Using AI to generate an ICO white paper
With ICOs, or token sales of cryptocurrencies, having reportedly raised $2.3 billion in 2017, the time seemed right to use AI to generate a fake ICO white paper programmatically using a recurrent neural network (RNN) in the cloud.
I’ve titled it “RNN-Coin” and have published the full white paper here (pdf).
RNNs are a machine learning technique that have frequently and hilariously been used to generate arbitrary text from text-based training data. Some of my favorite examples are generating new British town names or a Game of Thrones book. However, unlike most town names or the work of George R.R. Martin, there are plenty of people that claim that ICO whitepapers are mostly incomprehensible anyway. This post shares some the technical background and some of the excerpts from the paper (it’s painful to read in full).
Building the model
Using Tensorflow, Google Cloud ML engine, and a modified RNN example I came across on Github from Maximilian Clarke, I created a model using a GPU-powered cloud training job that ran for a few hours (my MacBook Pro was too slow and RNNs are resource-intensive to train). The trained model could then output text that eerily echoed many ICO white papers.
For the published white paper, I generated around 2,000 characters at a time from 40-character sentence seeds I pulled from real white papers (like the Bitcoin white paper). I then manually organized the output into sections — formatted in LaTeX for full white paper legitimacy, of course.
One of the first outputs became the basis of the white paper abstract:
A purely peer-to-peer version of the company system with the price of the system to develop and the contract agent servers and the price of the system of the system which is a simple decentralized system which is the contract to provide the contract […] can be considered as a result of the blockchain.
The algorithm created peer-to-peer, blockchain-backed alternative to corporate structure! It’s just a simple decentralized system. The price is still developing, however.
The local revolution of experience is here!
RNNs have been described by some people as “magical” or “uncanny” and after going through this exercise I agree. There are points in the white paper that sound perfectly reasonable — and even decent marketing. What millennial, including myself, wouldn’t want to be part of “the local revolution of experience”?
It often generates language constructions that don’t sound natural. This could probably be tuned further by trying cleaner training data and different model parameters. Yet strange-sounding sentences also have gems:
This is the highest commitment wallet that a confident of money for medical currencies of the project in the initiative instruments of the project is the farmer of the cost of the system.
A “highest commitment wallet” might be a reassuring concept for an ICO investor in medical currencies.
“A pledge to the smart token” (and having fun with ML)
The first social advance is a problem for control of the platform.
I don’t see RNNs replacing real white paper writers or creating ICOs, but the experience of building and training fairly complex machine learning models that were previously the exclusive domain of academics and researchers does seem to be getting much easier for the average developer.
Much of the code for sample tensorflow projects using different ML techniques is freely available on Github and it seems cloud providers are now racing to rent GPUs to developers for model training purposes. I think as more tensorflow projects get shared the potential for collaboration also goes up: if a real AI/ML researcher comes across this post, pull requests are welcome to refine the model.
There is much speculation and uncertainty around the future of ICOs, but it also seems like there has never been a better time to generate fake ICO white papers with AI.
