Hacker News new | past | comments | ask | show | jobs | submit login
Safe Artificial General Intelligence via Distributed Ledger Technology (arxiv.org)
10 points by godelmachine 14 days ago | hide | past | web | favorite | 9 comments

I retain faint naïve optimism for blockchain's long-term impact, but I really feel this paper overstretched trying to impose blockchain on the AI safety domain.

An example is the author's 3rd proposed axiom for Safe AGI in Table 3: "Morality, defined as no use of force or fraud, stored via DLT."

I applaud the sheer optimism, but found it difficult to take the paper seriously beyond page 6.

Blockchain is in 99.9% of cases a solution looking for a problem

It is hard to grasp for non technical/economics minded folk so it is easy to exploit as a magic snake oil solution for hard problems

I’ve had very painful conversations with people who control very large budgets at companies that are convinced blockchain is the next big thing when they only have a rudimentary concept of what it is and can do, but think it will revolutionize the payments industry or how supply chains are structured

It’s a distributed database that is more expensive and time consuming to maintain than a centralized database but if structured correctly can remove the need for a set of players to trust a central authority or one another (you pay for this in speed of read/write and storage making it nearly useless in payments industry)

There are a lot of solutions looking for problems out there that are getting a lot of funding due to the glut of cash looking to seed the next big thing in tech, and the ease of finding and investing in small companies now

AI is perhaps the best meaningless umbrella term for raising money or generating corporate interest currently

AI safety is a very loosely defined term because of the vague nature of AI

Two welders, accountants, plumbers, farmers are going to have similar jobs, two AI safety engineers could be coding anything from video game algorithms to nuclear plant turbine shutdown code and failsafes

Anything hardware running any code could be broadly considered AI

Combining blockchain and AI safety is just mashing together two buzzwords and hoping something interesting comes out with no seed of true inspiration germinating a fruitful idea

> you pay for this in speed of read/write and storage making it nearly useless in payments industry

You are ignoring the so-called “off-chain” solutions embodied by lightning.

Lightning is for continuous or repeated transactions between two parties because it doesn’t settle immediately/at end of day like the vast majority of one off transactions conducted in the world

Ultimately using lightning, “Bitcoin banks” would develop because that is the most efficient way to do it with large volumes of trades between disparate parties involving cash, credit, and gold

Instead of opening individual lines of credit with the 1000 people I paid or who paid me last month (lightning wallets) and adding/subtracting from those lines of credit instead of settling (nothing is written to the blockchain), it all goes through central authorities and arbitrator - my banks and credit card providers, who tally all the transactions

Lightning is not a scalable solution without centralized lightning banks which defeats the entire purpose of the blockchain in the first place - it’s just added complexity to arrive at virtual banks as a solution

What you describe is payment channels, not lightning. Lightning allows multi-hop payments so you don’t have to have channels with your endpoints, just be connected to the same network graph.

This might honestly be one of the stupidest things I've ever heard of. This was clearly written by someone who understands neither AI nor blockchain technology. Can't wait to hear about their ICO...

I think the author could benefit from getting feedback from some people who work in ai safety full time.

This is the greatest grant money magnet of this generation.

Is work like this commonly funded? I have seen many papers like this spread out over many fields and they seem to have been tailored to support pitching for grants.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact