首页 > 失业生 > 正文

同桌的你

Edmunds recommends the five best SUVs for under $40K_我的网站

阿里巴巴

A |     贾浩楠 发自 副驾寺     智能车参考 | 公众号 AI4Auto          ChatGPT能记住你的聊天历史,知道你的表达习惯,甚至可以根据过去的交流调整回答方式。    The automotive experts at Edmunds test SUVs of all shapes, sizes and prices. During these evaluations, certain models rise to the top with qualities such as roomy seating and cargo areas, a comfortable ride, and easy-to-use touchscreens and technology features. But with many prospective buyers looking to make every dollar count, it’s important to consider value as well.Working with a budget cap of $40,000, Edmunds picked five SUVs to highlight. Some have city-friendly proportions that take the stress out of parking in tight spots, while others have cavernous cargo holds and handy third-row seating. They are listed below in order of price, which includes the destination charge.2024 MAZDA CX-50The Mazda CX-50 small SUV earns a spot in this list thanks to its driving refinement and upscale cabin. From inside it’s not hard to believe that you’re behind the wheel of a luxury-badged model. The design is first-rate and the seats are supportive in either row. The CX-50 rides a bit more firmly over bumps than other SUVs, but in return it has stable handling and is enjoyable to zip around town in or take out on a scenic drive. All-wheel drive is standard on every trim, as are driver assist features such as lane keeping assistance, traffic-adaptive cruise control and blind-spot warning. A 186-horsepower four-cylinder serves as this Mazda’s capable and budget-minded base engine. There’s an optional turbo four-cylinder that’s more powerful, but it nudges the price at or above a $40,000 spending limit.Starting manufacturer’s suggested retail price: $31,6752023 KIA SORENTOThe Kia Sorento is a compelling choice for shoppers wanting a low-priced seven-passenger SUV. Its third-row seating area is smaller than those of bigger models — such as the Hyundai Palisade listed below — but it is still suitable for occasional use for kids. Most of the Sorento’s trim levels come well stocked with helpful convenience and technology features.There is a choice of a fuel-conscious, non-turbocharged 191-horsepower four-cylinder engine or a punchier 281-horsepower turbo-four available in higher trims. And it’s possible to opt for the more powerful engine and keep the Sorento’s price well below $40,000. A Sorento Hybrid is also available for shoppers seeking higher fuel economy. Starting MSRP: $31,4152024 HONDA CR-VThe Honda CR-V is about as well rounded as a small SUV can get. It delivers a smooth and comfortable ride, great outward visibility, and loads of cargo space. Fuel economy is another plus point. The CR-V’s base 190-horsepower turbocharged four-cylinder gets up to an EPA-estimated 30 mpg combined, or you can opt for the 204-hp hybrid powertrain that gets up to 40 mpg combined. Thanks to its recent redesign for the 2023 model year, the interior of the CR-V is attractive and functional. Picking the midgrade EX-L trim stays under our budget limit and still gets you some upscale features such as leather upholstery, a power liftgate and a bigger center touchscreen. Starting MSRP: $30,8252024 SUBARU OUTBACKThe Subaru Outback is ideal for SUV shoppers who want a vehicle that’s suitable for occasional off-road adventures. Thanks to its standard all-wheel drive and generous ground clearance, the midsize Outback has above-average capability on dirt roads or rocky trails. The Outback’s wagon-like shape also makes it easier to load recreational gear or cargo on the roof.Acceleration from the Outback’s standard 182-horsepower four-cylinder can be slow when you need a burst of speed on the highway, but in exchange the Outback gets a respectable EPA-estimated 28 mpg combined. A more powerful turbocharged four-cylinder is also available, though it exceeds our budget cap.Starting MSRP: $30,1902024 HYUNDAI PALISADEThe three-row Hyundai Palisade is the most expensive SUV among Edmunds’ choices here, and the majority of trims are priced beyond a $40,000 budget. But it’s important to know even the base Palisade is equipped with a long roster of tech features and driving aids. Highlights include a big 12.3-inch center touchscreen, first- and second-row USB ports, traffic-adaptive cruise control and blind-spot warning.All Palisades come with a 291-horsepower V6 engine that delivers acceptable acceleration and fuel economy. Despite its relatively large size, the Palisade is relaxing to drive and easy to maneuver. Also worth noting is its third-row seating has ample space even for adults, which is a rarity among many midsize SUVs. Starting MSRP: $37,735_______This story was provided to The Associated Press by the automotive website Edmunds.Nick Kurczewski is a contributor at Edmunds and is on Twitter and Instagram。         现在,这种“记忆能力”开始进入物理世界。         马斯克一条推文,透露了Tesla FSD正在发生的新变化:          FSD开始拥有“千人千面”的能力。         从停车偏好,到驾驶干预,再到长期驾驶习惯,FSD正在尝试记住每一个人的不同。         从“一个AI服务所有用户”,走向“每个人都有自己的AI”。

B |          马斯克确认:FSD开始拥有“长期记忆”          事情起源于马斯克最近两次公开回复。         6月的一次,是关于停车习惯。

C |          有用户提出,希望FSD能够记住自己在固定地点的停车方式,比如回家停哪里、去公司停哪里。         马斯克回应:          FSD很快会记住用户的停车偏好。         最近的这一次,更加关键。         有用户反馈,FSD在一些驾驶决策上不符合自己的习惯,比如车道选择、变道时机等。         马斯克表示,未来车辆会记住用户具体的接管行为,并逐渐匹配每个人的驾驶偏好。         两次回复本质上是同一件事:          FSD开始从一次性的驾驶任务,走向长期理解驾驶员。         过去,自动驾驶面对的是一个抽象用户。         无论谁坐进车里,背后的模型都是同一个。         它知道:这里是什么道路,前方有什么车辆,应该如何规划路线。         但它不知道:你喜欢提前变道,还是宁愿多等几秒;你喜欢跟车近一点,还是更关注舒适;你回家时习惯停在哪里。         这些属于人的驾驶习惯,一直没有进入自动驾驶系统。         而现在,Tesla让这些信息成为AI决策的一部分。         和之前我们谈论FSD不同的是,“长期记忆”不是让FSD重新学习“怎么开车”,这件事已经由基础模型完成。         它学习的是:在安全范围内,你希望它如何开。         这也是Tesla第一次公开走向Personalized FSD:          从“一个模型服务所有司机”,走向“基础模型+个人驾驶偏好”。         为什么是现在?          “千人千面”“越开越懂你”,智能汽车行业其实喊了很多年。         但真正落地很难,用户体验完全依赖车企OTA的“自觉性”和能力高低,系统更新后体验下降的事也不罕见。         因为截至目前,绝大多数自动驾驶玩家竞争,还停留在模型、数据层面,本质上都是在解决:如何让AI成为一个更好的司机。         而特斯拉在V14完成端到端重构之后,开始率先探索另一个问题:          如何让AI成为你的专属司机。         这背后,对应的是近几年自动驾驶领域兴起的几个研究方向:          Driver Modeling、Preference Learning以及Human-in-the-loop Learning。         最早出现的是Driver Modeling,也就是“驾驶员建模”。

D |          简单来说,就是给每个驾驶员建立一个数字画像,比如:有人开车更激进,喜欢更短跟车距离,喜欢更快完成变道;有人更保守,倾向提前减速,更少主动超车。         传统自动驾驶系统会把这些差异忽略,默认存在一个“平均驾驶员”。         而Driver Modeling希望解决的问题是:AI应该知道,坐在驾驶位上的到底是哪一种司机。         技术上,这类方案通常会从历史驾驶数据中提取特征,比如方向盘操作、油门刹车变化、车速变化、变道频率、跟车距离。

E | 然后通过一个额外的模型,生成驾驶员特征向量(Driver Embedding)。

F |          这个向量不会改变基础驾驶能力,而是作为规划系统的输入,影响后续决策。         但Driver Modeling仍然存在一个问题:它更多是在判断“你是什么类型的人”,而不是理解:“你具体喜欢什么”。         于是近几年出现了Preference Learning,也就是偏好学习,它关注的不再是这个司机属于激进型还是保守型,而是这个司机在具体场景下更喜欢哪个选择。例如同样遇到慢车,方案A是提前变道,提高效率;方案B是保持车道,更加舒适。

G |          AI通过长期观察用户选择,逐渐学习这个特定用户更倾向哪一种。         这与大模型的Preference Learning逻辑非常类似,模型能力是基础,用户反馈决定个性。         而Tesla这次提到的Intervention,其实更接近这一方向。         当然还有第三个方向,是Human-in-the-loop,它解决的问题是如何让人在自动驾驶学习过程中持续参与。         传统机器学习通常依赖大量标注数据。但自动驾驶里,真正有价值的数据,往往不是正常驾驶,而是:AI不知道该怎么办的时候。         特斯拉过去大量收集车队数据,本质上就是利用全球车辆作为数据采集网络。而未来,如果把这些反馈进一步和个人账号绑定,那么数据价值会发生变化:          它不只是帮助Tesla优化一个统一模型,还可以帮助车辆理解某一个用户。         Tesla目前没有公开FSD个性化功能的具体架构,但结合FSD V12之后的端到端路线,可以做一个合理推测:单独训练一个“每人一个模型”的可能性最低。         原因很简单,如果每个用户都有独立模型,那么会导致训练成本巨大,更新困难,也会破坏Tesla统一模型的优势。         更符合FSD现有体系的方案,可能是:基础端到端模型 + 用户偏好记忆层。也就是底层模型负责视觉理解、道路判断、安全驾驶、通用规划。         上层增加一个轻量化的User Preference Layer,记录停车习惯、驾驶节奏、路线偏好、历史接管行为。         基础模型负责“能不能这么开”,偏好模型负责“你是不是喜欢这么开”。         这其实和今天大模型的发展路径高度类似:GPT不会因为每个人重新训练一次。而是通过Memory、User Profile等机制实现个性化。         不过,真正的难点也在这里:个性化不能突破安全边界。AI要明白不能学习你喜欢超速、喜欢危险变道等等行为。         这也是特斯拉此次探索最值得关注的地方:它不是让AI变成另一个你,而是“在安全约束内,越来越像用户”。         AI汽车进入“千人千面”时代          说过去几年,自动驾驶竞争的核心是:谁的数据更多,谁的模型更强,谁的能力更接近人类驾驶员。         但特斯拉的Personalized FSD一出,下一阶段,竞争可能会发生变化。         今天,大多数智能汽车仍然依靠统一策略服务所有人。

H | 但未来,AI司机可能会越来越像真人司机,不是因为它拥有情绪,而是因为它了解你的习惯。         这意味着,User Model可能成为自动驾驶新的能力层。         马斯克一石二鸟,解决了两个问题,首先是再次给智能汽车新卖点指出方向,而且是避免高度同质化、低效内卷的路径。         在车位到车位自动驾驶逐渐普及之后,量产车上的AI司机如何建立差异,会成为新的问题。         所有车辆都能完成基础导航和智能驾驶,那么下一步比拼的就是:谁更懂你、谁更符合你的驾驶习惯、谁更像一个长期陪伴你的司机。         当然,对技术能力的要求更高了。         尤其是对于“one for all”的自动驾驶供应商来说,可能很难腾出资源去开发这样的功能,而即使他们愿意开放权限给车企在基础模型上自由开发,有代码级AI能力的OEM却寥寥无几。

I |          所以,利好全栈自研玩家!          其次,之前有一句话把老马问道语塞:特斯拉造车卖车的模式,跟自动驾驶追求的目标,是不是本质上矛盾?          完全自动驾驶普及后,好像的确没有再买私家车的必要。

J |          所以,Personalized FSD也给马斯克解了一个燃眉之急——Robotaxi落地如火如荼,但私家车,以及把AI司机搭载在私家车上依然有价值、有必要。

K |          毕竟Robotaxi用一个AI服务所有用户就是最优解,不可能给出千人千面的高级定制体验。         Personalized FSD路径,其实与生成式AI的发展非常相似。         ChatGPT通过Memory变得更加个性化,AI Agent通过长期记忆完成复杂任务。         未来,汽车作为连接物理世界的重要入口,也可能拥有自己的长期记忆,当然,这并不意味着自动驾驶已经进入成熟阶段。         Tesla的具体实现方式、安全约束机制,以及个性化能力的边界,都还有待进一步验证。         但至少,一个趋势已经出现。

L |

Current article:http://b9p.xuanliaruamibinmilangca.shop/list_ju1fo0j/hm8u0.html

Published on:03:42:04