Andrew Wang

@generalsequence - Infrastructure for General Physical AI

Building General Sequences | Master & Undergrad at @Penn | Noe-matrix Robotics| WM Research Fellow

[Andrew Wang @Andrew_WangI]

Distributed Cloud Inference is the Future for Physical AI

BENCHMARK results from @generalsequence

Physical-AI models are moving to the cloud: bigger models, more of them, running for real robot deployments. Serving them well turns out to be a distributed-systems problem, not just a modeling one. Here's what the data shows.

Latency is a deadline and we can solve it.
Decompose a single robot-inference call and the model is a small slice: on an open policy like pi0.5-DROID, ~106 ms. A naive cloud path (Runpod, etc.) spends ~1,494 ms per call, dominated by connection setup, observation upload, and routing overhead. Optimize the path and it drops to ~144 ms (a 10.4× reduction, only an additional 38ms) from our networking optimization, which holds across models (pi0.5, Cosmos3-Edge, GR00T, DreamZero). (See Image 1 and Image 2)

And regional coverage works.
Your customer might carry your robot around the US and things gets unstable. By routing it to the closest data centers with the multiple neocloud providers we support, we provide low latency and reliability on the table. (Image 3)

@generalsequence believe that robotics inference should be distributed, routed across providers to the nearest, warmest GPU, with an optimization layer on top.

This is exactly what we've providing as a platform service. If you're running inference at scale and across the US, contact us.

Shout out to our investor @bcristei from OPTIM VC for making this possible.

[Andrew Wang @Andrew_WangI]

@shoucccc ��是内网access，有ssh也没用啊，也就只能偷偷别人apikey

[Andrew Wang @Andrew_WangI]

@HanPaoao 这个其实是风投

[Andrew Wang @Andrew_WangI]

definitely agree, the next wave of papers will be agent + physical ai in all kinds of application scenarios.

[ClaraChengGo @ClaraChengGo]

Purely atomic task manipulation is one thing. But VLM calling action expert to work is another.

As Open AI and Anthropic brought up the wave of the second approach, there will be more labs publishing how VLMs expand the agentic power to physcial world (Gemini, Llama, Qwen, Kimi, MiniMax etc).

[Andrew Wang @Andrew_WangI]

very interesting, we do believe llm will act as the "brain" for robotics, but executor-wise, the 20x speeding of the video is concerning..

[Jay Chooi @chooi_jeq]

GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. 🧵

[Andrew Wang @Andrew_WangI]

Launching our pilot!

[General Sequences @generalsequence]

Robot foundation models are getting good.
Running them is still the hard part.

General Sequences is a router HTTPS endpoints for all of it.

Six models. Per-robot-hour pricing. No GPU.
https://t.co/CSLwzedl9T

[Andrew Wang @Andrew_WangI]

@MaxForAI 一般出现这种词 “人类黄金时代” 大概要马上要完了

[Andrew Wang @Andrew_WangI]

@chenreason b2c是这样的

[Andrew Wang @Andrew_WangI]

@qianWuyun 在做DataPlot，丢一个短剧链接,秒出整部剧情提示词万行。给做 AI 短剧的团队用。接合作

[Andrew Wang @Andrew_WangI]

@msjiaozhu DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@wangzaibibibi DataPlot —— 我们在做，丢一个短剧链���,秒出整部剧情提示词。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@GrayNoteLab DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@williamlab DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@ZhanweiC DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试！

[Andrew Wang @Andrew_WangI]

@wanerfu DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@BTCqzy1 DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试！！

[Andrew Wang @Andrew_WangI]

@jiroucaigou DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@NoraX2026 DataPlot —— 我们在做，丢一个短剧链接,秒出整部剧情提示词，结构(钩子、反转)。给做 AI 短剧的团队用。DM 一个链接试试

[Andrew Wang @Andrew_WangI]

@wangray 信这个的这辈子有了

[Andrew Wang @Andrew_WangI]

pretty powerful to me, Claude Workflow