Levandowski Dumps LiDAR at Pronto, Calls Prediction Real AV Barrier

Anthony Levandowski on lessons learned at TC Sessions: Robotics+AI — TechCrunch

TechCrunch

Autonomous vehicle disengagements stem from software failing to forecast human behavior rather than sensor deficits, making direct-to-Level 4 moonshots unviable compared to iterative commercial deployment.

AI-assisted summary of Anthony Levandowski on lessons learned at TC Sessions: Robotics+AI — original source. Check the original for full context.

Source published:

Key ideas

  • Anthony Levandowski has dropped LiDAR at his startup Pronto, citing behavioral prediction and motion forecasting—not sensor hardware—as the primary bottleneck in autonomous vehicle development. (0:00)
  • Disengagement data from mature autonomous driving programs reveals that hardware failures account for virtually zero percent of disengagements, which stem almost entirely from software misunderstandings and miscommunications. (2:00)
  • Levandowski rejects the direct 'moonshot' path to Level 4 autonomy, arguing that an iterative, evolutionary rollout in consumer hands is the only pragmatic route to ensure safety. (4:33)
  • Dismissing extreme AI doom narratives that equate machine learning to weapons of mass destruction, Levandowski contends technological progress is inevitable and urges society to build without fear. (6:41)

Detailed notes

## Background Published on YouTube by TechCrunch on April 22, 2019, this video captures an interview with former star Google engineer Anthony Levandowski at TC Sessions: Robotics+AI. Following his involvement in the high-profile Waymo v. Uber trade secrets lawsuit, Levandowski sits down to discuss autonomous vehicle development and his new startup, Pronto.ai.

## TL;DR Anthony Levandowski explains why his startup Pronto deliberately avoids LiDAR, admitting past errors and arguing that the primary autonomous bottleneck has shifted from sensor hardware to behavioral prediction and software communication. Reflecting on his departure from corporate friction, he describes how a forced clean slate gave him the perspective to discard legacy assumptions. He dismisses the moonshot approach of leaping straight to Level 4 autonomy, advocating instead for an evolutionary deployment model where driver-assist systems iteratively learn and prove safety in the wild. Finally, addressing wider anxieties about artificial intelligence, Levandowski pushes back against apocalyptic panic, urging developers and society to approach inevitable progress without fear.

## Atmosphere & style The session has an engaging, candid, and slightly self-deprecating atmosphere, with Levandowski noticeably relaxed as he eats 'humble pie' regarding his previous dogmatic defense of LiDAR. The dynamic between the interviewer and guest is playful yet incisive, freely touching on past controversies like his AI-focused church and high-stakes legal battles, which Levandowski meets with chuckles and historical analogies. Listening to the exchange feels like watching a chastened pioneer who has traded Silicon Valley moonshot hype for hard-won, pragmatic engineering realism.

## Bulletin - [00:00](https://www.youtube.com/watch?v=fNgEG5rCav4&t=0s) Anthony Levandowski has dropped LiDAR at his startup Pronto, citing behavioral prediction and motion forecasting—not sensor hardware—as the primary bottleneck in autonomous vehicle development. - [02:00](https://www.youtube.com/watch?v=fNgEG5rCav4&t=120s) Disengagement data from mature autonomous driving programs reveals that hardware failures account for virtually zero percent of disengagements, which stem almost entirely from software misunderstandings and miscommunications. - [04:33](https://www.youtube.com/watch?v=fNgEG5rCav4&t=273s) Levandowski rejects the direct 'moonshot' path to Level 4 autonomy, arguing that an iterative, evolutionary rollout in consumer hands is the only pragmatic route to ensure safety. - [06:41](https://www.youtube.com/watch?v=fNgEG5rCav4&t=401s) Dismissing extreme AI doom narratives that equate machine learning to weapons of mass destruction, Levandowski contends technological progress is inevitable and urges society to build without fear.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=fNgEG5rCav4&t=0s) Pronto's decision to forgo LiDAR is a technical choice rather than a legal constraint from the Waymo v. Uber lawsuit. - [00:35](https://www.youtube.com/watch?v=fNgEG5rCav4&t=35s) The primary technical hurdle in autonomous driving has shifted from range sensing and mapping to behavioral understanding and motion forecasting. - [00:52](https://www.youtube.com/watch?v=fNgEG5rCav4&t=52s) Autonomous systems must predict not only what other actors will do, but also model what those actors expect the autonomous vehicle to do. - [01:25](https://www.youtube.com/watch?v=fNgEG5rCav4&t=85s) Levandowski concedes his previous position as LiDAR's champion, agreeing with Elon Musk's assessment that LiDAR acts as a crutch. - [02:00](https://www.youtube.com/watch?v=fNgEG5rCav4&t=120s) Hardware components such as fail-operational steering and braking account for virtually 0.0% of autonomous vehicle disengagements. - [03:33](https://www.youtube.com/watch?v=fNgEG5rCav4&t=213s) Stepping away from major corporate environments gave Levandowski a clean slate to reassess technical assumptions without internal political friction. - [04:33](https://www.youtube.com/watch?v=fNgEG5rCav4&t=273s) Safety is the sole metric that determines when autonomous vehicles will successfully deploy commercially. - [04:58](https://www.youtube.com/watch?v=fNgEG5rCav4&t=298s) Skipping straight to Level 4 moonshot autonomy is unrealistic because solving every rare edge case upfront takes too long. - [05:15](https://www.youtube.com/watch?v=fNgEG5rCav4&t=315s) The Silicon Valley analogy that 'a ladder cannot reach the moon' fails because driving to the corner store is an evolutionary, grounded task rather than a moonshot. - [07:00](https://www.youtube.com/watch?v=fNgEG5rCav4&t=420s) Attempting to halt or cage AI development out of fear is counterproductive; progress is inevitable and must be actively guided. - [07:15](https://www.youtube.com/watch?v=fNgEG5rCav4&t=435s) Society should approach machine intelligence as a beneficial tool for connection rather than comparing it to nuclear weapons.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=fNgEG5rCav4&t=0s) Abandoning LiDAR and Re-Evaluating the Autonomous Bottleneck

The interviewer opens by pointing out that Anthony Levandowski is widely known for his pioneer work on LiDAR, yet his new startup, Pronto, is deliberately not using it. She asks whether this decision is due to legal restrictions stemming from the Waymo v. Uber lawsuit or another technical reason. Levandowski clarifies that while he has no legal constraints preventing him from using LiDAR, he holds a strict personal rule: “I do have restriction personally of not doing things that I know aren’t going to work.” He recalls that around 2009 or 2010, early LiDAR units lacked the necessary range and performance for autonomous software to react safely. However, he argues that the fundamental challenge today is no longer about sensing further or mapping more accurately. Instead, the real technical deficit lies in behavioral understanding and motion forecasting—predicting not only what surrounding vehicles and pedestrians will do, but also understanding what other actors expect your own vehicle to do.

Levandowski openly concedes his past stance, admitting he was once LiDAR's biggest champion and believed it was the secret sauce for autonomous driving. Acknowledging that Elon Musk had faced public ridicule for dismissing the technology, Levandowski concurs: “A wiser person than me, and I’ll eat some humble pie here, said that LiDAR is a crutch.” He stresses that his departure from LiDAR is not out of inability to build it, but because it misses where value is generated. Looking at autonomous vehicle disengagement data across mature companies, he highlights that hardware failures—such as lack of redundant steering or brakes—account for virtually zero percent of disengagements. Instead, failures stem almost exclusively from software misunderstandings and miscommunications with surrounding actors. Comparing the shifting consensus to Galileo overturning geocentrism (joking that he can relate to facing an "Inquisition"), he summarizes the core pivot: the industry's bottleneck is no longer sensor hardware, but prediction.

```mermaid flowchart TD accTitle: Autonomous Vehicle Failure Analysis accDescr: Diagram contrasting sensor and hardware factors with behavioral prediction failures.

subgraph Hardware["Perceived Hardware Issues (0.0% Real Bottleneck)"] A["Sensor Range & Accuracy (LiDAR)"] B["Fail-Operational Steering / Brakes"] end

subgraph Software["Actual Autonomous Bottleneck"] C["Behavioral Perception"] D["Predicting Surrounding Actor Motion"] E["Modeling What Others Expect You to Do"] end

Hardware -.->|"Not the real cause of disengagements"| F["System Failure"] Software ==>|"True source of miscommunications & disengagements"| F ```

[03:33](https://www.youtube.com/watch?v=fNgEG5rCav4&t=213s) Clean Slates, Level 4 Realities, and Evolutionary Deployment

The discussion moves to whether the high-profile Waymo v. Uber lawsuit served as a blessing in disguise by forcing Levandowski to discard assumptions he had pursued for over a decade. He reflects that when an engineer is entirely consumed within an organization, day-to-day internal friction and team politics obscure the broader picture: “It’s very difficult to get perspective and see like you think you’re there, but actually you’re missing this other path that’s going to block you completely.” Stepping away allowed him time for personal reflection and a clean-slate reassessment of the problem. When the interviewer jokingly brings up his controversial AI-focused church (Way of the Future), he clarifies with a laugh that the church was actually founded earlier, in 2015 while he was still at Google.

Turning to the broader industry landscape, the interviewer asks if Levandowski has abandoned the belief that Level 4 autonomy is achievable in the near term, given that dozens of competitors are still pursuing it directly. Levandowski emphasizes that safety—not the choice of sensors, car manufacturers, or software vendors—is the sole criterion for when autonomous vehicles will actually deploy. He rejects the moonshot, "straight to Level 4" approach as unrealistic in the short term because resolving every edge case upfront takes too long. Addressing the popular silicon valley counterargument that "building a ladder doesn't get you to the moon," Levandowski counters: “The thing is that we’re not going to the moon, we’re just trying to drive to the corner store and back.” He contends that an evolutionary path, where consumer systems iteratively improve in the wild, is the only pragmatic way to gather lessons, chip away at safety problems, and achieve true reliability.

- Safety serves as the only real metric dictating when commercial AV deployment occurs. - Developing incremental, evolving driver-assist products uncovers operational insights that direct-to-Level 4 development pipelines miss.

[06:41](https://www.youtube.com/watch?v=fNgEG5rCav4&t=401s) AI Anxiety and the Path to the Singularity

To close the session, the interviewer humorously asks what humanity should do to appease its future artificial intelligence overlords ahead of an impending singularity apocalypse. Levandowski plays along, quipping that people can enroll in a Udacity course or similar online classes to understand the technology better.

Shifting to a serious note, he argues that technological progress cannot be stopped and that panic is counterproductive. He pushes back against extreme doom-mongering narratives that equate machine learning to weapons of mass destruction: “Progress is an inevitable, and you know you want to make sure that you help create the progress and understand where that’s going as opposed to deny it or try to cage it up or you know, say that it’s more dangerous than nuclear weapons.” He concludes that society has a choice between engineering AI from a place of bias and fear or building it as a beneficial tool for human connection, boiling his philosophy down to a direct three-word takeaway: “Don't be scared.”

## Notable quotes > [03:33](https://www.youtube.com/watch?v=fNgEG5rCav4&t=213s) The thing is that we’re not going to the moon, we’re just trying to drive to the corner store and back. — Anthony Levandowski

## Mentioned Anthony Levandowski, Elon Musk, Galileo Galilei, Google, Pronto.ai, TechCrunch, Uber, Udacity, Way of the Future, Waymo

## Video description Anthony Levandowski, the former star Google engineer and serial entrepreneur who was at the center of a trade secrets lawsuit between Uber and Waymo, is back with a new startup. TechCrunch sits down with Levandowski for a wide-ranging discussion on autonomous vehicles and his new startup Pronto.ai

[Source](https://www.youtube.com/watch?v=fNgEG5rCav4)

中文简报

莱万多夫斯基在Pronto弃用激光雷达，称预测才是自动驾驶的真正障碍 自动驾驶脱离源于软件无法有效预测人类行为而非传感器缺陷，这使得直接跃升至L4级的登月式构想相比渐进式商业部署不再可行。 安东尼·莱万多夫斯基解释了他的初创公司Pronto为何特意避开激光雷达，他承认过去的失误，并指出自动驾驶的主要瓶颈已从传感器硬件转向行为预测和软件通信。回顾自己脱离大公司内耗的经历，他描述了被迫从零开始如何让他获得全新视角，从而摒弃传统假设。他否定了直接跃升至L4级自动驾驶的登月式路径，转而主张采用渐进式部署模式，即让驾驶辅助系统在真实环境中迭代学习并证明其安全性。最后，针对更广泛的人工智能焦虑，莱万多夫斯基驳斥了末日恐慌论，呼吁开发者和社会毫无畏惧地迎接不可避免的技术进步。 安东尼·莱万多夫斯基在其初创公司Pronto弃用了激光雷达，称行为预测和运动轨迹预判——而非传感器硬件——才是自动驾驶汽车开发的主要瓶颈。 (0:00) 来自成熟自动驾驶项目的脱离数据显示，硬件故障在脱离原因中几乎占零比例，脱离几乎完全源于软件的理解偏差与沟通不当。 (2:00) 莱万多夫斯基拒绝直接走向L4级自动驾驶的“登月”路径，认为在消费者手中进行迭代、渐进式推广才是确保安全的唯一务实途径。 (4:33) 莱万多夫斯基驳斥了将机器学习等同于大规模杀伤性武器的极端AI末日论，坚称技术进步不可阻挡，并敦促社会无惧前行、积极建设。 (6:41)

Source quotes

[3:33] Anthony Levandowski: The thing is that we’re not going to the moon, we’re just trying to drive to the corner store and back.

Original transcript

[0:00] so you're speaking of Hardware though [0:02] you are known for your work on ldar and [0:06] yet Pronto is not going to be using lar [0:10] and I'm wondering why is it is it [0:13] because a res are you just staying away [0:16] from it because of the wh Uber lawsuit [0:19] or is there some other reason why liar [0:22] is not in the picture with Pronto yeah [0:24] really good so that's a great question [0:25] too uh so I would say uh so I don't have [0:29] any restrictions on not doing lar but I [0:31] do have restriction personally of not [0:33] doing things that I know aren't going to [0:34] work um so uh in the past uh you know I [0:40] I I understood I could see what we were [0:42] doing and basically back in 200 you say [0:46] N9 or 10 you could see that the Lars [0:49] didn't have the performance for what you [0:51] needed because you couldn't see far [0:52] enough to actually safely have the [0:56] system react to the software uh safe [1:00] have the system see what's going on to [1:02] have a correct reaction it turns out [1:04] that even with lar um what's missing [1:07] today uh is not seeing more accurately [1:11] or seeing further what's missing is [1:14] understanding what is happening to those [1:15] vehicles around you and being able to [1:17] predict what the motion is uh so that [1:20] you can react accordingly and be safe [1:22] and not just predict what their motion [1:23] is going to be but what they think [1:24] you're going to be doing and so um we we [1:29] we mainly focus on that because that's [1:31] the that's the building block that is [1:33] unlocking the uh the future now I'll say [1:38] I was a huge proponent of lar and it was [1:41] the sauce right well that was not my [1:42] quote but I I definitely [1:45] um believed that that was actually the [1:48] unlocking Factor uh you know a a wiser [1:51] person than me and I'll eat some Humble [1:53] Pie here said that lar is a crutch yeah [1:55] I was just about to ask you about that [1:57] so you would agree with Elon Musk he [1:59] said and he's been been ridiculed for [2:01] that a bit um right for calling lar a [2:04] crutch and that's what you believe as [2:05] well I believe that as well and I don't [2:07] believe that because I can't do it [2:08] because I'm happy to do that it's just [2:10] not where I think the actual value is [2:13] coming from and and you could look at [2:14] that right so if you were to like [2:16] analyze all the disengages that people [2:17] have done and try to break it down like [2:19] what's the actual real reason behind it [2:22] is the real reason because the brakes [2:24] weren't redundant or the steering wasn't [2:25] redundant No in fact that's like [2:27] 0.0% of all this ages are because there [2:30] was no fail operational system all of [2:33] them are software failure and they're [2:34] mainly a software failure um you know on [2:37] mature companies um at trying to [2:40] understand what the vehicle is going to [2:41] doing or the pedestrians around you and [2:43] you having a misunderstanding or [2:45] miscommunication with them and so that's [2:47] the that's where the value is going to [2:49] be now we could have this conversation [2:51] in a couple years and be like that [2:52] wasn't the problem the problem is this [2:53] other new thing that we don't know about [2:55] but in like all scientific process you [2:58] make a hypothesis you test it and you're [3:00] wrong and then you move on you know like [3:02] before Galileo we used to think that the [3:05] Earth uh was the center and you know sun [3:07] went around us and then he realized it [3:08] was opposite that was a major Discovery [3:11] back then and they were like oh we're [3:12] going to this whole like Inquisition and [3:14] all kinds of thing that can relate to a [3:15] little bit about that [3:17] um and and so you know I would say going [3:21] forward you know I'm all about finding [3:23] where the problem is uh and solving that [3:27] and right now the problem is not better [3:29] lasers the problem [3:30] is perception a prediction [3:33] prediction it's interesting because you [3:36] talk about the willingness to discard [3:39] things that don't work and yet you you [3:41] really did work on a single premise for [3:44] a decade or so and I'm wondering if this [3:47] is going to sound crazy but was the [3:49] lawsuit and everything that happened [3:51] sort of a blessing did it cause you to [3:54] rethink things or were you already [3:56] starting to rethink that maybe this [4:00] level four first approach isn't the way [4:04] to go yeah so I would think that it's [4:06] hard once you're in the middle of [4:08] something and you're like dedicated [4:09] fully to it it's very difficult to get [4:11] perspective and see [4:13] like you think you're there but actually [4:15] you're missing this other path that's [4:17] going to block you completely so I [4:19] definitely would say that taking a step [4:21] back and you know taking some time to [4:24] reflect on my life and um you know [4:27] things I could do better and and you [4:28] know work started to church Church yeah [4:30] well that's not when the church was [4:31] started but that's uh that was while I [4:33] was at Google in 2015 [4:36] uh but um to think about really what is [4:41] the actual problem and and and be able [4:43] to start from a clean slate is very [4:45] difficult to do at a company [4:47] specifically when you're dealing with [4:48] the day-to-day about like this team and [4:50] that team don't get along and all kinds [4:51] of other problems at that point it's [4:53] very difficult to like really set the [4:55] strategy what you need to do you're just [4:56] focus on executing along that path so [4:59] we'll know if it was a blessing disguise [5:01] or not but uh you know I think I'm [5:04] pretty hopeful about the future and I'm [5:05] optimistic about what we're [5:07] doing so while you have now pursuing [5:11] Pronto there's dozens of companies that [5:13] are still pursuing this level four and [5:16] I'm just wondering in your view now [5:19] um and I wonder how they would respond [5:22] to this but they're not on stage so we [5:23] can't ask them but do you no longer [5:25] believe that level four is achievable [5:28] safely in the short term well so I would [5:32] say is I encourage people to have [5:33] different approaches because that could [5:34] totally be wrong and so it's great that [5:36] other people have different points of [5:38] views and you know what I would say is [5:40] like the true metric for when AVS [5:43] actually get to be deployed is safety [5:46] right so safety is the only metric that [5:48] determines when an AV will come out not [5:50] whether they use lar or not or or car [5:52] maker or software vendor and [5:55] so I don't personally think that [5:57] straight to level four is going to be [5:59] something that is doable because you'll [6:02] need to have a very long time before you [6:04] actually solve the problems and by [6:06] providing technology that evolves and [6:08] gets better over time you'll be able to [6:10] like chip away at all the problems [6:12] people use an example of if you're going [6:14] to the Moon you know building a ladder [6:16] doesn't really help you well the the [6:18] thing is that we're not going to the [6:19] Moon we're just trying to drive to the [6:21] corner store and back and so we [6:23] shouldn't be building you know that's an [6:25] analogy that while you should be [6:27] building the right technology with the [6:28] right strategy it's not necessarily that [6:31] you can't make an evolutionary step to [6:33] get there and I think that the [6:34] evolutionary step will uncover a lot of [6:36] lessons learned that will make it you [6:38] know safer and better so what should we [6:41] do to appease our AI overlords prior to [6:44] the singularity [6:46] apocalypse it's phenomenal question well [6:48] you can sign up for a Udacity course uh [6:50] and learn how to do that uh I'm sure [6:52] there's other classes that that you [6:53] could do as well no I mean ultimately uh [6:56] I think it's a funny question but you [6:58] know progress is an evitable and you [7:00] know you want to um make sure that you [7:05] help create the progress and understand [7:08] where that's going as opposed to uh deny [7:11] it or try to cage it up or you know say [7:15] that it's more dangerous than nuclear [7:16] weapons you know and just be fearful [7:18] basically AI has a huge opportunity to [7:20] connect us in a way uh that you couldn't [7:24] before um and we can build that in a [7:26] fearful way or in a way that has bias uh [7:29] or we can build that in a way that's [7:31] going to be helpful for us so don't be [7:33] scared is the you know threel answer or [7:36] three-word answer