Peng Su

Building Physical AI.

[Andrei Bursuc @abursuc]

The #WorldEngine tech report is now up!
It's our post-training framework to deal w/ scarcity of long-tail safety-critical scenarios.
Post-training strategies for AD are rarely tackled and discussed in the open. We hope that this will open more this area.
https://t.co/5E8Ta131xa

[皓樂芒 @howlemont]

这篇 arXiv 的《Dive into Claude Code》，感觉最有用的地方是讲清楚了，其实Claude Code这种系统一旦进入真实环境，工程重点立刻就变的很现实。

Claude Code当然是个 coding agent，会跑 shell、会改文件、会调外部服务。 模型循环本身并不复杂。

文中直接写了Claude Code的核心就是反复调用模型、执行工具、再根据结果继续往下走。

论文里整理得出，真正占分量的是围在外面的那一圈系统设计:

权限控制、审批模式、上下文压缩、插件和 skill 的扩展面、subagent 的委派、会话持久化。

分析结论：

今天决定agent 产品成败的，是“怎么在真实世界里动手”，而非“会不会想”，

很多人谈 agent，还停留在 prompt、loop、planning 这些表层词汇。

Claude Code 这类产品越来越在意的一点却是一旦真的要进项目、进终端、进文件系统，马上就碰到很现实的问题。

文件能不能碰。
命令能不能跑。
风险怎么分级。
用户什么时候该介入。
上下文太长了怎么办。
会话断了怎么接回去。

这些问题不花哨，但恰恰这些才是生产系统的骨架。

文章给 Claude Code 提炼了五个出发点，

-人始终保留决策权，
-系统要守住安全和隐私，
-执行要稳定，
-人的能力得要真的放大，
-还得慢慢适应不同用户的项目环境和信任关系。

Claude Code 的设计品味清楚得很。就是没有把大量精力放在“用显式状态机把模型一步步管死”这种方向上。
更像是在给一个越来越强的模型准备工作现场，让它有护栏，有工具，有恢复机制也有边界。

还有就是把 context 当成主约束来看。

很多人总觉得上下文管理只是优化项，模型更大一点就好了。

但这篇分析的判断却是：context 根本就是成本、稳定性和任务质量的交叉点。如果带着错的上下文开工，后面会越跑越偏。

所以压缩、筛选、延迟加载也是主工程，不是边角料。

跟OpenClaw对比这一段也挺有意思。
同样一组问题，放在不同产品形态里，答案会完全不同。没有什么谁比谁更先进，

CLI 里的 coding agent和一个多通道的 assistant gateway，天然就不会长成一样。

权限边界不同，扩展方式不同，运行时位置不同，整套架构的重心也会变。

所以这篇文章帮助你更确定了一件事，那就是agent architecture 说到底是在设计一套运行秩序。

模型就像发动机。能不能上路，靠的是刹车、方向盘、仪表盘和路权规则。

虽然基于公开源码和公开资料，这篇第三方分析终归不是 Anthropic 自己写的官方架构说明。
拿来理解设计思路，很有价值。
但不能拿来把所有细节都当定论。

Claude Code 这种系统，已经很会替人把事情往前推了。 短期能力放大，这点没什么好怀疑的。

最后留下的问题在于，系统做得越多，人有没有因此理解得更深，沉淀出一些经验，代码库有没有变得更清楚，

随着工具在变强，人是不是也在一起变强，

这些现在离被彻底解决还尚早。

https://t.co/8TgvjNY56Y

[Hongyang Li @francislee2020]

Introducing #WorldEngine, https://t.co/UK3vsulS2u, a two-year long project. The missing infrastructure for Physical AI post-training in Autonomous driving. Open-source. Production-validated.

[Peng Su @jayden_so]

@JingxiangSun42 Nice work, Jinxaing. Are you interest in building L5 self-driving agent at industry-level dataset , i.e.~1 M fleet .

[Peng Su @jayden_so]

@huajian_xin @EdinburghUni @WendaLi8 Great. Are you interest in building L5 self-driving agent at industry-level dataset , i.e.~1 M fleet .

[Peng Su @jayden_so]

@ZhengYinan2001 Nice work, Yinan. Are you interest in developing real self-driving system at industry-level dataset ,i.e.~1 M fleet .

[TSLA99T @Tsla99T]

实用却不烧脑，值得周末静下来慢慢看

How I use LLMs https://t.co/NIm5gDh2ZI via @YouTube

[Peng Su @jayden_so]

2024华为智能汽车解决方案发布会 | 2024 Huawei Intelligent Automotive Solutions Launch ... https://t.co/RfRg93J9FF via @YouTube

[Yue Wang @yuewang314]

Introducing EmerNeRF, our answer to the challenging dynamic NeRF in-the-wild problem. EmerNeRF is the best-ever project I've got involved in, led by our only @JiaweiYang118 (stay tuned for his even more impressive works), in collaboration with @NVIDIAAI colleagues @iamborisi @orlitany @xinshuoweng @seungkim0123 @Boyiliee Tong Che @danfei_xu @FidlerSanja @drmapavone. EmerNeRF is a spatial-temporal scene representation for autonomous driving and robotics. It learns static-dynamic decomposition and flow estimation, all from self-supervision. No magic, all you need is a good combination of static, dynamic, and flow fields parameterized by instant NGPs.

Paper: https://t.co/aA5GXvbAsX
Website: https://t.co/4LjZSQBgNb
Code: https://t.co/K36ePoNx85

EmerNeRF can perform high-quality scene reconstruction and scene decomposition.

We also visualize the learned 3D voxels with flow (check out our website for more visualizations).

In addition, EmerNeRF lifts foundation model features to 4D space-time, to enable a range of semantic tasks. In doing so, we discover that self-supervised trained VITs exhibit unpleasant positional embedding noises (in line with the Register paper from Meta FAIR). We design a feature fitting module to denoise the 2D foundation model feature maps for better feature lifting.

We also introduce a NeRF On-The-Road (NOTR) dataset (a derived subset of the Waymo Open Dataset) to facilitate future research in this domain.

[Nuro @nuro]

We've witnessed the amazing power of LLMs to accelerate everyday tasks. But, how does that translate to autonomous vehicles?

As an AI-first autonomous technology company, Nuro has invested in state-of-the-art research to apply LLMs across our AV stack.

[42HOW @42how_]

The sales of Wenjie's new M7 have made a major breakthrough, reaching 80,000 units.

This data is somewhat leading by a wide margin, isn't it?

[Peng Su @jayden_so]

New Huawei technology keeps drivers safe https://t.co/XnJArmVQUV

[Peng Su @jayden_so]

AITO M7 Launch Ignites Buzz: Huawei’s Advanced Autonomous Driving Goes Nationwide https://t.co/BRMdouk0I0

[42HOW @42how_]

"Avatr announces a phased roll-out of its Autonomous Driving (NCA) nationwide, progressing from 6 to 16 cities across China, including Beijing, Shanghai, Guangzhou, Chongqing, Shenzhen, and Hangzhou, by the end of the year."

[42HOW @42how_]

How Did AITO's Surge in M7 Pre-orders Shock the Chinese Car Market?
Link → https://t.co/R2pl9a1Tzs

[Peng Su @jayden_so]

Cool!

[Tesla Optimus @Tesla_Optimus]

Multiple fully Tesla-made Bots now walking around & learning about the real world 🤖

Join the Tesla AI team → https://t.co/dBhQqg1qya

[AI at Meta @AIatMeta]

Code Llama is free for both research and commercial use and we've made three different models available:
- Code Llama
- Code Llama - Python
- Code Llama - Instruct

More details on each of these models and how you can download them ➡️ https://t.co/yLBEKVJhU5

[MIT Technology Review @techreview]

Today in The Download, our daily newsletter:

🚗 Using chatbots to make driverless cars smarter
💉 Updated covid vaccines
🩻 How AI can help screen for cancer https://t.co/gRssYCmxeR

[ThinkerCar @thinkercar]

The #AITO New M7 incorporates Huawei's developed GOD (General Obstacle Detection) and RCR (Road Cognition Reasoning) technologies, making it the first model in the AITO brand that does not rely on high-definition maps.