I’m feeling like I’m struggling a bit so sometimes it’s useful to just type out a thread with you all, here we go

The only problem is externality pricing.

Building https://negationgame.com

So, broadly, I’m trying to get mechanisms for governance and public goods funding up and running. 

It might seem like those are difficult to do at the same time, why not just pick one?

There’s a good reason

A compelling argument on behalf of there being lots of headroom if we figure this out is that our body works really hard to pump energy and oxygen to our extremities, and yet through market mechanisms we largely rely on something much more akin to mere osmosis for allocating resources to our extremities

Well, you can be the judge of whether that’s compelling. But I find it to be a useful argument.

The thesis with pg in the context of decentralized governance is that there will be more latitude for experimentation and, importantly, that that experimentation will be profitable from an energy perspective for the whole.

The challenge though is that its really hard to evaluate the value of a contribution in the absence of “counterfactual evidence”.

Of course, this is why markets are so lovely. Every moment is an experiment in testing the counterfactual question of “if this pizza slice wasn’t here, would people miss it?

But public goods (which don’t exist by the way) are way harder. 

E.g. you’re quite thrilled that there’s a police force that prevents a cartel from making you pay for protection, but we don’t want to have to find that out by having a trial period where we let the cartel come knocking at your door.

Oh yeah pg don’t exist is just a bit of cheek.

I’ll happily argue that every possible good is excludable for at least someone in the world.

The concept of public goods is still useful but I’m finding in designing mechanisms it’s more useful to think in gradients of accessibility and value superlinearity

This problem of needing to know the value of a contribution rears it’s head practically when we try to answer questions like “how much should we give as a grant to this project”

Retrofunding tries to avert this by looking in the rear view mirror, but that hardly answers the “what if not for this project” q

It also shows up in policy. E.g. let’s say you’re worried about the risk of fatalities on a jobsite, so you preemptively mandate the use of hard hats. If no one has a head injury  it becomes difficult to justify your policy choice, but of course it could be due to the policy that people aren’t getting hurt

So what you actually want to do is to infer what would happen if not for a certain policy.

The difference between those tells you the value of your policy 

E[injury | do(hard hat)] - E[injury | do(no hard hat)]

Of course, this is difficult

It’s difficult for like lots of reasons. The possibility space is massive. There are second order consequences for any policy (maybe workers decide you’re a sissy and disobey because there haven’t been injuries yet). You’d like to minimize cost (injuries) but also know the policy is necessary, etc.

This is where you could turn to statistics and historical data, or to storytelling, or to expertise.

What I’m interested in is game theoretic inference. We all know that markets are a game to infer the value of a stock.

What ideas like futarchy argue is that you can use markets to create games for finding the value of a plan.

I’m some sense, this is what markets are already doing, they’re putting a leader in position, who make some plans, and then the market says “yes I like this plan” or “no I don’t”

But futarchy goes further than this. It says the market should figure out the whole plan

This is both cool and crazy. It’s cool because if you achieve it then you massively parallelize human and artificial intelligence

Now you don’t work at one company, you can contribute to the workings of many companies. In some sense it’s like if traders were also calling the shots, in some sense they already do

It’s also crazy for that exact same reason.

We’ve all heard about famous trades that were detrimental to the company (or government) they bet on, but we’re lucrative to the trader.

This is the dark side of markets

Perhaps a good example would be if the policy of a company were on auction then you can imagine a competitor having a bit of strategic fun by spending a bunch of money picking something bad

“Exxon announces its plans to drill dry holes”

“Apple will now exclusively sell to people living above sea level”

One sparkling and vastly underutilized mechanism in preventing this outcome is staking.

Really if you haven’t read the PoS papers you should. They’re marvelous.

There’s something magical about only offering downside. 

(Fwiw asymmetry is also important in ml, as seen in activation functions and kl divergence)

Buuut all the behavioral value of staking comes from the threat of slashing. And you should slash if the person was “wrong”

In PoS, due to the ease of checking what’s right or wrong (hashes are deterministic) you can come to a consensus and slash

But most topics are not so clear, apparently even Earth’s shape

Another way to ask this question is “how might we create a generic PoS, one that can be used for any kind of question?”

There are so many good ideas in ml on how to do inference, counterfactual, and policy search, that I guarantee there’s a way to do it game theoretically

And I want to be the one to do it

But, it’s proving hard. Haha. The only information that we have access to is player’s stakes, and perhaps their willingness to slash a stake, and from that we’re trying to animate an entity

It’s proving hard. I need help

To mitigate risk and maximize learning rate, we’re building a much less ambitious tool. Just a discussion product that reveals information about users’ preferences and rationale. If we succeed, we’ll have a useful bottom up discussion tool that will be for daos and perhaps as a slack-like product for remote teams

And so far it’s been fun. Lots of learning, front end has been fun to write which I had no experience with

But I’d really much prefer to have a clear path for the underlying mechanism set. Naturally it’s hard, it’s arguably the most ambitious thing you could build, lim like turning all of humanity into one AGI

Or would that just be an HGI?

It seems unlikely that it’s impossible. I struggle to see an argument that there can’t be a bijection (or at least isomorphism) between AGI and HGI, especially given how sophisticated our basic units are (us).

So if AGI is possible then this is

And it gives me hope that humans can be part of the final machine we’ll find on the other side. 

For reasons I don’t know how to articulate, to get this to work without it becoming capturable you kinda have to spread the power around to the nodes, and that’s great news for us (and probably non-human animals too)

The parts I’m excited and proud about are:
1) we’ve maintained good velocity under resource constraints
2) this is way more plausible than 8 years ago when it started
3) index wallets are fucking cool, and the negation game’s algorithm is Turing complete and I think will be able to do online structural adaptation

The places I need help:
1) index wallets should be a totally viral idea, the fact that no one is talking about them means the comms aren’t sufficient, we need example stories told
2) the game theory of the negation game is so damn technical, it has to achieve active inference w mechanisms & |{⚙️} ∩ {ml}| ≈ ε

This was useful, the framing of “AGI using mechanism design” is new as of this thread, and is easy to follow.

I wanted to talk about other things, like how markets incent dishonesty (manifold is great at demoing this) and more about the issue of power diffusion, but it didn’t make it

Thanks for reading! Night

It turns out that you can basically simplify the problem to be solved down to the issue of energy allocation 

Pg funding is the part where you allocate the available energy of the group toward behavior that’s likely to return more energy

Governance is where you figure out which behaviors likely return a surplus