PhD dropout. Ex-Story Protocol engineer. Building whatever I find interesting next.
Planet Earth
Joined January 2021
- Tweets6.7K
- Following1.3K
- Followers820
- Likes21.7K
Pinned Tweet
我把 Haggai Roitman 的 The Hitchhiker’s Guide to Agentic AI: From Foundations to Systems 翻译成了简体中文：
《智能体式 AI 漫游指南：从基础到系统》
这是一本系统介绍 Agentic AI 的书，从 LLM 基础、Transformer、GPU 与分布式系统、RL、PPO / DPO / GRPO、奖励模型、SFT、大推理模型，一直讲到 RAG、记忆系统、agent orchestration、MCP、A2A、多智能体系统、开发框架和 Agentic UI。
这份译本经原作者授权，依据 CC BY-SA 4.0 发布。它不是官方中文版；如果涉及关键技术细节，仍建议以英文原文为准。
我希望这份中文译本能帮助更多中文读者系统地理解 Agentic AI，也欢迎大家通过 GitHub Issues 反馈错别字、术语建议、漏译或排版问题。
Authorized Simplified Chinese translation of The Hitchhiker’s Guide to Agentic AI.
Bo Wu retweeted
Data in the Wild, ep. 004.
Task context
• Clean several pairs of shoes, then arrange by type + use frequency
• 11 minutes
• A home in India
Capture
• Handheld monocular Realme RMX3867
• 1280×720, H.264
Embodied behavior
• Repeated handling across shoes, cloth, cleaner + storage
• Actions: wipe_shoe, rotate_shoe, apply_cleaner, change_pair, carry_shoes
• Challenges: object identity, state tracking, tool changes + long occlusions
Video sourced from Numo. Numo collects real-world data that Poseidon refines into training-ready datasets.
Long-horizon tasks require more than recognizing an action frame by frame. Models need to track which objects they’ve interacted with, how their state has changed, and what still needs to happen.
Bo Wu retweeted
Data in the Wild, ep. 005.
Real-world tasks rarely happen from a fixed viewpoint. This one captures manipulation in a constrained space, with frequent camera movement, occlusion, and re-entry as the task progresses.
Task context
• Clean one area of a car interior + return supplies
• 15 minutes
• A parked car in residential United Kingdom
Capture
• Handheld monocular Infinix X665B
• 1280×720, H.264
Embodied behavior
• Cloth, bucket, dashboard, controls, storage, console + seats
• Actions: rinse_cloth, wring_cloth, wipe_dashboard, wipe_control, open_storage, wipe_console, wipe_seat_gap, wipe_sill
• Challenges: constrained reach, shifting viewpoints + repeated exits
Video sourced from Numo, which collects real-world data that Poseidon then refines into training-ready datasets.
Bo Wu retweeted
See you at @corl_conf 2026.
Our paper, "When Crowdsourced Egocentric Video Goes Wrong: Failure Modes, Metadata, and Quality Control," was accepted to the "Oops, I Erred" workshop at CoRL, a leading robotics research conference.
If you'll be in Austin, say hi: @sebliu + team will present our research on a critical challenge for Physical AI: how to ensure crowdsourced, real-world video data is reliable and useful for training robots.
corl.org/
Apple in an alternate timeline where Jony Ive never discovered aluminum.
Introducing the Mac Studio G8.
Same power. More jelly. 🫐🍇🍊🍏🍓
Huge thanks to @Cloudflare for hosting us on Friday. They got a suite at Chase Center for Warriors vs Grizzlies and invited the @psdnai team along. Great game, great company, way too much AI talk between possessions.
Warriors didn't get this one, but anyone who trains models knows one game is a sample size of one.
I also couldn't stop seeing the court as a dataset. Ran an off-the-shelf detector on one phone photo: 10 players, 3 officials, 1 ball, every frame of every game. That's the lesson we live with at Poseidon. The model is the car, real world data is the fuel.
And that fuel needs somewhere to live. Our raw hand-task video sits in R2 and some of our QC runs on Workers. For an AI company doing real research on real data, Cloudflare stopped being "the CDN" a while ago. It's core infra.
What's the best real world dataset you've spotted somewhere you weren't supposed to be working?
Bo Wu retweeted
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
trillion of pictures you can choose from to promote speed limit, and you choose this one.
City of Toronto
New 30 km/h area speed limit signs will be posted at neighbourhood entry and exit points. Once you enter a signed area, the 30 km/h speed limit applies throughout the local road network until you leave that area.
Slower speeds help reduce the risk of serious collisions and make neighbourhood streets safer.
Learn more about the change: toronto.ca/LocalSpeed
ALT Two men walking and chatting on a sidewalk, one pushing a baby stroller near a white picket fence and brick house.
A Fruit Fly / A Dying Light
All gone?
All gone.
Why is that?
Well, sometimes there's a rationale behind it. Sometimes there isn't.
K, then what did you do exactly?
No clue, believe it or not.
A bend along Highway 1. Warm sand, greenish-blue water, 微风拂面. Cozy. It all made me extremely drowsy.
I took out the shovel.
Not a regular shovel you can buy on Amazon, you dumbass.
It was a shovel made of titanium, with a titanium coating.
I laughed. Funny, I thought. Who would put a titanium coating on a titanium shovel?
Silly, but funny.
The small cave in the sand was ready.
I'd been thinking of digging a small one, just big enough to put all the beautiful/bad/ugly/sweet memories in.
The shovel was shining, and 给力.
But still, it took me years to build this shovel, with my own hands, bit by bit.
Not perfect, even 简陋, but hey, I built it as best I could.
You know what?
The cave was more comfortable than I'd thought.
At one point, lying there, I thought: I REALLY, REALLY don't wanna get out.
Warm sand is like a blanket, warmed by a whole day of sun.
Warm, with the smell of the ocean.
But I had to get out.
It had been killing me for the past few years.
It was simply driving me crazy.
Is that so? A fruit fly whispered near my left ear.
YES, IT IS LIKE SO.
She smirked, without another word.
I turned my head to the right.
I mumbled, if there were no "fruit" in your name, you wouldn't look so charming, you know that?
I know. (Proudly.)
You'd look even sexier if I put the word "burning" before your name.
Burning? Wdym?
Burning Man.
Oh...
Well, I'll get burnt eventually, like all the others, when we reach our destination, expected or unexpected.
But have you reached your destination yet?
Well, it's about the journey, right?
Touché.
Touché indeed. I took the last sip of a cold beer.
It was the last IPA of the six-pack.
Relieved, I dusted the sand off my pants, making sure no one would notice this was where I'd buried it.
You should burn it.
As a fruit fly, you know too much, you talk too much, and you suggest too much.
A flicker of emotion crossed her face. Dead silence.
I turned around and, without hesitation, started the engine. The Z4 roared.
There was no way I would come back, no way.
Data, not architecture, is the bottleneck for low-resource languages.
Proud of the team for shipping 353K+ fully transcribed recordings for Indic ASR!!
How much speech data is enough to build reliable ASR models?
With Numo Indic Speech, we're adding 3,845+ hours of transcribed audio across Hindi, Telugu, Tamil, and Bengali.
That's roughly 6.3x the labeled hours in Kathbath across the same four languages, giving researchers greater depth for training and evaluation.
New #halloween theme pushed to $TSLA is very nice
(It was triggered when we parked our car in a dark underground parking lot
