justindrake

> how does the system recover if that were to happen?

The Ethereum 2.0 consensus layer gracefully falls back to a slightly biasable form of randomness called RANDAO. This is merely a "weakening" of the consensus, not a breakage. Applications built on Ethereum 2.0 (such as lotteries) may indeed break with RANDAO biasability.

> why is it assumed that someone won’t mount a long range attack to break the system by engineering a 100x faster VDF?

One of the things we are doing is trying to prove lower bounds on the circuit depth of modular squaring. (Ryan Williams is working on such lower bounds right now.) On the hardware engineering front, getting a 100x speedup is extremely difficult. Improvements to sequential speeds in traditional ASICs has tapered off at around 5% per year (see for example https://github.com/preshing/analyze-spec-benchmarks). And if possible, being able to increase the speed of multiplication by 100x would be an amazing development for humanity (multiplication is arguably the key basic operation in computing).

> Given the competition is only a ~$15m investment in a group of hobbyists

We can afford a few experts with $15m :)

> we’ve already seen order of magnitude improvements over and over again with proof of work in Bitcoin

Keep in mind that Bitcoin is an energy consumption game, not a latency game. Latency improvements do not follow Moore's law and other related laws.

> If there’s a reward for computing a VDF faster than everyone else

There's no direct reward for being a bit faster than everyone else. If you are much faster (say, 100x faster than the general public) then you may be able to bias the randomness, and therefore manipulate things like lotteries built on Ethereum.

> The article claims that proof of work is costing Ethereum hundreds of millions of dollars per year, but this VDF competition is also wasting millions of dollars in human capital

It's about orders of magnitude. Ethereum burns about $0.5B per year. The VDF project is about $15m, and the ASIC should last about 10 years.

> if it's not embarrassingly parallelizable, first, tell me why

It's a cryptographic assumption that repeated squaring in an RSA group is "inherently sequential". This assumption was first made in 1996 by Rivest, Shamir and Wagner for timelock puzzles. See https://people.csail.mit.edu/rivest/pubs/RSW96.pdf

> what's all this business about optimizing the algorithm

To link sequential computation with physical time we need the general public to evaluate the VDF at roughly the same speed as an attacker. By "roughly the same speed" we mean that the attacker (even with a huge budget) cannot build hardware that runs, say, 10x or 100x faster than a publicly available ASIC.

One efficient verification scheme is by Benjamin Wesolowski (see academic paper [here](https://eprint.iacr.org/2018/623.pdf)).

The scheme is surprisingly simple and fits on just one slide. See https://www.youtube.com/watch?v=zqL_cMlPjOI&feature=youtu.be...

ELI5 version: 100 people, one by one, (re)roll dice placed in a dark room. After the last person lights turn on, revealing a fair random number. The Verifiable Delay Function (VDF) ensures lights aren't turned on early.

For a video explanation of the randomness mechanism see https://www.youtube.com/watch?v=zqL_cMlPjOI&feature=youtu.be

The ultimate goal is indeed to build an ultra-fast ASIC. There's a $1M ASIC circuit competition planned for for 2020. This $100K FPGA competition is a "warm up" to the ASIC competition.

A bit more information here https://www.youtube.com/watch?v=9-77V1IareQ&feature=youtu.be...

Ballpark latency numbers per 2048-bit modular multiplication:

CPU—1100ns

FPGA—66ns

ASIC—1ns (expected)

If you're a hardware engineer (or mathematician) interested is building an ultra-low latency ASIC, please email [email protected]

Source: https://twitter.com/drakefjustin/status/1123143447551053824

Context: https://www.youtube.com/watch?v=zqL_cMlPjOI