AI 与机器学习2020pubmed.ncbi.nlm.nih.gov
自动驾驶汽车与人类驾驶员：安全水平相当远远不够
Self-driving vehicles against human drivers: Equal safety is far from enough
进阶 · 约 40 分钟读完前置知识：自动驾驶基本概念、风险评估基础、交通伦理与责任
导读
这篇论文讨论自动驾驶汽车应达到怎样的安全门槛，核心论点是：仅仅做到与人类驾驶员"同等安全"远远不够。作者从风险认知与伦理角度论证，自动驾驶需要显著更高的安全裕度才能被公众接受。具体论证方式与数据依据我无法从标题和评论确认。
HN 关注点：的焦点是："同等安全"是否公平（人类基准里包含酒驾、分心驾驶）、多高的安全倍数才够、事故责任如何划分，以及自动驾驶事故中的遇难者与获救者并非同一批人这一伦理与法律困境。
HN 评论精选
这几条评论里：支持 2 补充 2
补充安全数字再好看也绕不开责任与赔偿：最终门槛取决于精算和商业成本，而非统计上的死亡人数。
如果这属于与风险感知相关的主观问题，而非实际的统计风险，那这种感知是可以被引导改变的。挑战依然在于：涉及自动驾驶的事故中总会有人死亡。我们预期死亡人数会更少，也就是"拯救生命"；但死于自动驾驶事故的人，与原本会死于人类驾驶事故的并不是同一批人。会有案件让法院认定自动驾驶系统是肇因，遇难者家属要讨公道，而那些因这套系统获救的人永远不会在同一起案件里被听见。我预计最终这会归结为一个商业决策，并由精算结果来决定：利润和保险能否覆盖为此类案件辩护与和解的成本。谁知道呢，也许阈值是安全 5 倍。
11thEarlOfMar查看英文原文
If it's a subjective matter tied to perception of risk rather than actual, statistical risk, such perception can be swayed. The challenge remains that people will be killed in accidents involving autonomous control. And we anticipate that the number of people killed will be fewer, hence 'saving lives'. However, the lives lost in autonomous accidents will be a different set of people than those that would have died in human driven accidents. There will be cases where a court determines that the autonomous system was the cause. Families of those killed will want justice, while those separately saved by autonomous systems may never be heard from in the same case. I expect that in the end it will come down to a business decision, and that decision will be informed by an actuarial exercise: Will profits and insurance be able to cover the costs of defending and settling such cases. Who knows, maybe the threshold is crossed at 5x safer.
支持人类基准若包含酒驾、分心等违法失误，那要求自动驾驶更严格就合理。
这个基准是否包含分心驾驶、酒驾、毒驾等的全部事故？还是指一个没有故意违法的普通人类司机？如果基准把这些人為失误都算进去，那把机器人置于更高标准我认为没问题。想象一下，如果我们推出机器人警察，它们只以人类同样的比例无故处决黑人。
rozab查看英文原文
Would this baseline include all the accidents from distracted drivers, drunk drivers, drug drivers etc.? Or is it referring to an average human driver who isn't intentionally breaking the law? If the baseline includes all these sorts of human error, I see no issue with holding robots to a higher standard. Imagine if we rolled out robot policemen who only executed black people for no reason at the same rate as humans do.
补充除了驾驶行为，自动驾驶还允许重新设计车内布局，从被动安全角度降低伤亡。
很多评论都在关注如何把车开得更好来提升安全。但有了自动驾驶，我们不是可以让车内布局更安全，从而在事故中减少对车内人员的伤害吗？比如现在因为必须看路，我猜挡风玻璃带来的危险远大于两侧带软垫、乘客像火车车厢那样面对面而坐的布局。看起来很可能即使碰撞次数不变，仅靠改变布局就能让自动驾驶汽车安全得多。
gamerDude查看英文原文
A lot of comments are focusing on safety via driving better. But with self-driving vehicles, can't we make the layout of the car safer, and thus accidents cause less harm to the people inside? For example, right now because we need to see the road, I assume there is significantly more danger from the windshield vs. a padded back on both sides of the car with passengers facing each other like in a train car. It seems likely that we can make self-driving vehicles much safer, even with the same number of collisions, by just changing the layout.
支持公众不会接受与人类同等事故率的自动驾驶，且现有责任体系至少是可理解、可操作的。
这看起来相当合理，而且也很可能实现。允许一项技术上路、犯下和人类一样多的错误，那太荒唐了。如果自动驾驶汽车像今天的人类一样每年撞死三万人，我肯定不会坐。事故责任又怎么划分？我们现有系统远非完美，但至少是人们能理解、知道怎么应对的东西；而且确实有些司机比别人更优秀、更谨慎——所以这不只是控制的幻觉。
spaetzleesser查看英文原文
This seems pretty reasonable and also very possible to achieve. It would be insane to allow a technology on the streets that makes as many mistakes as humans are making. I certainly wouldn’t use self driving cars if they killed 30000 people per year like humans are doing right now. How would you assign responsibility for crashes? Our current system is far from perfect but at least it’s something people understand and know how to navigate. And there are drivers that are better and more cautious than others. So it’s not just an illusion of control.
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2020Hacker News 讨论帖147 分 · 448 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选