Actually, I think there is a link between AI and the blockchain

There is a character flaw in some people (eg, me) which means when they see something that is obviously wrong on Twitter they feel compelled to comment. This is why I couldn’t stop myself from posting a few somewhat negative comments about an “infographic” on the connection between AI and the blockchain, even though I could have just ignored the odd combination of cargo cult mystical thinking and a near-random jumble of assorted IT concepts and gone about my day.

When it came down to it though, I just couldn’t. So, naturally, I decided to write a blog post about it instead. The particular graphic made a number of points, none of which are interesting enough to enumerate in this discussion, but at its heart was the basic view set out, here for example, that blockchain and AI are at the opposite ends of a technology spectrum: one fostering centralised intelligence on closed data platforms, the other promoting decentralised applications in an open-data environment. Then, as the infographic “explained”, the technologies come together with AIs using blockchains to share immutable data with other AIs.

Neither of those basic views is true though. Whether an AI is centralised or decentralised is tangential to whether it uses centralised or distributed data, and whether “blockchain” is used by centralised or decentralised applications is tangential to whether those applications use AI. What is important to remember is that decentralised consensus applications running on some form of shared ledger technology can only access consensus data that is stored on that ledger (obviously, otherwise you couldn’t be sure that all of the applications would return the same results). An AI designed to, for example, optimise energy use in your home would requires oracles to read data from all of your devices and place it on the ledger and then another set of factotums to read new settings from the ledger and update the device settings. What’s the point? Why not just have the AI talk to the devices?

There is, however, one part of the shared ledger ecosystem—of consensus applications running on consensus computers—that might benefit considerably from a shift to AI and this is the applications. People are very bad at writing code, by and large, and as the wonderful David Gerard observed in the chapter “Smart contracts, stupid people” in his must-read “Attack of the 50 foot blockchain”, they are particularly bad at writing smart contracts. This is clearly sub-optimal for apps that are supposed to send anonymous and untraceable electronic cash around. As David says, “programs that cannot be allowed to have bugs … can’t be bodged by an average JavaScript programmer used to working in an iterative Agile manner… And you can even deploy fully-audited code that you’ve mathematically proven is correct — and then a bug in a lower layer means you have a security hole anyway. And this has already happened”.

It seems to me that one thing we might expect AIs to do better than people is to write code. Researchers from Oak Ridge National Laboratory in the US foresee AI taking over code creation from humans within a generation. They say that machines, rather than humans, “will write most of their own code by 2040”. As it happens, they’ve started already. AutoML was developed by Google as a solution to the lack of top-notch talent in AI programming. There aren’t enough cutting edge developers to keep up with demand, so the team came up with a machine learning software that can create self-learning code… Even scarier, AutoML is better at coding machine-learning systems than the researchers who made it.

When we’re talking about “smart” “contracts” though we’re not talking superhuman programming feats, we’re really talking about messing around with Java and APIs. Luckily, last year saw the arrival of a new deep learning, software coding application that can help human programmers navigate Java and APIs. The system—called BAYOU—was developed at Rice University with funding from the US Department of Defense’s Defense Advanced Research Projects Agency (DARPA) and Google. It trained itself by studying millions of lines of human-written Java code from GitHub, and drew on what it found to write its own code.

Putting two and two together then, I think I can see that if there is an interesting and special connection between AI and “blockchain” then it’s not about using the blockchain as a glorified Excel spreadsheet that AIs share between themselves, it’s about writing the consensus applications for the consensus computers. They still wouldn’t be contracts, but they would at least work.

Leave a Reply

Your email address will not be published. Required fields are marked *