Nishi

Physical AIするぞぅ💪 wandb.connpass.com/event/388863/

My visit to SF is coming to a close this week. I spent the last month meeting with robotics builders and founders. Here's a dump of my takeaways: 1. Hard to imagine there's a higher density of robotics talent in any other US city. 2. The vast majority of founders are hyper focused on technical problems as opposed to customer/market ones. 90% of "products" are just tools for researchers. 3. World models. 4. 99% reliability is still an elusive metric but 90% is reasonably within reach. 5. Large labs are starting to build out ecosystems of deployment partners (eg @physical_int & @weaverobotics @Ultraroboticsco). Big first mover advantages here to get the data flywheel going with real-world deployments. 6. Reliable and affordable arms are still scarce. YAM Ultra (with active cooling) overheats making it untenable for prolonged service. Expect a literal arms race. 7. Teleop with leader arms is going to be obsolete by EOY, if not already. 8. The case for human-like hands as end effectors grows stronger by the day in order to get the most out of egocentric human data (the most scalable data collection format by far). 9. Software engineers love robotics and many existing tools aren't built for them, a great place to search for worthwhile problems to solve cc @mexitlan 10. There's still hype around "general purpose" robots but vertical robotics startups are finding traction in narrow use cases and will build fantastic businesses.

@yukimamax 今からだとおすすめはどこですか？

Claude Code 向けの decomposition コマンド🎉 Plan Mode から実行するときにTaskを細かく分割してくれる君 より規模のデカめなタスクを投げる時に、Claude Code の実装グリップ力を補完してくれる。（既存だと、分解が足りないw github.com/fumiya-kume/cl… 頑張って作ったのでRTして拡散してほしい！