the most interesting find for me was around systems -- latency and performance matter a lot more than the average observer might think often improving task completion directly correlates w improving things like latency, not just scaling training

cash flow positive on my account?

day 2-3 built a gaussian splatting pipeline from scratch based on ObjectGS. Initially, outputs kept converging to one grey blob (see below) so I spent a lot of time debugging

day 1 of documenting my project I'm going to create a near-perfect reconstruction of my apartment + learn the all the individual objects within it last night I took some videos on Record3D of my bedroom & reconstructed a sparse, filtered point cloud follow along this thread

The future shouldn’t be driven by a fixed set of rules encoded in the latent space that updates every 6 months. Models must learn as fast/faster than people.

If the rate of change of a foundation model is gated by {more compute, more data, more evals, more engineering hours}, then its ability to change is limited to the speed of the organization.

I want the world to be more dynamic than a foundation model can fundamentally provide.

becoming a pro wechat user so feel free to add me there too: aptni27

Our plan is to write another deep dive on the chinese robotics ecosystem and are currently chatting with folks across (resource, batteries, actuators, sensors/cameras, pcb, CMs/JDMs, full-stack humanoids, and foundation model labs)

I am heading to SZ/HZ at the end of the year to meet a bunch of robotics, hardware, and manufacturing companies across the entire stack. If you or anyone you know can help connect me with folks in the ecosystem, I'd love to chat.

that stone was shenzhen, in reference to their special economic zone created by deng xiaopeng. it was a window into the outside world/technology and represented freedom for those willing to move.

cross the river by feeling for the stones

two minutes of magic 🪄 eight years of obsession robot that changed everything 👇

<lightning.jpg> Retiring this guy for the time being. Served me well for the past 10 years. Generated in cs class in high school under the legendary @gcschmit

My intuition is that BotCo will focus heavily on BOM bc in a world where robots aren’t that useful, price point will be a huge sticking point for consumers. Price point + UX will probably be more important than actually task performance to kick off the GTM flywheel for this

📍

270,000 hours of fine-grained manipulation data is absolutely insane. One metric I'd be interested in is Generalists hrs/hr (how much data can they produce per working hour). This could tell us a bit about how they got to 270k so quickly (if it's all real-world/not sim).

Clipping is critical in all areas of RL, but intuitively feels incorrect. I wish there were more beautiful ways to handle this.

https://t.co/KITiDT3SHI Even streaming actions over BLE sounds pretty hard if you wanted to make a robot do something useful while running blind.