[ben_w]

> FABs are another issue entirely. Re terafab. It's very difficult to audit silicon. Just cutting the top off a chip layer by layer and using an electron microscope, isn't going to fully give you trust. People have been worried about their designs being modified at the fab for a long time. Ah well.

Sure, sure, but that's not why I used terafab as a question for "what if competent AI?"

The details on that are… ugh. They were clearly done for glossy magazines and posters rather than with thought, wouldn't be surprised if it was vibe-written or done by a teenage intern or by someone high on drugs.

But if you ignore all that and treat it as a serious proposal, a billion(!) Optimus robots with a competent AI is Musk having personal control over a workforce the size of China or India's adult population.

[b112]

Yes, it can be disconcerting to consider such. In anyone's singular control. And that number of robots is indeed sort of the plan.

This is a loooong podcast, but was just great from start to end. Multiples of the neurolink team were interviewed, including Musk at the start.

https://podcasts.happyscribe.com/lex-fridman-podcast-artific...

I think Musk's interview is maybe 30 minutes? Certainly less than an hour. In it he does dive a bit into robotics, and how he'd like robots everywhere. I believe he talks about how valuable they could be for things not considered yet, like just walking deep into forests and monitoring the local environment, sampling, etc.

Anyhow, it's the first thing which came to mind when you mentioned a billion robots. On that front, I think we're maybe 8 years out from mostly-competent robots. Certainly, local LLM compute will be super cheap even in a few years, so no issues with a robot having multiple LLMs onboard. But I do wonder, if we can't do competent self-driving, will we have competent robots?

I'd peg robots at years after perfect self-driving, and we're years from that. Robots would need to navigate far more random, uncertain terrain/locations.

[ben_w]

> On that front, I think we're maybe 8 years out from mostly-competent robots.

I think this depends on what is meant by "mostly-competent".

As you say:

> I'd peg robots at years after perfect self-driving, and we're years from that. Robots would need to navigate far more random, uncertain terrain/locations.

I'd put a 10 year gap between "self-driving car of quality X" and "humanoid robot able to get into driving seat of car, drive it at quality X", just on a power-envelope basis.

For this part, I'd have to disagree:

> Certainly, local LLM compute will be super cheap even in a few years, so no issues with a robot having multiple LLMs onboard.

While some LLM will be able to fit (and already can because some fit on a phone), the AI we have now in LLMs (and VLMs) is much too spiky for this use. Recent Anthropic and OpenAI models have just given me half a dozen different confident identifications for the same plant, many of which were easily falsified even by comparing what was in the image to what the AI's text output claimed it had seen in the image, like the leaf shape and how many came out of each node.