Hacker News Reader

Comment by SR2Z | Hacker News Reader

There is some value in producing a lot of solid hardware, but nowhere even close to Tesla's absurd valuation.

I think they are perfectly capable of writing software to drive the robot - if Musk doesn't stick his head in like he did with LIDAR/FSD and impose some stupid requirement that handicaps the product.

View on HN

But the whole shtick with Optimus is that they aren't writing software. It's supposed to be all LLM training so when you buy your robot you can give it orders like "do the dishes", "clean the gutters", "dig a backyard pool for me", or "build me another Optimus" and you can go off to do whatever while it completes the task.

Elon thinks it would be too expensive to have to write code for every task you might ask one of these to do, they want it to be fully autonomous.

Their engineers aren't behind keyboards typing C++, they're wearing VR headsets and feeding the data to a LLM, although even that is probably too specific for Elon's long term plans. Obviously he doesn't want to have to have people repeat actions hundreds of times before the dumb robots figure it out. Especially for "simple" tasks like serving drinks at press events.

You say this like it's a bad idea. These VLA models are going to be even more disruptive than the coding models because otherwise it's prohibitively expensive to set up an industrial robot for most uses.

My main doubts about Tesla's plan are that they will sell enough of these to get benefits of scale or that Musk will force the engineering team to "skip lidar" again and compromise the design.

It's yet another gamble where if it works out he will look like a visionary and if it doesn't he'll look like an idiot. The exact sort of bet that Elon never fails to go for.

> My main doubts about Tesla's plan are that they will sell enough of these to get benefits of scale or that Musk will force the engineering team to "skip lidar" again and compromise the design.

Indeed.

Even with 9 million total cars sold, Tesla still has yet to solve for driving safely with no interventions across just the contiguous USA.

With a similar approach, a million robots operating for years is still a long way short of gathering the data needed for training an AI to autonomously operate safely in a full range of industrial environments.

(That said, IMO remote-controlled humanoid robots still make a lot of sense, they'd only need a little bit of AI to assist rather than to do everything; if I was in Musk's position, I would be selling that vision of the future rather than claiming fully autonomous AI-driven androids are anywhere near).

I feel like society is decades away from being comfortable with "you can go off to do whatever while it completes the task"...regardless of whether or not the tech is there.

It's just the AI singularity discussion again. AI Techbros insist it will be here before the end of the decade. Like you I am skeptical about it. I tend to think AI capability is already plateauing and ever more effort is going to be spent chasing smaller and smaller returns.

I'm experiencing AI that is very fast, but also kinda dumb and thoughtless.

But how would we evaluate "perfectly capable" without evidence, there's just been no evidence they've done anything so far right? Am I missing something? I guess looking closer it was only announced four years ago. But it seems like it's only been smoke and mirrors so far.

I think FSD is very impressive, even if it is still pretty unsafe.

Tesla clearly has at least some AI chops, and if Musk can bullshit for long enough, they might have enough time to make these robots more than just props.

Reply on news.ycombinator.com