从 AV 1.0 到 2.0：端到端神经网络在自动驾驶中的崛起

今天我们邀请到 Wave 的 CEO Alex Kendall，讨论从软件 1.0 到 2.0，即从经典机器学习到端到端神经网络在自动驾驶中的转变。Wave 向汽车 OEM 厂商销售自动驾驶系统，类似于特斯拉的 FSD，但用于非特斯拉汽车。全球主要汽车制造商如日产选择 Wave 来驱动其自动驾驶系统。Alex 在 2017 年创立 Wave，当时大多数自动驾驶软件栈是庞大的手工编码 C++ 代码库，覆盖每一个边缘情况，比如绕行双排停车。Alex 从一开始就押注端到端神经网络方法，以及使用合成数据和世界模型作为泛化和规模化的终极路径。如今，这一架构正在重塑自动驾驶和所有物理 AI，包括机器人技术。享受节目。Alex，感谢你参加我们的节目。

Today we're talking with Alex Kendall, CEO of Wave about the shift from software 1.0 to 2.0 or from classical machine learning to end-to-end neural networks in autonomous driving. Wave sells an autonomous driving stack to auto OEMs similar to Tesla FSD, but for non-Tesla automobiles. Major car manufacturers globally like Nissan are choosing Wave to power their AV stacks. Alex started Wave back in 2017 when most self-driving software stacks were massive handcoded C++ code bases covering every possible edge case like navigating around double parked cars. Alex bet the farm from the beginning on an end-to-end neural net approach to self-driving and on the use of synthetic data and world models as the ultimate path to generalization and scaling. Today that architecture is reshaping AV and all of physical AI including robotics. Enjoy the show. Alex, thanks for joining us on the show.

嗨，Pat。嗨，Sonia。

Hey Pat. Hey Sonia.

贵公司非常特别的一点是，它代表了 AV 2.0，即一种新的架构方法，我认为它已被证明优于人们多年来苦苦钻研的 AV 1.0 方法。我们能否先定义什么是 AV 1.0？什么是 AV 2.0？

One of the things that is very special about your company is that it sort of typifies AV 2.0, meaning a new architectural approach that I think is kind of demonstrated to be superior to the AV 1.0 approach that people toiled with for so many years. Can we just start by defining what was AV 1.0? What is AV 2.0?

当然。我们在 2017 年创立公司时，种子轮融资的推介完全围绕当时的经典机器人方法：将感知、规划、地图、控制等自动驾驶问题分解成多个组件，并大量手工设计。我们的观点是，我们认为机器人的未来不是靠手工设计、依赖高精地图等基础设施的系统，而是拥有车载智能、能自主决策的智能机器。当然，我们知道的构建 AI 系统的最佳方法是端到端深度学习。所以在过去 10 年里，我们一直在推广一种下一代方法——AV 2.0，用一个端到端神经网络取代那个栈。当然，今天这看起来可能更明显，但多年来一直是逆向思维。不过我认为今天再做这种基本区分可能不公平，因为任何有见识的人都会在栈的各个部分使用深度学习。但在更主流的自动驾驶解决方案中，你看到的是深度学习用于感知，可能每个组件都用，但仍然有很多手工接口，大量依赖高精地图基础设施，或许还依赖大量硬件。所以我们的解决方案仍然有所进步，但今天不仅仅是端到端网络，我们开始谈论基础模型。我们开始谈论更通用的智能，不仅能理解如何驾驶那辆车，还能理解不同传感器架构、不同用例的多种车辆。所以归根结底，我们如何构建最智能的机器人，使其能够规模化，而不需要繁重的基础设施。

For sure. When we started the company in 2017, the opening pitch in our seed deck was all about the classical robotics approach at the time was to take a perception, planning, mapping, control, essentially break down the autonomy problem into a bunch of different components and largely hand-engineer them. And our pitch was that okay, we think that the future of robotics is not going to be a system that's hand-engineered to drive with a lot of infrastructure like high-definition maps but instead we thought that the future of robots would be intelligent machines that have the onboard intelligence to make their own decisions. And of course the best way we know how to build an AI system is with end-to-end deep learning. So for the last 10 years we've been promoting an approach, a next generation approach, AV 2.0, that replaces that stack with one end-to-end neural network. Now, of course, that may seem more obvious today, but it has been contrarian for many, many years. But I think today it's maybe unfair to make that basic distinction because of course, anyone who's worth a grain of salt will use deep learning in various parts of the stack. But what you see in more incumbent solutions to autonomous driving is of course deep learning for perception and maybe for each different component but still a lot of hand interfaces, still a lot of infrastructure on high-definition maps and perhaps reliance on a lot of hardware. So our solution is still somewhat moved on but today rather than just being an end-to-end network, today we start to talk about foundation models. We start to talk about more of a general purpose intelligence, one that can understand not just how to drive that car but many cars with different sensor architectures with different use cases. And so really it all boils down to how do we build the most intelligent robot that can scale without needing onerous infrastructure.

所以 Wave 是传感器输入、运动输出，中间是一个巨大的神经网络。

So Wave is sensor inputs, motion output, gigantic neural net in the middle.

没错。从非常简单的层面来说。但一些有趣的事情可能与我们在大型语言模型中听到的故事不同，自动驾驶当然有一些有趣的新因素。一个是安全性。我们需要确保系统是安全设计的。这意味着我们不能只是输入更多数据，希望幻觉消失。我们需要设计一个架构，它仍然是端到端数据驱动的，但既功能安全，又能构建一个稳健的行为安全案例。这带来了一些有趣的架构挑战，当然我们还需要在机器人上、在车辆上实时运行。所以处理车载计算和车载传感器的限制是一个有趣的挑战。但是的，这与我们在机器人领域看到的叙事相同，也与其他 AI 领域如语言或游戏智能体一样。端到端数据学习的解决方案正在超越任何手工编码的东西。我们很兴奋能率先在机器人和自动驾驶领域推动这一叙事。

That's right. At a very simple level. But some of the interesting things you see that are maybe different from the story we've all heard with large language models is with autonomous driving of course there are some interesting new factors. One is of course safety. The system we need to make sure is safe by design. And what that means is that we can't just pump more data in and hope that hallucinations go away. But we need to design an architecture that is still end-to-end data-driven but is both functionally safe and we can build a robust behavioral safety case. So that introduces some interesting architectural challenges and then of course we also need to run real time on board a robot, on board a vehicle. And so dealing with the onboard compute and onboard sensor limitations make it an interesting challenge. But yes, it's the same narrative we're seeing playing out in robotics that we've seen play out in all these other AI fields like language or gameplay agents. It's that an end-to-end data-learned solution is outcompeting anything we can handcode. And what we're excited to be pioneering is that exact same narrative here in robotics and autonomous vehicles.

当你们在 2017 年起步时，这是一种非常逆向的方法，业内有人说‘这行不通，因为……’他们是怎么说完这句话的？

And when you guys started this in 2017 and it was a very contrarian approach when people from the industry said, 'Well, that'll never work because...' How did they finish that sentence?

我能数出几百种这样的说法。典型的论点是：看，它不安全。它不可解释。无法理解它在做什么。甚至简单地说，这说不通。我们没听说过这种 AI 东西。我认为 5 或 10 年前，说端到端深度学习不可解释可能还算合理。但今天我不这么认为。我认为今天我们有很多很好的工具来理解和回应这些深度学习系统推理方式的见解。但更重要的是，我认为如果你有雄心构建任何智能机器，认为你能构建一个复杂的智能机器并让它严格可解释，以至于你能指出一行代码或一个东西因果导致了结果，那是天真的。智能机器的美妙之处在于它们如此奇妙地复杂，我认为我们不仅要设计它们，还要理解它们的方式是通过数据驱动的结构。

I could count hundreds of those footings. Typical arguments were: look, it's not safe. It's not interpretable. Can't understand what it's doing. Or even simply, it doesn't make sense. We haven't heard of this AI thing. And look, I think 5 or 10 years ago, it was probably reasonable to say end-to-end deep learning wasn't interpretable. But I don't think that's true today. I think today we have a lot of really great tools for understanding and responding to insights about the way these deep learning systems reason. But moreover, I think if you have the ambition to build any intelligent machine, I think it's naive to think you can build a complex intelligent machine and actually make it strictly interpretable to the point where you can point to a single line of code or a single thing that causally made the outcome occur. The beauty of intelligent machines is that they are so wonderfully complex and I think the way that we're going to not just design them but understand them is through a data-driven structure.

你能详细说说 AV 1.0 栈的前后对比吗？那些系统有数百万甚至数十亿行代码，而今天的 2.0 系统呢？变化有多快？因为我的感觉是，深度学习、大型神经网络进入实体经济，比人们可能意识到的要新得多。

Can you say more about the before and after of the AV 1.0 stack and the millions or billions of lines of code that go into those systems versus the 2.0 systems today? And how quickly is that changing? Because my sense is that deep learning, large neural nets hitting the physical economy is a much more recent phenomenon than people might appreciate.

嗯，尤其是当你想到这些系统的分发和部署路径时。汽车行业已经经历了一场巨变，推出了软件定义的汽车和合适的硬件来实现自动驾驶。一个常见的争论点是，应该只用摄像头，还是摄像头加雷达和激光雷达。明确我们的立场：我们希望构建一个能理解各种传感器架构的 AI。有时纯摄像头方案是合理的，有时则需要摄像头、雷达和激光雷达。我们在所有这些组合上训练我们的具身 AI 模型，数据来源非常多样。我们刚才开的车就是纯摄像头方案。我们还有其他与合作伙伴合作的车辆，配备了雷达和激光雷达，各有不同的权衡。更广泛地说，全球顶级制造商量产的车都配备了 GPU、环绕摄像头、环绕雷达，有时还有前置激光雷达。美妙之处在于，现在有机会让这个 AI 惠及全世界的人们。这种软件定义的基础设施正在汽车领域发生，可能在其他机器人垂直领域还没有达到同样的程度，但我相信市场也会朝这个方向发展。总的来说，以可扩展的方式拥有合适的计算基础设施，并向 AI 开放这些平台，我认为这才是真正让这一切成为可能的原因，而这在过去几年里已经达到了一个临界点。

Well, especially when you think about the path to distribution and deploying these systems. The automotive industry has gone through a seismic shift in bringing out software-defined vehicles and the right hardware to make them drive. A common point of debate is whether to use camera-only or camera, radar, and lidar. To be clear on our position, we want to build an AI that can understand all kinds of sensor architectures. Sometimes a camera-only solution makes sense, sometimes camera, radar, and lidar. We train our embodied AI model on all those permutations from diverse data sources. The car we just drove is a camera-only stack. We have other cars with partners that have radar and lidar, with different trade-offs. More generally, mass-produced cars from the best manufacturers have a GPU, surround cameras, surround radar, and sometimes a front lidar. The beauty is that there's now an opportunity for this AI to benefit people worldwide. This software-defined infrastructure is happening in automotive, perhaps not yet to the same degree in other robotics verticals, but the market will move that way. Having the right compute infrastructure in a scalable way and opening up platforms to AI is what's making this possible, and it's gone through a tipping point in the last couple of years.

你的 AV2.0 观点在过去两三年里从逆向变成了共识。是 FSD12 促成的吗？还是这个心态是什么时候开始转变的？

Your perspective of AV2.0 has flipped from contrarian to consensus in the last two or three years. Was it FSD12 that did it, or when did that mindset start to shift?

我怀念那些逆向的日子。但即使在今天，我仍然看到很多人说我们需要带有硬约束或安全保证的端到端 AI。有些人相信结合基于规则的堆栈和端到端学习堆栈的混合方法是可行的，但这种方法往往得到的是两方面的缺点，或者增加成本和复杂性。所以我仍然认为市场上存在分化：一部分人正在快速推进，另一部分人则需要迎头赶上。当然，我们要归功于深度学习，特别是大语言模型的突破，它们让这项技术变得改变世界并成为主流。它们启发了世界，打开了市场的好奇心。但同样，我们在 Wayve 所做的：一年前我们还在伦敦市中心驾驶，那是一个非结构化、极其复杂和动态的城市，我们的 AI 学会了平稳、安全、可靠地导航。在过去一年里，我们把它带到了高速公路、欧洲、日本、北美。我们的车上周在纽约市行驶。把它带到全球，带到不同制造商的车辆上，并展示出类似产品的体验，这确实在世界范围内激发了很多灵感。

I miss the contrarian days. But even today, I still see many folks saying we need end-to-end AI with hard constraints or safety guarantees. Some believe a hybrid approach combining a rules-based stack and an end-to-end learned stack is the way to go, but often these approaches get the worst of both worlds or add cost and complexity. So I still think there's a distribution in the market between those leaning in and moving fast and those that have some catching up to do. Of course, we credit the breakthroughs in deep learning, especially large language models, for making this world-changing and mainstream. They've inspired the world and opened up the market's mind. But also, what we've been doing at Wayve: a year ago we were driving in central London, an unstructured, complex, and dynamic city that our AI learned to navigate smoothly, safely, and reliably. In the last year, we've taken it to highways, Europe, Japan, North America. Our cars were in New York City last week. Bringing it global, to different manufacturers' vehicles, and showing a product-like experience has really opened up inspiration worldwide.

为什么你能够在全球数百个城市推出，而一些 AV1.0 公司却需要构建高清地图？请谈谈差异，以及技术差异如何导致机器学习和部署方式的不同。

Why is it that you're able to launch in hundreds of cities worldwide, while some AV1.0 companies need to build HD maps? Say a word on the difference and how technical differences lead to differences in how the machine learns and how you roll out.

自动驾驶完全关乎泛化。泛化意味着能够推理或理解你从未见过的东西。每次开车，你都会看到新事物。今天我们看到一个道路工人在人行横道前铺地毯，但没有走出来，我们必须推理能否在不礼让的情况下通过。你永远无法在训练数据中看到所有经历。所以你必须能够推理并泛化到未见过的场景，才能在全球范围内安全且有用。这激励了我们的整个方法。无论制造商给我们一辆车，几个月内我们就能让它上路。今年 9 月，我们在东京与日产一起向媒体展示了一辆车。仅仅四个月前，我们才第一次在东京驾驶并拿到这辆车。四个月后，媒体就驾驶了它，在一个新的国家和车辆中体验。这表明我们的 AI 能够泛化。它是在来自全球的多样化数据、多样化的传感器组和车辆上训练的，因此它能理解那辆车的新传感器分布以及在东京市中心驾驶的复杂性。这是泛化的绝佳展示。如果你构建一个垂直整合的机器人解决方案，你可以深入。但我们的目标是成为全球所有最佳车队和制造商的具身 AI 基础模型。要做到这一点，除非我们想为每个应用构建单独的神经网络而让公司超负荷，否则我们需要泛化。我们需要将成本摊销到一个大型智能上，并快速适应客户关心的每个应用。这就是我们正在推动的。

Autonomous driving is all about generalization. Generalization means being able to reason about or understand something you've never seen before. Every time you drive, you see something new. Today we saw a road worker rolling out carpet in front of a pedestrian crossing but not stepping out, and we had to reason whether we could pass without yielding. You never see every experience in your training data. So you must be able to reason and generalize to unseen situations to be safe and useful worldwide. That motivates our entire approach. Whether a manufacturer gives us a vehicle, within a couple of months we can drive it on the road. In September this year, we unveiled a vehicle to media with Nissan in Tokyo. Just four months earlier was the first time we drove in Tokyo and got hands on that vehicle. Four months later, media drove it and experienced it in a new country and vehicle. That shows our AI can generalize. It's trained on diverse data from around the world, diverse sensor sets, and vehicles, so it can understand that vehicle's new sensor distribution and the complexity of driving in central Tokyo. That's a great demonstration of generalization. If you build a vertically integrated robotic solution, you can go deep. But our ambition is to be the embodied AI foundation model for all the best fleets and manufacturers worldwide. To do that, unless we want to overload the company by building a separate neural network for each application, we need to generalize. We need to amortize our cost over one large intelligence and quickly adapt to each application our customers care about. That's what we're pushing.

你提到了推理，比如模型如何推理，作为建筑工人，我现在该做什么？在 LLM 领域，推理是一个独立的轨道，有很多扩展推理时算力的技术。你是刻意训练模型去推理，还是说这是模型的涌现行为？请多谈谈你所说的推理是什么意思。

You mentioned reasoning in terms of how the model is reasoning through, you know, as a construction worker, what do I do now? In the LLM world, reasoning is its own separate track with lots of scaling inference-time compute techniques. Are you deliberately training your models to reason, or is it an emergent property? Say more about what you mean by reasoning.

是的，我认为物理世界中的推理可以很好地用世界模型来表达。2018 年，我们将第一个世界模型方法上路测试。那是一个非常小的 10 万参数神经网络，能够模拟我们前方道路的 30x3 像素图像。但我们能够将其用作内部模拟器，训练基于模型的强化学习算法。如果你想了解这段历史，有一篇有趣的博客文章。快进到今天，我们开发了 GIA。它是一个完整的生成式世界模型，能够模拟多个摄像头和传感器，在非常丰富多样的环境中运行。你可以控制它，并提示其中的不同智能体或场景。这是推理的一个例子，我们可以训练模拟世界如何运作以及接下来会发生什么的能力。当这种表示被带到道路上时，你会得到一些非常好的涌现行为。比如今天，当我们看到车辆在无保护转弯时，你会看到汽车向前挪动，直到它能自己看清，然后完成转弯。或者在伦敦有雾时，你会看到汽车减速，并根据它能推理的情况行驶。通过以这种理解水平进行训练，它产生了那种涌现行为，帮助它真正理解特别复杂的多智能体场景。我认为这是实现安全平稳自动驾驶的关键。

We are, and I think reasoning in the physical world can be really well expressed as a world model. In 2018, we put our very first world model approach on the road. It was a very small 100,000-parameter neural network that could simulate a 30x3 pixel image of a road in front of us. But we were able to use it as this internal simulator to train a model-based reinforcement learning algorithm. There's a fun blog post if you want to see the history on that. Fast forward to today, and we've developed GIA. It's a full generative world model that's able to simulate multiple cameras and sensors in very rich and diverse environments. You can control it and prompt the different agents or scenes in it. That's an example of reasoning where we can train the ability to simulate how the world works and what's going to happen next. What happens when you bring this kind of representation on the road is you get some really nice emergent behavior. Like today, when we saw driving around unprotected turns, you saw the car nudge forward until it could see for itself and then completed the turn. Or when it's foggy in London, you see the car slow down and drive to what it can reason about. By training it with that level of understanding, it gives that level of emergent behavior that helps it really understand particularly complex multi-agent scenarios. I think that's key for getting safe and smooth autonomous driving.

所以世界模型确实是教会模型如何推理新场景的关键。

So the world models are really key to teaching the model how to reason through new scenarios.

完全正确。

100%.

你之前提到了数据的多样性。请谈谈这些数据都来自哪里。数据量变得非常庞大，因为与语言领域或图像领域不同，当我们处理一辆典型的自动驾驶汽车时，它有十几个百万像素摄像头、雷达，可能还有激光雷达，当你把这些数据汇总起来，很快就会达到几十或几百 PB。所以你需要训练的数据量非常巨大，但多样性才是关键。你们是如何解决多样性问题的？

You mentioned earlier the diversity of your data. Say a word about where all the data comes from. It's becoming an enormous amount of data because, unlike the language domain or image domain, when we're dealing with a typical self-driving car that has a dozen multiple-megapixel cameras, radar, maybe a LiDAR, when you aggregate that up, it's very quickly tens or hundreds of petabytes of data. So it's an enormous amount of data you have to train on, but it's the diversity that's really key. How have you solved for diversity?

我们通过两种方式解决了多样性问题。第一种是成为行业内的可信合作伙伴，聚合来自许多不同来源的数据，从行车记录仪到车队，再到制造商和机器人操作员。第二种是能够过滤并真正理解数据。在这方面，我们非常努力地开发了不同的无监督学习技术，以便聚类并发现异常或异常体验。当然，还要找出系统表现不佳的场景，并据此推动学习课程。但没错，今天我们学习自多样化的车辆、多样化的传感器架构和不同的国家。这确实是推动泛化水平的关键因素之一。

We've solved for diversity in two ways. The first one is by becoming a trusted partner across the industry and aggregating data across many different sources, from dash cams to fleets to manufacturers to robot operators. The second one is being able to filter and really understand the data. Here we've really worked hard to develop different unsupervised learning techniques to be able to cluster and find unusual or anomaly experiences. And of course, find the scenarios that our system is performing poorly at and then drive the learning curriculum on those. But yeah, today we learn from a diverse set of vehicles, a diverse set of sensor architectures, of countries. That's really one of the key things that drives the level of generalization.

世界模型和模拟数据的增长是否意味着你不需要那么多实际的道路里程？

Does the increased growth of world models and simulated data mean that you just don't need as many actual on-road miles?

我认为这个问题有两面。一方面，效率确实很重要，但另一方面，你不能只依赖学习效率。在极限情况下，如果我们采用当前的方法并直接扩展，我确信它会产生通用的 L5 级驾驶，在极限情况下，如果你有无限的训练数据，这实际上只是一个带有一些先验经验的查找数据表。但这不是经济上或技术上可行的。所以问题是，你如何训练出最高数据效率的系统？因为我认为效率不仅会降低成本，还会加快上市时间并提高智能水平。效率来自许多不同的因素。最重要的是你制定的数据课程，然后是学习算法。你如何放大已有的学习？我认为世界模型是一个很好的机会。它们生成合成数据和合成理解，这不会取代真实世界数据，而是以新的方式重新组合并放大它。它让你引入有趣的见解。我认为这类方法可以真正提高数据效率。但总的来说，我认为在资源限制下工作迫使我们的团队开发了许多创新。我还要提到工作流程，因为在传统机器人学中，当你调整参数或算法或设计几何地图时，有非常成熟的文化和工作流程。我们的团队，当有 50 名模型开发人员在一个主要生产模型上工作时，或者当我们有一个需要理解和内省的端到端网络时，甚至当我们将这些系统部署到仿真或道路上并获取反馈时，我们从零开始建立了整个文化。Wayve 的文化是为具身 AI、为端到端深度学习驾驶、数据基础设施、仿真、在系统上路前的安全许可而建立的。这对我们来说不是对冲或副业，而是我们文化的全部精髓。我认为在资源限制下并以使命驱动的信念做这件事，带来了一系列有趣的创新。走到今天，一切都关乎迭代速度。

I think there are two sides to that question. On the one side, yes, efficiency really matters, but on the second, you can't only rely on learning efficiency. At the limit, if we take our current approach and just scale it up, I'm sure it'll produce generic level five driving, at the limit, if you have unlimited training data, this is really just a lookup data table with some prior experience. But that's not economically or technically feasible. So the question is, how can you train this to be the most data-efficient system? Because I think efficiency will lead to not just improved cost, but faster time to market and more intelligence. Efficiency comes from a number of different factors. There's most importantly how the data curriculum you put in place, but then the learning algorithms. How do you magnify the learning you have? I think world models are a really great opportunity for that. They generate synthetic data and synthetic understanding that doesn't replace real-world data but recombines it and magnifies it in new ways. It lets you pull in interesting insights. I think these kinds of approaches can really improve data efficiency. But across the board, I think working under resource constraints has forced our team to develop so many innovations. I'd also call out just the workflow because in traditional robotics, when you're tuning parameters or algorithms or designing geometric maps, there are very well-established cultures and workflows. Our team, when we have 50 model developers working on one main production model, or when we have an end-to-end net that we need to understand and introspect, or even the way we deploy these systems to simulation or to the road and get feedback, we've developed the entire culture from the ground up at Wayve. It has been developed for embodied AI, for end-to-end deep learning for driving, the data infrastructure, the simulation, the safety licensing before we put systems on the road. This has not been a hedge or a side bit for us, but this is the entire essence of our culture. I think doing this under resource constraints and with full mission-driven conviction has led to a bunch of interesting innovations. Getting to where we are today, everything is about iteration speed.

说到你的文化，我想象的是一群 AI 研究人员、机器学习工程师之类的人。考虑到你们服务的客户群，考虑到你们专门针对汽车行业，以及其围绕供应链的所有特点和安全性要求，你们组织的文化与类似的应用实验室环境有何不同？这对你们的业务文化有何影响？

Speaking of your culture, I'm picturing a bunch of AI research types, machine learning engineers, that sort of thing. How does the culture of your organization differ from similar applied lab environments, given the customer base you serve, given that you're going after the automotive industry specifically with all of its quirks around supply chain and requirements around safety? How does that influence the culture of your business?

影响巨大。事实上，在 Wayve 的头几年，我们真的是一群充满热情的具身 AI 研究人员。但在过去几年里，我为我们团队在理解汽车行业方面建立起的深厚专业知识，以及可靠地向合作伙伴交付的能力感到非常自豪。

Hugely. In fact, for the first few years of Wayve, we were really a group of passionate embodied AI researchers. But in the last couple of years, I'm really proud of how our team has built out deep expertise in understanding the automotive industry, but also the ability to reliably deliver to our partners there.

那是一种不同的文化。我真的很尊重这种文化，因为当你制造数百万辆汽车时，所需的可靠性和平均无故障时间（MTTF）水平是非凡的。你从他们那里学到了什么？我敢肯定你的部分工作是向他们介绍 AI 领域的最新进展。你从他们那里学到了什么？

And that's a different culture. It's a culture I've really grown to respect because when you're building millions of cars, the level of reliability and MTTF you need is extraordinary. What have you all learned from them? I'm sure part of your job is to teach them about what's going on in the world of AI. What have you learned from them?

我认为我提到的一些主要方面是效率和可靠性。产品中技术的差异是一些主要主题。所需的可靠性水平，以及在部署前真正稳健地验证这些系统所需的质量水平，这些公司对此的自豪感是非凡的。另一件事可能是品牌差异化的意识和对……的渴望。

I think some of the main things I'll call out have been efficiency and reliability. The difference between technology in a product would be some of the main themes. The level of reliability required, but also the level of quality that is seen to really robustly prove these systems out before deployment, and the pride that these companies take in that has been exceptional. Another thing has been perhaps the sense of brand differentiation and the desire for...

你希望你的车怎么开？你的驾驶个性如何真正匹配品牌的偏好？

Do you want your car to drive how? How can your driving personality really match the brand's preferences?

你如何提供真正带来品牌差异化的体验？好消息是，我认为我们已经能够在此基础上即兴发挥和集思广益，并沿着这条思路提出一些非常巧妙的技术想法。但最终，安全、高质量和可个性化的 AI 是我们从行业中获得的一些很好的反馈。

How can you provide that experience that really gives brand differentiation? The great news is that I think we've been able to riff and brainstorm off these and come up with some really neat technical ideas down that vein. But ultimately, safe, high quality, and personalizable AI has been some great feedback we've got from the industry.

你能谈谈你与汽车 OEM 合作的市场路径吗？你是如何决定这样做的？然后你认为自动驾驶的市场格局将如何发展？

Can you talk about your path to market actually in partnering with the auto OEMs? How did you decide to do that? And then how do you think the market landscape will play out for how autonomy rolls out?

当然可以。好问题，Sonia。自从 Wave 成立以来，我们一直专注于我提出的关于端到端深度学习作为自动驾驶方法的理念，但多年来我们尝试了许多不同的市场策略。但在过去几年里，我对与全球最大最好的汽车消费制造商合作感到非常振奋。为什么？嗯，我提到他们已经开始引入软件定义汽车。所以他们有处理自动驾驶的基础设施。市场相信这是一项能够真正蓬勃发展的技术。而且这也是一个机会，可以扩展到远超我们目前看到的逐个城市机器人出租车的规模。但更重要的是，这些 OEM 正在投资正确的基础设施，从仅仅是驾驶辅助发展到无需注视的自动驾驶，你可以真正承担驾驶责任，让用户安全并节省驾驶时间。所以这很棒。我认为当你考虑市场时，每年生产 9000 万辆汽车。像特斯拉这样自己构建自动驾驶系统的制造商生产了几百万辆，但市场的绝大部分，我认为有机会与一些创新平台合作，将我们的 AI 推向市场，使这些自动驾驶产品成为可能，而且只会从那里增长。这些制造商不想止步于驾驶辅助。我们正在合作构建无需注视和无人驾驶的机器人出租车产品。但关键是通过避免在这些车辆上改装我们自己的硬件，而是作为软件集成原生植入，我们可以快速规模化。我们可以制造低成本的车辆，并在全球范围内获得认证。我认为这将是看到数十万甚至数百万辆机器人出租车以可承受的价格在全球推出的途径。当然，这一切之所以可能，是因为这种 AI 实现的泛化水平。

Yeah, of course. Great question, Sonia. Since the beginning of Wave, we've been focused on the pitch I gave around end-to-end deep learning being the approach to autonomy, but we've tried a number of different go-to-market approaches over the years. But in the last couple of years, I've been hugely energized about working and partnering with the biggest and best automotive consumer manufacturers around the world. Why is that? Well, I mentioned how they've begun to introduce software-defined vehicles. So they have the infrastructure to work with autonomy. There's the market belief that this is a technology that can really thrive. And also it's the chance to get to scale far beyond what we're seeing with the city-by-city robotaxis right now. But moreover, these are OEMs that are investing in the right infrastructure to go from not just driver assistance but to eyes-off autonomy where you can actually take liability for the drive and give the user a safe and give them time back from their driving experience. So that's awesome. I think when you think about the market, there are 90 million cars built each year. Some manufacturers that are building the autonomy systems themselves like Tesla build a couple of million, but the vast majority of the market, I think there's an opportunity to partner to work with some of these innovative platforms and to bring our AI to market to make these autonomous products possible, and it'll only grow from there. These manufacturers don't want to stop at driver assistance. We're working together to build eyes-off and driverless robotaxi products. But the key thing is that by avoiding retrofitting our own hardware on these vehicles, by putting them in natively as a software integration, we can move fast at scale. We can build low-cost vehicles that can be homologated all around the world. I think this is going to be the path to see tens and hundreds of thousands of robotaxis rolled out around the world at an affordable price. And of course, this is all possible because of the level of generalization that this AI enables.

特斯拉的 FSD 真是一个改变游戏规则的产品。我那些拥有它的朋友无法想象其他驾驶方式。你将赋能每年售出的 8800 万辆其他车辆也能提供这种体验，这真的很酷。

Tesla FSD is just such a game-changing product. My friends that have it, they can't imagine driving any other way. It's really cool that you're going to empower the 88 million other vehicles sold every year to be able to sell that experience as well.

百分之百。很多人跳上我们的车去兜风，有些人对自动驾驶持怀疑态度，但无一例外，他们下车时都面带微笑。这是一种神奇的体验。我迫不及待地想让全世界的人都能尝试它，让自动驾驶不仅仅是机器人出租车的旅游体验，而是最终将这种体验带给每个城市的人们。

100%. It's one of those things that a lot of people jump in our car and come for a drive, some being skeptical about autonomy, but without exception, they step out with a smile on their face. It's a magical experience. I can't wait for people to be able to try it around the world and make autonomy not just a robotaxi tourism experience, but bring this experience to people in eventually every city.

你怎么看传感器融合的困惑争论？就是每年在 Twitter 上上演的那种，如果同时有摄像头和雷达输入，特斯拉会感到困惑。

What do you make of the sensor fusion confusion debate? The one that plays out on Twitter every year or so of Tesla gets confused if there's both camera and radar coming in.

我认为这是错误的争论。这不是前沿问题。行业，我想，除了特斯拉之外，已经真正围绕一个通用架构凝聚起来：环绕摄像头、环绕雷达和前置激光雷达堆栈。现在，这成本低于 2000 美元。所以，它是汽车级组件，而不是你今天看到的改装机器人出租车组件。但在车上拥有前沿 GPU 计算、汽车级 GPU 和那种传感器架构，是构建 L3 L4 无需注视或无人驾驶自动驾驶的绝佳平台。它提供了必要的冗余。它让你处理边缘情况，我同意，仅摄像头可以达到人类水平，但我们想要超越人类水平。所以我认为这种架构是负担得起的、可扩展的，拥有大规模制造的供应链，并且可以消除所有事故，真正实现超人类水平的性能。这就是我们看到许多制造商在他们的车辆上推出的，也是我们集成 AI 的地方。当然，对于驾驶辅助系统，仅摄像头可以工作。对于人类水平的无人驾驶系统，或者我应该说，90%多，你可以看不同的统计数据，但 95%或以上的事故是由人为错误造成的。所以不仅可以达到人类水平，还可以消除很多由人类注意力引起的事故。但仍然有一些事故需要超越视觉的感知能力才能解决。如果我们想解决那个长尾问题，有很多方法。其中一种方法是引入其他传感模式，如雷达和激光雷达。所以我们很高兴能与这类平台合作，但关键是原生集成到 OEM 的车辆本身。

I think it's the wrong debate to be having. It's not the frontier question. The industry has really, I guess, outside of Tesla, has really coalesced around a common architecture of a surround camera, surround radar, and a front-facing lidar stack. Now, this costs under $2,000. So, it's automotive grade components, not the retrofit robotaxi components you see today. But having a frontier GPU compute, automotive grade GPU on the car and that kind of sensor architecture is a really great platform to build L3 L4 autonomy eyes-off or driverless. It gives you the necessary redundancy. It lets you deal with edge cases that cameras alone, I agree, can get you to human level, but we want to go beyond human level. So I think this kind of architecture is affordable, scalable, it's got the supply chain for mass manufacture, and it can eliminate all accidents and really drive superhuman levels of performance. So that's what we're seeing many manufacturers bring out on their vehicles and where we're integrating our AI. Of course, for a driver assistance system, camera only can work. For a human-level driverless system, or of course I should clarify, 90-something percent, you can look at different stats, but 95% or above accidents are caused by human error. So not only can you be human level, but you can eliminate a lot of human attention and accidents caused by that. But there are still accidents that to be able to solve would require perception capabilities that go beyond vision. And if we want to tackle that long tail, there are many ways to solve it. One of the ways would be to bring in some other sensing modalities like radar and lidar. So we're excited to be working with those kind of platforms, but crucially natively integrated into the OEM's vehicles themselves.

同一个神经网络能驱动不同车企的汽车吗？这到底是怎么做到的？因为我想每辆车摄像头的摆放位置之类都有细微差别。

Is it the same neural net that can drive on one OEM's car and another's car? And how does that even work? Because I imagine each vehicle has slightly different positioned cameras, things like that.

它们来自同一个家族。我们训练一个非常大规模的模型——我们定期训练超大规模模型。当然我们每月都会迭代。但那是所有合作车队共用的一个模型。不过，当你针对特定传感器组或特定嵌入式目标进行优化时，就可以开始专门化模型。美妙之处在于，99%以上的成本、时间和精力都花在训练基础模型上，然后我们可以为特定客户构建非常高效的个性化方案。这让我们既能规模化，又能将其压缩到非常高效的实时平台，并适配特定用例。

It comes from the same family. So we train a very large scale model—we regularly train very large scale models. Of course we iterate on them monthly. But that's one model common to all the fleets we work with. However, as you optimize for a specific sensor set or a specific embedded target, you can start to specialize the model. The beauty is that 99% plus of the cost, time, and effort is training that base model, and then we can build very efficient personalization for the specific customer. This lets us scale but gives us the ability to squeeze it to very efficient real-time platforms and adapt it to a specific use case.

你会让 Pat 个性化一个超级激进的驾驶模型吗？我肯定需要。

Are you gonna let Pat personalize a super aggressive driver model? I'm gonna need to.

Pat，你想要哪种驾驶风格？嗯，相当激进、安全、非常安全。但你知道，我们可以做到。我们发现围绕驾驶行为构建分布真的很有趣。从我们拥有的人类训练数据中，你确实能看出什么时候从所谓的有益自信变成了无益的激进。我们可以在那里划出一条清晰的界线。

What driving style would you like Pat? Yeah, pretty aggressive, safe, very safe. But, you know, we can do that. We find it really funny when you build distributions around driving behavior. You can really tell from the human training data we have, you can really tell when it goes from being helpfully assertive, let's say, to unhelpfully aggressive. And we can draw a clean line there.

这就对了。

There you go.

你呢，Sonia？我们刚才的驾驶体验怎么样？

What about you, Sonia? How was the drive we just had?

太棒了。很舒适，很安全。而且实际上感觉非常人性化，比如它在转弯时看不到情况会慢慢试探。非常像人。

Fantastic. It was comfortable. It was safe. And it felt very human actually, like the way it was kind of nudging up when it couldn't see on the turn. It was very human.

是的。嗯，在硅谷我们可以做到很复杂，但到了东京或伦敦——我这周末在旧金山市区——你真的需要预测和推理周围其他人的能力，才能以类人的方式驾驶。我们发现，如果你不能平稳地绕过双排停放的车辆或处理其他动态障碍，甚至如果车流的主流方向并不完全对齐车道，但有一种类人的驾驶方式，那么我们构建的智能的厉害之处在于它能推理这些情况，保持交通流畅，以非常类人的方式与其他道路使用者互动。我认为这将是社会接受并喜爱无人驾驶出租车的关键。我迫不及待想让它成为现实。

Yeah. Well, as complex as we can get in Silicon Valley, but come to Tokyo or London or—I was in downtown San Francisco this weekend—and yeah, you really need the ability to predict and reason about other folks around you to be able to drive in a humanlike way. What we find is that if you're not able to smoothly go around double-parked vehicles or deal with other dynamic obstacles, or even if the prevailing row of traffic might not be aligned to the specific lane but there's a humanlike way of driving, then what's awesome about the intelligence we built is it's able to reason about these things and keep the traffic flowing, keep interacting with road users in a very humanlike way. I think this is going to be key for societies to accept and love robotaxis. I can't wait to make that a reality.

目前你们的汽车有没有哪些特定的边缘场景处理起来很困难？

Are there any specific corner cases that your cars have a hard time with today?

有很多，而且很难泛泛地谈某一个，因为它们太罕见了。如果我说‘哦，总是这类情况’，那总是几个边缘情况在角落里凑到一起，当你遇到非常罕见的情况时，总是混杂因素。但我们在 500 个城市行驶。所以当你在这种规模下驾驶时，当然会看到从未见过的东西。用新语言写的路标。实际上，也许一种分解方式是，我们常把驾驶分解为安全性、实用性和流畅性。安全性指安全关键行为，流畅性指驾驶风格——是否平稳？是否愉悦？——实用性指导航和道路语义。我们发现安全性和流畅性在全球范围内泛化得非常好。在我们运营的每个国家，安全性和流畅舒适度的指标几乎一致。但实用性是我们走向全球时非常有趣的一点。如何导航？如何处理路标？如何阅读不同语言？如何处理不同的驾驶文化？所以这一点很有趣。根据我们发表的结果，从英国到美国，我们需要数百小时的数据才能达到前沿性能的 10%以内。但当我们去欧洲和德国时，当然我们已经学会了靠右行驶。来到美国，我们学会了红灯右转。然后到德国，我们得学会仍然靠右行驶，但那里不能红灯右转。但在高速公路上，你需要开到——我们目前最高开到 140 公里/小时。所以那里速度很快。但每次效率都更高，每个新市场所需的数据呈指数级减少，因为你以前见过其中一些情况。

There are loads, and it's really hard to generically talk about one because they're so rare. If I was to say 'oh, it's always these types,' it's always a couple of edge cases coming together in a corner, and it's always confounding factors when you get something really obscure. But we're driving in 500 cities. So when you're driving at that level of scale, of course you see things you've never seen before. Road signs written in a new language. Actually, maybe one way to break it down is often we talk about driving broken down into safety, utility, and flow. Safety being safety-critical behavior, flow being the style of driving—is it smooth? Is it enjoyable?—and utility being the navigation and road semantics. Safety and flow we found generalize exceptionally well throughout the world. We get almost uniform metrics in every country we operate in terms of safety and flow comfort. But utility has been the really interesting one as we've gone global. How do you navigate? How do you deal with road signs? How do you read different languages? How do you deal with different driving cultures? So that's the one that's been interesting. From our published results, when we went from the UK to the US, we needed hundreds of hours of data to be able to drive within 10% of our frontier performance. But then when we went to Europe and Germany, of course we'd already learned to drive on the right side of the road. Coming to the US, we learned to do right turns at red lights. Then coming to Germany, we had to learn to still drive on the right side of the road, but you can't turn right at a red light there. But on the autobahn, you need to drive—we drive today up to 140 km/h. So pretty fast there. But it gets more efficient each time, with exponentially less data in each new market because you've seen some of those things before.

你一开始提到，大语言模型是让你从逆向思维转向共识的一部分。你们有没有将大语言模型整合到你们的模型中？我知道现在一些刚起步的机器人公司是从 VLA、VLM 基础开始的。这是你们架构的一部分吗？

You mentioned at the beginning that large language models were part of what flipped your approach from contrarian to consensus. Are you integrating large language models at all into your models? And I know some of the robotics companies that are getting started now are starting from this VLA VLM base. Is that part of your architecture?

100%。2021 年，我们开始研究驾驶语言。我记得当时我的团队来找我说，我们应该启动一个语言项目。我说不，不，不，创业公司要专注。保持专注。但他们确实给出了一些很有说服力的论据。所以我们开始摆弄这些东西。大约一年后，我们发布了 Lingo，这是自动驾驶领域第一个视觉-语言-动作模型。这个模型的特别之处在于，它不仅能驾驶汽车——观察世界、驾驶汽车——还能用语言交流。它可以让你跟它说话，问它问题：‘你觉得哪里危险？接下来会发生什么？’甚至可以为你的驾驶做解说。有趣的是，这有几个好处。一是将语言引入预训练当然能提升表征能力，给你比仅靠图像更丰富的信息来学习。但其次，将表征与语言对齐开启了许多有趣的产品功能。它实现了 chauffeur 体验，你可以真正与你的司机对话。你不再需要机器人学博士学位来理解系统，而是可以直接跟它说话，让它驾驶。Pat，如果你想在通勤路上飙车，你可以直接要求。

100%. In 2021, we started working on language for driving. I remember my team came to me at the time and said we should start a project on language. I said no, no, no, startup's all about focus. Keep focused. But they actually gave some pretty compelling arguments. So we started to play around with these things. A year or so later we released Lingo, which is the first vision-language-action model in autonomous driving. What was special about this model was it could not only drive a car—see the world, drive a car—but also converse in language. It'll let you talk to it, ask it questions: 'What are you finding that's risky? What's going to happen next?' Or even it could commentate your drive. What's interesting about this is that there are a few benefits. One is bringing language into pre-training of course just improves the representation's power, gives you more interesting information to learn from than just imagery alone. But then second, aligning the representation with language opens up a ton of interesting product features. It enables a chauffeur experience where you could actually talk to your driver. No longer do you need a PhD in robotics to understand the system, but actually you can just talk to it and ask it to drive. Pat, if you want to race around the commute super fast, then you can demand that.

但第三，它为你提供了一个非常好的内省工具，你可以想象监管机构或我们的工程团队用语言与系统对话，真正诊断它为什么这样做，或者让它解释其推理。所以我认为这些都是非常明确的好处，我们很兴奋在推动这些。

But then third, it gives you a really nice introspection tool where you can start to actually, you know, you could imagine regulators or our engineering team converse with the system in language to really diagnose why it's doing what it's doing or get it to explain its reasoning. So I think these are really clear benefits which we're really excited to be pushing.

这太酷了。而且你是在嵌入式算力上运行它。

That's super cool. And you're running it on the embedded compute.

是的。我们已经发布了在车外运行的演示。在今天的汽车市场上，车载运行很有挑战性，但一些下一代算力，例如我们下一代开发车辆将搭载的 Nvidia Thor，将足够大以在车上运行。

We are. So, we've put out demos that run offboard. Onboard's challenging with what's in the automotive market today, but some of the next generation compute, for example, the Nvidia Thor that our nextG development vehicle is going to be built with will be large enough to run it on board.

那会很酷。

That's going to be cool.

非常酷。

Very cool.

你谈到自动驾驶如何为更通用的具身 AI 提供路径。你能为我们描绘一下吗？如何从自动驾驶到人形机器人或其他你想让 AI 具身的东西。

You've talked about how autonomous driving sort of provides a path to more generalized embodied AI. Can you paint that picture for us? How you go from autonomous driving to humanoid robots or whatever other things you might want to embody AI.

我认为未来我们会看到大量有趣的机器人应用。我们看到的是，移动性比操作更早成为可能。操作在数据获取、硬件全球供应链甚至硬件设计本身方面都很有挑战性，比如触觉传感仍然是一个很难的挑战。但不可避免的是，这将是一个巨大的变革性事物，也许它处于 2015 年自动驾驶的成熟度。但今天，我们的系统正在迅速成为一个通用的导航智能体：给它任意视角和目标条件，它就能生成安全轨迹。所以我认为我们将看到快速进步，不仅限于消费级自动驾驶出租车，还有卡车运输和其他应用，而且这种 AI 将使制造商和车队能够在任何移动应用中构建机器人。当然，我们非常兴奋能与前沿开发者和应用合作，随着我们深入机器人栈。我预计未来几年，制造和操作用例也会更加成熟。但最终，我认为拥有大型基础模型的好处是，在汽车领域我们拥有最大的机器人和数据供应链，所以我们在这方面很幸运，能够推动那里的智能。但将这种智能推广到新应用，我认为模型能够体验多个不同垂直领域将带来好处，只会让它更通用。

I think we're going to be in the future looking at a ton of interesting use cases for robotics. What we're seeing is that mobility is becoming possible much before manipulation. Manipulation is challenging in terms of access to data, global supply chains for hardware and actually even the hardware designs themselves like I think tactile sensing is still a really hard challenge. But inevitably it'll be a massive transformative thing but maybe is at the maturity of where self-driving was in 2015. But today our system is rapidly becoming a general purpose navigation agent: give it an arbitrary sense of view and a goal condition, it's able to produce a safe trajectory. So I think we're going to see a rapid advancement from not just consumer automotive robo taxis, think about trucking and other applications, but this AI will enable manufacturers and fleets who want to build robots in any kind of mobility application. And of course we're really excited to be working with frontier developers and applications over time as you go out across that robotic stack. And I expect we'll see more maturity in the coming years from manufacturing and manipulation use cases as well. But in the end I think the benefits of having a large foundation model that certainly in automotive we have access to the largest robot and data supply chain and so we're really lucky in that regard to be able to push forward the intelligence there. But generalizing that intelligence to new applications, I think there'll be benefits from the model being able to experience multiple different verticals and it will only make it more general purpose.

有什么你特别兴奋的应用吗？

Any applications you're excited about?

我真的很兴奋看到人形机器人到处走动。

I mean, I'm psyched to have humanoid robots walking around.

是的，我也是。我认为它们会很棒。无论哪种形态，人形机器人都会扮演重要角色。其他运动形式以及操作也是如此。这些领域有一些非常有趣的挑战，但我认为同样的故事会重演：在狭窄应用上工作，比如自动驾驶去亚利桑那州凤凰城，投入大量基础设施和昂贵硬件来使其工作，跑道有限；但致力于通用、精简、低成本的硬件栈，真正专注于让系统最智能和最鲁棒，我认为这是规模化的秘诀。所以，让我们关注这个领域。

Yeah, me too. I think they're going to be neat. Whichever form factor, I think humanoids will play a big part. I think other forms of locomotion as well and then manipulation. There's some really interesting challenges in those spaces but I think the same story is going to play out: working on a narrow application like when self-driving went to Phoenix Arizona and put in a ton of infrastructure and expensive hardware to make it work is going to have limited runway, but working on general purpose lean low-cost hardware stacks that really focus on making the system most intelligent and robust, I think this is the recipe for scale. So yeah, let's watch that space.

你认为达到所谓的物理 AGI 是否需要重大的研究突破？如果是，你认为最有希望的方向是什么？

Do you think there are major research breakthroughs needed to reach kind of physical AGI so to speak and if so what do you think is the most promising direction?

绝对需要。我认为当前方法还有很大的扩展空间，我们会这样做，但我认为我们将从驱动性能的四个因素中获得复合回报。当然有数据和算力，但还有算法能力和具身：机器人的硬件和能力。我认为我们需要推动所有四个。在算法方面，有很多增长机会。我认为关键之一是测量。你如何实际测量和量化这些系统？如何快速响应、发现回归、拥有一个能大规模缩小现实差距且高效运行的模拟器？这些生成式世界模型非常消耗算力，这已不是秘密。但拥有一个好的测量系统将推动效率和迭代速度。所以这是一个关键。人们经常谈论这是一个先有鸡还是先有蛋的问题。如果你有一个完美的模拟器，你就解决了自动驾驶，反之亦然。我确实相信这一点。AlphaGo 表明，当你有一个完美的模拟器时，你可以通过蒙特卡洛搜索解决问题。所以我认为在机器人领域也会如此。所以一个是测量。另一个支柱是在模型中构建更多通用性。你如何构建更多模态并使其推理对齐？我认为这将开启新的用例，特别是在人机交互和导航方面。我之前提到了效用问题。其中一些事情我超级兴奋。最后一个是工程效率。训练这些系统和数据需求是巨大的。所以我不低估，我认为这个问题最性感的部分是高效的基础设施来训练和服务这些模型。并且把它做对。我认为这是一个真正的竞争优势或劣势。

Absolutely I do. I think there's so much more ground to scale up the current approaches and we'll do that but I think we will get compounding returns from four factors that drive performance. There's of course data and compute but then also the algorithmic capabilities and the embodiment: what is the hardware and capability on the robots. And I think we need to push all four. On the algorithmic side, there are so many opportunities for growth. I think a key one is measurement. How do you actually measure and quantify these systems? How do you respond quickly, find regressions, be able to have a simulator that closes the real world gap at scale and can run efficiently? It's no secret that these generative world models are very compute intensive. But having a good measurement system will just drive efficiency and iteration speed. So that's a key one. People often talk about being a chicken and egg. If you have a perfect simulator, you've solved self-driving and vice versa. And I really believe that. AlphaGo showed that when you have a perfect simulator you can just solve problems through Monte Carlo search. So I think that's going to be the case in robotics as well. So one is measurement. Another pillar is building more generality into the model. How can you build out more modalities and align those different modalities in their reasoning? I think this is going to open up new use cases particularly when it comes to human robot interaction and navigation. I was going back to the utility problem before. Some of these things I'm super excited about. And then the last one is just engineering efficiency. I mean training these systems and the data requirements is extraordinary. So I wouldn't understate, I think the most sexy part of this problem is the efficient infrastructure to train and serve these models. And getting that right. I think it's a real competitive advantage or disadvantage.

我们从谈论 AV2.0 开始。有一天，我想我们可能会谈论 AV3.0。AV3.0 会是什么样子？如果展望 5 年、10 年、15 年，你认为这个行业还会有其他重大飞跃吗？

We started by talking about AV2.0. Someday, I imagine we might be talking about AV3.0. What could AV3.0 look like? If you go 5, 10, 15 years in the future, are there any other big leaps in this industry that you think we'll see?

你那么面无表情地说 AV……AV 2.0 的整个前提就是把智能放在车上，不需要基础设施和大量过度的硬件，而是真正让系统变得智能。所以我认为我们现在看到这一点正在显现，系统能够通过所有车载可扩展智能和算力泛化到世界。

You said that with such deadpan. AV... So the whole premise of AV 2.0 was all about putting the intelligence on the car and not needing infrastructure and a ton of overcooked hardware, but really making the system intelligent. And so I think we're seeing that emerge now with the system that can generalize to the world with all of the onboard scalable intelligence and compute.

如果要推测 AV 3.0 的走向，一个想法是把智能延伸到车外。当大多数车辆都是自动驾驶时，它们可以通信和互动。如果它们能协调，为什么还需要红绿灯？如果能与前方车辆通信看到拐角，为什么还需要传感器？这会带来网络安全和延迟问题。我完全支持具身 AI，如果能通过将智能扩展到车外来构建更安全、更易用的系统，那或许是一条路。

If I were to speculate where AV 3.0 goes, one idea could be taking the intelligence outside the car. When you have a majority of autonomous vehicles, they can communicate and interact. Why do we need traffic lights if they can coordinate? Why do we need sensors if you can communicate with the AV in front to see around corners? It opens up cybersecurity and latency questions. I'm all for embodied AI, and if we can build a safer, more accessible system by taking intelligence beyond the car, maybe that's a path.

我觉得这很有意思。如果 AV 3.0 是一个网状网络，也许人类不被允许驾驶，因为他们无法像机器人那样与网络通信。或者有专门的娱乐驾驶场所，但交通出行全是自动驾驶。

I think that's really interesting. If AV 3.0 is a mesh network, maybe humans aren't allowed to drive because they can't communicate with it. Or there are special places for recreational driving, but for transportation it's all autonomous.

在如今 AI 市场如此火热的情况下，你们如何招聘和吸引人才？

How do you hire and attract people with how hot the AI market is these days?

我喜欢这个问题。团队就是我们的产品。在 Wayve，我们致力于成为能让你做出职业生涯最佳成就的地方。在具身 AI 领域，这意味着拥有鼓舞人心的同事、合适的资源以及能为你扫清障碍的文化。独特之处在于，我们将前沿 AI 环境与汽车领域的近期产品机会结合在一起。如果你想研究智能机器，并看到你的系统像 ChatGPT 在机器人领域那样大规模部署，这里就是你的选择。我们已走向全球，在伦敦、斯图加特、特拉维夫、温哥华、东京、硅谷都有团队。我们正在构建一种全球文化，将产品推向世界。任何对开创具身 AI 并将其转化为颠覆性产品感兴趣的人，欢迎来聊。

I love that question. Our team is our product. At Wayve, we talk about being a place where you can do the best work of your career. In embodied AI, that means having inspiring colleagues, the right resources, and a culture that unblocks you. Uniquely, we bring together a frontier AI environment with a near-term product opportunity in automotive. If you want to work on intelligent machines and see your system deployed at the scale of ChatGPT in robotics, this is the place. We've gone global with teams in London, Stuttgart, Tel Aviv, Vancouver, Tokyo, Silicon Valley. We're building a global culture to bring this product to the world. Anyone interested in pioneering embodied AI and turning it into a game-changing product, come chat.

Alex，你比几乎所有人都更早相信端到端神经网络在自动驾驶和实体经济中的未来。看到这个愿景开始成为现实，一定非常令人满足。祝贺你，感谢你加入我们。

Alex, you've believed in the future of end-to-end neural nets in self-driving and the physical economy longer than almost anyone. It must be incredibly fulfilling to see that vision start to come to life. Congratulations and thank you for joining us.

谢谢 Sonia，谢谢 Pat。这是我的荣幸。

Thank you, Sonia. Thank you, Pat. It's such a privilege.