jamesough

Here's an intuitive explanation for L1:

Imagine you have two planetary bodies fixed in space. A free mass is attracted to both, but it's possible to find an unstable equilibrium somewhere between them. Put the mass here, and it'll stay still-- but it's like balancing a ball on the top of a sphere; it'll roll off towards one of the two planets unless you stabilize it.

Now imagine displacing the mass from the equilibrium, but in a direction perpendicular to the line between the planets-- it'll swing back and forth like one of those slingshot rides. There are two dimensions perpendicular to the line between the planets, so two dimensions with stable oscillations, therefore you can get circular orbits.

> I thought it’s a given that since both CNNs and CAs are Turing complete, you can use one to simulate the other.

They are both Turing complete, yes, but the similarities are deeper. Recurrent CNNs are a very natural way to describe CAs because they both impose exactly the same prior: that the system's underlying laws are spatially quantized, local, and translationally invariant.

There hasn't been much work done in this area. People have made the connection before (e.g. this paper from Sutskever: https://arxiv.org/pdf/1511.08228.pdf) but AFAIK this is the first work to actually look deeply into how well you can recover CA rules with neural nets and gradient descent.

If we are to believe Steven Wolfram's proposition that the universe is a cellular automaton at its core, then learning the rule by gradient descent seems pretty obviously useful, and not something that anyone has really tried yet.

To be more explicit: applying gradient descent to cellular automata has a reasonable likelihood to being the only way to uncover the fundamental laws of reality, and essentially nobody is working on it except OP. So it's a useful paper!

'Alright so here’s the gameplan:

1. Elon removed as CEO by SEC order - can’t be CEO of a public company

2. Tesla stock drops 50%

3. Tesla stock becomes so cheap they take it private

4. Musk reinstated as CEO'

https://twitter.com/AustenAllred/status/1045554529914089472

Yes! This was very much inspired by Shannon and Thorp. I love that story.

Yeah, it's not anon: the game was retired from machines and I left the ROM out. Wouldn't recommend to do for other than fun (and tbh hourly wage would be very small), but was a fun side-project!

Yep.

It's not anon. The game was retired from machines and I left the ROM out. Hourly rate probably wouldn't be very high if you did try to actually make money. Fun challenge though!

The game was recently retired from machines and I left the game ROM out. Fun as a challenge though!

Yeah, it's not anon: the game was retired from machines and I left the ROM out. Wouldn't recommend to do for other than fun (and tbh hourly wage would be very small), but was a fun side-project!

Nah this was really just a fun side-project, just posting as proof-of-hack. Also this specific game was retired from machines. Fun as a challenge, though!

V2 will be available as a custom suppository.

Many hours of searching through internet archives of dead forums. I don't think you can find it any more.

I got asked to be on a University Challenge team after someone saw me playing... but completely flopped when they asked independent questions at trials.

This kind of unlinked data leaks very fast, too: at my peak I had to top up with Anki for about 3 hours a day. I've forgotten almost all of it now.

A small consolation is that I can still tell you the 'exact number of gallons of water' in most major lakes.

Yes, and the decryption code had been bundled into the binary! I tried inspecting it as you say but it seemed easier to just solve it by hand.

They mix up the answers. And sometimes choose 4 answers from a possible selection of 20.

I tried this initially and put all 30K questions in Anki. I learnt about 97% of them, but it still wasn't good enough to win

It's legally not a gambling machine -- it's a 'game of skill', which means you can put it in a pub without a license. Gambling laws don't apply.

Looks like a case for <git-filter-branch>

V2 will be available as a custom suppository.

These are typically found in university towns in the UK.

Yes! This is because of UK legislature. You don't need a gambling permit if the game is skill-based.

Accuracy is high enough!

Heard on grape vine that IBM claim to have pulled first silicon for a memristor RPU (design like below) just over a month ago.

https://www.nature.com/articles/s41586-018-0180-5

Thanks. Can you recommend any books on these topics?

What technology needs to be developed to get the data on synapses and neurotransmitters? A high-resolution Raman imaging microscope? And is a dead brain sufficient, or would you need a real-time noninvasive scan of a living one?

The interesting meta-point here is the chicken-egg problem between cheap robotic hardware and large-scale data collection.

We don't have general-purpose robotics because there aren't enough robots being built for them to become cheap.

And the robots can't do anything in the real world because they cost too much for someone to gather 1m hours of training data.

Presumably this is the Softbank game plan -- build a ton of general-purpose robotic hardware platforms, and hope that they (or another group) figures out how to build general-purpose algorithms in the next few years.

It's a calculated risk that Softbank have the resources to bankroll.

Yep, it all ends up as heat. Energy is conserved, so if it didn't end up as heat, it would have to go somewhere, and there isn't really anywhere else it could go.

Related topic: reversible computing. You might wonder if a computation need take any energy at all: it turns out that an irreversible computation (e.g. NOR, XOR, AND etc) is physically guaranteed to waste energy, but if you make sure your compute steps are always reversible (i.e. each input maps to one and only one output) then you can theoretically compute for free (the Feynmann lectures on computation cover this well).

But these energies are negligible compared to the wattage that goes through a standard CPU.

Depends on UWB bandwidth+power, how good your processing is + how many antennas you have... You could also get way better with a fleet of trucks spaced over a large area, or a city-sized swarm of drones, or satellites...

Yeah this is why you'd use UWB -- the capacity is insane. Also the raw information theory calculation makes sense only if you have zero prior information about the sample you're reconstructing. But 'people in buildings' is a pretty strong prior. And it's not like you have to do any extra work, just put a simulated raw signal through a neural net and regress to groundtruth data.

But let's do the math, I'm curious!