Thread by @alexandrosM

I've been hinting in various conversations about a distributed news aggregator design I've been sitting on for about a decade or so. I'll describe it here so I have somewhere to point to, for feedback, and if I never get a chance to build it, maybe it will inspire someone else.🧵

My quest begun with writing this paper back in 2009. We didn't know about the bitcoin white paper that had gotten released a months earlier, but we sketched out several pieces of what came out of the crypto ecosystem over the next dozen years. arxiv.org/abs/0907.2485

The obvious weakness was governance in the human layer. So long as moderation of any community can't be distributed, all the distributed infrastructure in the world won't help you when your Admins with God Powers get subjected to rubber hose cryptanalysis. h-i-r.net/2009/02/rubber…

As I kept digging deeper, I described the main problem: any algorithmically guarded objective "center" would become gamed by intelligent and motivated actors. The arms race between human and machine is lopsided, but not in the direction you might think. researchgate.net/publication/24…

What I didn't know at the time, but learned later, is that I was describing a sub-case of Goodhart's law, as applied to large-scale networked systems. en.wikipedia.org/wiki/Goodhart%…

Spending the next few years thinking about this problem, I concluded that there are 2 possible ways around the problem.

1. peer-to-peer systems with skin-in-the-game
2a. multi-channel triangulation
2b. predictions (aka. motion parallax of time)

Let's walk through them:

1/ Peer-to-peer systems with skin in the game are exemplified in the Web of Trust concept. The main innovation here is that there is no central object to capture, and every agent speaks only for itself and its connections. en.wikipedia.org/wiki/Web_of_tr…

With no central "home page" or other global object, Sybil attacks lose their potency, especially when the agents in the system keep track over time and losing a long-aged/highly-trusted account becomes costly as the agent has to start over from scratch. sciencedirect.com/topics/compute…

2. While this kind of approach will get us to a place somewhat similar to Twitter (minus the "trending" centralized object), we want to build a community that is more truth-oriented than Twitter.

This is where triangulation and "the motion parallax of time" (aka. predictions) come into play. The shared, un-gameable global object all participants have access to is called "underlying reality", featuring the passage of time.

https://twitter.com/alexandrosM/status/1423646231390081028?s=20

In the same way a pigeon bobbing its head simulates binaural vision, someone building a predictive record with a history of being more correct than others, creates a hard-to-fake signal for the objective value of this account.

And much in the spirit of the ELO rating system in chess, observers can then use these predictions to construct a mostly-objective ranking of the various pundits in their field of vision. It may require them to be able to mostly-correctly identify whose predictions have...

...been correct in the past, but that is a far easier task than predicting the future or choosing between different pundits arguing over an issue without access to prior track records. It's up to each agent to evaluate others' track record, and they have skin in the game to lose.

The system I'm describing aims to be more like science than religion, where the crucial difference is that different schools of scientific thought can diverge for a while, but ultimately converge, whereas religions only ever diverge and never converge (barring extinction).

With the philosophical matters out of the way, let's get more practical as to how such a thing could be implemented: I'm imagining a single-user node that would run in a Raspberry Pi, very similar to PiHole or many other projects we support at @balena_io pi-hole.net

These nodes would identify themselves using a distributed DNS solution such as .eth or unstoppabledomains.com. @brave browser supports these domains so by installing that browser in the users' phone or laptop, they can access the network from anywhere, through their own domain.

Each node starts life with a very small set of known other nodes (much like a Web of Trust setup, possibly needing someone else to induct/vouch for a new person joining the network). However, the nodes all expose their "activity feeds" to their "followers". (sounds familiar?)

Through these feeds (probably implemented with something like ActivityPub), the nodes beome aware of more and more other nodes, until they have a rich set of connections to many diverse nodes in the network.

The UI each user sees may look very similar to Hacker News, but the voting would work differently. An upvote or a downvote implies strengthening or weakening the connection of the user to the other user(s) that were the conduits to get any particular item visible to this user.

The ultimate result of such a dynamic, distributed, weighted and directed graph would be for the overall system to start operating as a bayesian network. Cycles may occur, and can be resolved using known methods from BN literature. en.wikipedia.org/wiki/Bayesian_…

This level of implementation gets us up to the "echo chamber" level, a better version of Twitter that is also fully distributed. The next layer to implement is the prediction layer.

Users should be able to ask for concrete, falsifiable predictions on active questions, and others should be able to give responses, with a % confidence, much like @slatestarcodex does every year, but daily. slatestarcodex.com/2019/01/22/201…

Over time, users' nodes should be able to compute the accuracy other users in the network by leveraging resolutions that their user has made of specific past questions with what they perceive to be the answer.

In fact, the nodes should be able to amplify a small amount of past resolutions into far stronger signal by assuming transitivity over the network (trust the resolutions of people I trust). As a result, the user should have to anchor just a few facts during any period of time.

In this way, the network can act as a CAPTCHA: A user that gives nonsensical answers eventually gets isolated from the network, while users giving reality-correlated answers find each other and continue to increase their entanglement with reality.

The most important thing to remember here is that in all instantiations of this network, users have skin in the game: sharing low-quality material gets others to distance themselves. Even having a bot-network voting in your favor just means you get isolated with your bots.

And on the predictive axis, which acts as a counterbalance to the well-understood strategies of sharing clickbait etc, predicting badly should have a similar effect. The only way to have longevity on the network is to predict at least decently, and make quality contributions.

If one happens to destroy their account by making silly predictions or contributing negative-value content, they can always start over, but they have to start over from zero.

On top of this social infrastructure, I believe many other applications can be built, not just a twitter/hacker-news clone as I've described here. But it's important to have a concrete use case to start from.

If this is something you'd like to work on please DM me, ideally with evidence of high-quality work you've done in the past.

Alternatively, if you have other feedback, I'd be glad to hear it!

Correction: binocular vision, not binaural. Obviously. Here's another pigeon gif to compensate you for the suffering I inflicted upon you.