
What's there's, there's pros and...
Elon Musk (25:00.660)
It's like ADD.
Lex Fridman (25:02.580)
Yeah.
Elon Musk (25:04.300)
That's good.
Lex Fridman (25:04.740)
That's a, that's a good line.
Elon Musk (25:06.660)
Uh, what was the design decision, uh, what, in that different philosophy
Lex Fridman (25:13.060)
of thinking where there's pros and cons, what we see with, uh, a wide ODD
Elon Musk (25:18.820)
is drive Tesla drivers are able to explore more the limitations of the
Lex Fridman (25:22.900)
system, at least early on, and they understand together with the instrument
Elon Musk (25:26.860)
cluster display, they start to understand what are the capabilities.
Lex Fridman (25:30.180)
So that's a benefit.
Elon Musk (25:31.740)
The con is you go, you're letting drivers use it basically anywhere.
Lex Fridman (25:38.180)
So anyway, that could detect lanes with confidence.
Elon Musk (25:41.100)
Was there a philosophy, uh, design decisions that were challenging
Lex Fridman (25:46.020)
that were being made there or from the very beginning, was that, uh,
Lex Fridman (25:51.300)
done on purpose with intent?
Lex Fridman (25:54.100)
Well, I mean, I think it's frankly, it's pretty crazy giving it, letting people
Elon Musk (25:57.380)
drive a two ton death machine manually.
Lex Fridman (26:01.340)
Uh, that's crazy.
Elon Musk (26:03.580)
Like, like in the future of people who are like, I can't believe anyone was
Lex Fridman (26:07.740)
just allowed to drive for one of these two ton death machines and they
Elon Musk (26:12.780)
just drive wherever they wanted.
Lex Fridman (26:14.100)
Just like elevators.
Elon Musk (26:14.980)
He was like, move the elevator with that lever, wherever you want.
Lex Fridman (26:17.780)
It can stop at halfway between floors if you want.
Elon Musk (26:22.060)
It's pretty crazy.
Lex Fridman (26:24.140)
So it's going to seem like a mad thing in the future that people were driving cars.
Lex Fridman (26:32.500)
So I have a bunch of questions about the human psychology, about behavior and so
Lex Fridman (26:36.380)
on that would become that because, uh, you have faith in the AI system, uh, not
Elon Musk (26:46.140)
faith, but, uh, the, both on the hardware side and the deep learning approach of
Lex Fridman (26:51.180)
learning from data will make it just far safer than humans.
Elon Musk (26:55.260)
Yeah, exactly.
Lex Fridman (26:56.900)
Recently, there are a few hackers who, uh, tricked autopilot to act in
Elon Musk (27:00.780)
unexpected ways with adversarial examples.
Lex Fridman (27:03.020)
So we all know that neural network systems are very sensitive to minor
Elon Musk (27:06.900)
disturbances to these adversarial examples on input.
Lex Fridman (27:10.420)
Do you think it's possible to defend against something like this for the
Lex Fridman (27:13.700)
broader, for the industry?
Lex Fridman (27:15.140)
Sure.
Lex Fridman (27:15.860)
So can you elaborate on the, on the confidence behind that answer?
Lex Fridman (27:22.900)
Um, well the, you know, neural net is just like a basic bunch of matrix math.
Elon Musk (27:27.820)
Or you have to be like a very sophisticated, somebody who really
Lex Fridman (27:31.620)
understands neural nets and like basically reverse engineer how the matrix
Elon Musk (27:36.620)
is being built and then create a little thing that's just exactly, um, causes
Lex Fridman (27:42.700)
the matrix math to be slightly off.
Lex Fridman (27:44.740)
But it's very easy to then block it, block that by, by having basically
Lex Fridman (27:49.540)
anti negative recognition.
Elon Musk (27:51.100)
It's like if you, if the system sees something that looks like a matrix hack,
Lex Fridman (27:55.460)
uh, exclude it, so it's such an easy thing to do.
Lex Fridman (28:01.860)
So learn both on the, the valid data and the invalid data.
Lex Fridman (28:05.340)
So basically learn on the adversarial examples to be able to exclude them.
Elon Musk (28:08.980)
Yeah.
Lex Fridman (28:09.480)
Like you basically want to both know what is, what is a car and
Lex Fridman (28:13.020)
what is definitely not a car.
Lex Fridman (28:15.260)
And you train for this is a car and this is definitely not a car.
Elon Musk (28:18.340)
Those are two different things.
Lex Fridman (28:20.180)
People have no idea neural nets really.
Elon Musk (28:23.020)
They probably think neural nets are both like, you know, fishing net only.
Lex Fridman (28:28.460)
So as you know, so taking a step beyond just Tesla and autopilot, uh, current
Elon Musk (28:36.260)
deep learning approaches still seem in some ways to be far from general
Lex Fridman (28:42.660)
intelligence systems.
Lex Fridman (28:43.940)
Do you think the current approaches will take us to general intelligence or do
Lex Fridman (28:49.820)
totally new ideas need to be invented?
Elon Musk (28:54.500)
I think we're missing a few key ideas for general intelligence, general artificial
Lex Fridman (28:59.740)
general intelligence, but it's going to be upon us very quickly.
Lex Fridman (29:07.700)
And then we'll need to figure out what shall we do if we even have that choice?
Lex Fridman (29:14.580)
But it's amazing how people can't differentiate between say the narrow
Elon Musk (29:18.700)
AI that, you know, allows a car to figure out what a lane line is and, and, and,
Lex Fridman (29:24.140)
you know, and navigate streets versus general intelligence.
Elon Musk (29:29.420)
Like these are just very different things.
Lex Fridman (29:32.020)
Like your toaster and your computer are both machines, but one's much
Elon Musk (29:35.340)
more sophisticated than another.
Lex Fridman (29:37.460)
You're confident with Tesla.
Elon Musk (29:39.340)
You can create the world's best toaster.
Lex Fridman (29:42.580)
The world's best toaster.
Elon Musk (29:43.420)
Yes.
Lex Fridman (29:43.920)
The world's best toaster. Yes. The world's best self driving. I'm, I, yes.
Elon Musk (29:52.240)
To me right now, this seems game set match.
Lex Fridman (29:54.880)
I don't, I mean, that sounds, I don't want to be complacent or overconfident,
Lex Fridman (29:57.760)
but that's what it appears.
Lex Fridman (29:58.880)
That is just literally what it, how it appears right now.
Elon Musk (30:02.600)
I could be wrong, but it appears to be the case that Tesla is vastly ahead of
Lex Fridman (30:08.960)
everyone.
Lex Fridman (30:09.480)
Do you think we will ever create an AI system that we can love and loves us back
Lex Fridman (30:14.960)
in a deep, meaningful way?
Elon Musk (30:15.960)
Like in the movie, her, I think AI will be capable of convincing you to fall in
Lex Fridman (30:22.360)
love with it very well.
Lex Fridman (30:24.360)
And that's different than us humans.
Lex Fridman (30:27.840)
You know, we start getting into a metaphysical question of like, do emotions
Lex Fridman (30:31.560)
and thoughts exist in a different realm than the physical?
Lex Fridman (30:34.160)
And maybe they do.
Elon Musk (30:35.040)
Maybe they don't.
Lex Fridman (30:35.600)
I don't know.
Lex Fridman (30:36.100)
But from a physics standpoint, I tend to think of things, you know, like physics
Lex Fridman (30:42.740)
was my main sort of training and from a physics standpoint, essentially, if it
Elon Musk (30:50.100)
loves you in a way that is, that you can't tell whether it's real or not, it is
Lex Fridman (30:53.940)
real.
Elon Musk (30:55.940)
That's a physics view of love.
Lex Fridman (30:57.380)
Yeah.
Elon Musk (30:59.180)
If there's no, if you cannot just, if you cannot prove that it does not, if there's
Elon Musk (31:04.780)
no, if there's no test that you can apply that would make it, allow you to tell the
Elon Musk (31:14.900)
difference, then there is no difference.
Lex Fridman (31:17.340)
Right.
Lex Fridman (31:17.860)
And it's similar to seeing our world as simulation.
Lex Fridman (31:21.420)
There may not be a test to tell the difference between what the real world
Lex Fridman (31:24.900)
and the simulation, and therefore from a physics perspective, it might as well be
Lex Fridman (31:28.780)
the same thing.
Elon Musk (31:29.540)
Yes.
Lex Fridman (31:30.540)
And there may be ways to test whether it's a simulation.
Elon Musk (31:33.220)
There might be, I'm not saying there aren't, but you could certainly imagine
Elon Musk (31:36.420)
that a simulation could correct that once an entity in the simulation found a way
Elon Musk (31:40.900)
to detect the simulation, it could either restart, you know, pause the simulation,
Elon Musk (31:46.620)
start a new simulation, or do one of many other things that then corrects for that
Elon Musk (31:50.340)
error.
Lex Fridman (31:52.380)
So when maybe you or somebody else creates an AGI system and you get to ask
Lex Fridman (32:00.260)
her one question, what would that question be?
Lex Fridman (32:16.260)
What's outside the simulation?
Elon Musk (32:20.900)
Elon, thank you so much for talking today.
Lex Fridman (32:22.660)
It was a pleasure.
Elon Musk (32:23.500)
All right.
Lex Fridman (32:24.000)
Thank you.
🔗 相关节目