> It's hard not to extrapolate 50 years down the line with this tech. Stuff gets complicated quick.

Indeed, though looking at the rate of change in various fields, I feel everything goes weird some time around 2032 or so: https://benwheatley.github.io/blog/2024/03/23-17.24.34.html

> How many human neurons do you have to use before you get an actual human consiousness?

How many grains of sand makes a heap.

(I don't know the relative importance of organic neurons as hardware vs. whatever specific connectivity architecture they happen to have in us, so the same might apply to perceptrons).

From Your Blog: > No Musk ⇒ No Mars

Sputnik 2.0 is near.

China AI and chip investments are off the charts. Note that China is scheduled to attempt a Mars sample return mission with Tianwen-3. They are scheduled before any other competing team. China has the industrial support of the non-democratic government. They are willing to bear the risk for their future. This might be more stable funding then commercial SpaceX or high-deficit-USA-NASA.

There's a massive gap between "probe" and "colonise". I meant the latter, and I don't see that kind of risk-taking from China — their first few crewed launch was not even announced as such until after the launch itself had succeeded, implying they were not willing to risk the optics of that failing.

I think there's a reasonable chance the actual Sputnik 2.0 will be them making a permanent lunar base before the US does. Musk doesn't seem to care about going to the moon beyond it helping prove (and fund) the Starship project, even though Starship could put a station there in a handful of landings.

I also think the moon is a better choice for the same reason I think China will prefer it over Mars: when things go wrong, it's much, much easier to act like your emergency rescue mission was the actual mission plan all along — "No, the 阿波罗-十三 module didn't have an explosion in the oxygen tank, we were simply venting unnecessary resources as part of a planned exercise. This mission was never intended to land on the Moon.", that kind of thing happens on a Mars mission, everyone just dies.

> China AI and chip investments are off the charts

That's as may be, but it doesn't change my point with any of that. The best reaches atomic resolution in (by my guess) 2032… and then what? China isn't going to beat that.

AI is limited by electrical power, both for training (data centres) and inference (for cars and other real-time robotics). The power envelope is whatever it is, but I'm expecting 5x hardware efficiency improvements for a fixed power envelope by 2030 (slower than Moore's Law used to suggest). The algorithmic efficiency also improves at the same time, which I'm assuming brings it back in line with Moore's rate, but that still means going from a car (with ~1kW spare for compute) to a robot (with 100W spare for compute) will take 5 years (or 10 years if you assume 3kW for the car and 30W for the robot). And the global power grid is 2TW, which is 250 W/person, so humanoid robots can be in an awkward place of driving up demand for electricity so much people literally can't keep the lights on while taking all our jobs and yet still be less than half of all labour.

And that's equally true regardless of if China does or doesn't take a lead over the US, or if the EU gets organised and has its own EUV fab , etc.

Fun :P

>How many grains of sand makes a heap

A minimum of four, usually much more though.

(yes that is a tetrahedron joke)

I thought it was a Banana Splits theme tune reference.

The best way I heard it was:

How many times N can one lie before being called a liar?

The question investigates a discrete phase or category or state change.

Asking how many neurons are needed before we “get” a consciousness is a bit like asking how many people are needed before we have a country.

The assumption that you obtain a human consciousness by using human neurons is itself flawed—physics does not say it works that way, and theory of mind philosophers are divided on the matter.

How does physics say it works?

I guess you could use any material, that behaves the same as the material, that is used in humans, right ?

Since we've yet to figure out a way to "measure" consciousness, it depends on what "theory" of consciousness you subscribe too. I believe something must make use of quantum effects such as entanglement in order to have qualia (eg. Penrose models). So, ChatGPT running on deterministic hardware is just a bigger mechanical clock, but the organoinds have qualia. At least insofar as a worm or chicken have qualia. Add enough of them of and you might get consciousness. I don't consider this line of research ethical, but philosophical objections won't provide much of a disincentive.

Physics says "we don't know what this word means", which is exactly why it's a flawed assumption to say "human neurons are necessary" or "human neurons are sufficient".

That’s a great way to put it.

Alex Jones will have a field day with that one.

This arguably has more functioning neurons than he does.

I love the comment on that video: "Imagine you wake up and you're just doomguy"

Do you have to use human neurons?

No, in the video they are using rat neurons (they're the cheapest)

You also have to keep them alive. Neurons don't last very long in vitro, particularly once you wire them up to electrical probes to measure them.

This sort of problem is also an issue with things like the Neuralink chip and other brain implant technologies.

Neural organoids can be cultured and maintained for months at a time (we grew fairly attached to the one my coworker was growing). Even the measurement wiring can be sustained for months.

I've worked in biology and computation for decades (and fascinated both with intelligence and artificial intelligence the whole time), but seeing a timelapse of a bunch neurosphere forming over a few weeks from some tiny stem cells, with all the neurons sort of self-organizing themselves by spreading lamellapodia (https://www.youtube.com/watch?v=fniW9EiOmUk) a few years ago flipped a switch in my head. I used to think the development of the hardware to support intelligence was a really tough, complicated problem. It's clear that the structure of neurons is absolutely primed to build semi-random networks that can be adapted to signal processing and computation in an almost "trivially easy". Obviously there's tons going on under the hood at the molecular, cellular, and organoid levels

It's funny but some technologies have a way of going from "sci-fi" to "ho-hum" quickly, and I think "rat neurons playing DOOM" is a great example of that.

Looking a little at the neuroplatform, it looks like they grow the organoid in a microfluidic container and can maintain it functionally for months. Seems like an opportunity to sell the computer as a subscription (new organoid package every month?)

> we grew fairly attached to the one my coworker was growing

This tickles me. One interpretation of the “when is it conscious” question is it’s really asking “when do we start to care about it,” and I love that the answer for humans is often “as soon as we start spending time with it.”

> It's funny but some technologies have a way of going from "sci-fi" to "ho-hum" quickly

I’m a programmer working in biotech, and the things that have become de rigor in the field are the stuff of sci-fi from a decade ago. It’s incredible.

And somehow related I think, but the lab grown meat industry is nowhere near feasible

The feasibility of the lab grown meat industry has very little in common with how hard it is to grow brains in petri dishes; the closest they get is "both are cells and both are possible".

Different cells, different arrangements, different volumes, different lifetime requirements before the cell lines are allowed to die, and different price concerns for the result.