🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

GPT-4.5 is the wall and also the most underwhelming release of all time. OpenAI is serving the purpose it was intended to back in 2015: it was meant to fund, accelerate, and disseminate the benefits of AGI to humanity. Today, OpenAI research has the world’s worst business

It’s becoming popular opinion among researchers (in many domains I’ve talked to) that there is no business model in foundation models. The model layer will get commoditized, as it has been happening with OpenAI, hence the poor business fundamentals, hence the purpose of the

Deep tech is tricky because product informs tech, but tech constrains the product's solution space. It requires forecasting long-term research progress and straddling product needs until both converge. PBRs are an algorithmic research problem but are finally becoming usable 🙌

Despite great advances in learning dexterity, hardware remains a major bottleneck. Most dexterous hands are either bulky, weak or expensive. I’m thrilled to present the RUKA Hand — a powerful, accessible research tool for dexterous manipulation that overcomes these limitations!

Closed-loop robot policies directly from human interactions. No teleop, no robot data co-training, no RL, and no sim. Just Aria smart glasses. Everyday human data is passively scalable and a massively underutilized resource in robotics...More to come here in the coming weeks.

I think the most interesting insight from EgoZero is the tradeoff between 2D/3D representations in human-to-robot learning. 2D inputs (images, VLMs) scale and encode strong visual priors—but mapping these to 3D actions is hard and data-hungry, especially given the human-robot

In AI, better models → higher productivity → higher margins → more output + capital deployment. In robotics, we often view outcomes as binary: solved or not. But if we instead frame partial success rates as boosting physical labor productivity, we can potentially increase

When we think of human data collection, we usually think of recording accurate 3D proprioception, but this alone does not contain force information. On the other hand, teleoperation data collects the human's input as action labels, so robots trained this way can learn how much