This was solid. Principled framing, opinionated take within it. I don’t agree with all of it, but none of it is trivial to counter. Learned some specific good insights, like why diffusion works better in robotics.

The big dichotomy of giant leap vs small steps perhaps overstates what the giant leap side is attempting. I suspect they’re doing their stunt demos in big leap ways but their deployment training in small steps ways. Giant leap feels like “AGI” misframe of robotics, so the right comparison is not to LLMs. Ie cue the “Giant Leap achieved internally” jokes. It may be a strawman. I don’t know enough about the companies classified as giant-leapers to judge. But there certainly is a camp that believes scraped video and simulation data are enough.

The commitment to vertical over horizontal seems more practically interesting. Ideologically, I’d prefer horizontal (and open source) stack layers, but the argument that historically things start vertical is correct. What’s more “horizontal” is kinda unsolved because transferring embodied learning laterally is unsolved. That said, neuroplastic adaptation in brains to changed dynamics is a thing, as are adaptive and robust control under at least parametric variation (mass, length etc. within a given structure). I’d bet fine-tunable horizontal models within broad morphology classes are possible.

Humanoids are a divisive topic. I need no reason to like them besides the cool factor (and hopes of cheap old age support) but they certainly need an updated case beyond Asimov’s “UX of built environment is human.” The unstable-under-power-loss point feels oddly trivial though. I can think of at least a few reasonable mitigations. It’s a totally manageable issue, not a show-stopper.

Commenting paywalled but in case nobody else caught it: the quote attributed to John Carmack was actually John Salvatier.