The new 1X clip is a compilation of super quick cuts with no indication of what parts are teleoperated (non-zero per the twitter thread.) Maybe it’s a capable hardware platform but this is not a serious robotics demo. Hope they will share more soon…
Doesn’t have to be ZIRP. As long as FAANG has more cash than they can spend on short term product development they will invest in research.
Incredible demo reel from Unitree!
youtu.be/X2UxtKLZnNo?...
Not saying that, just that it can distort incentives (and we don’t have a lot of insight into the o3 training process.) A long and varied list of AGI benchmarks is needed.
Every field has a deluge of low impact papers, that’s just academia. Imagine if the high impact papers came once every 5 years instead of a couple months.
I think it’s actually fine to argue that any AGI benchmark that comes with a training set will not survive Goodhart’s Law
Genesis has a lot going for it: delightful dev experience, clean implementation, many useful features and solid performance.
And yet, the benchmarks in their announcement are bogus, and the genAI features they put front and center are vaporware.
Not a good look.
github.com/Genesis-Embo...
Just my opinion here. It seems plainly untrue that most tracking applications are “negative”. e.g. every car on the road should have pedestrian tracking. Also strange that you call out quadrupeds as primarily military, that is the least developed of any real quadruped use case right now.
It happens. The only cure for impostor syndrome is knowing it exists.
Thanks!