Artificial intelligence (AI) is going to be the most powerful tool/weapon in the world very soon. It is inevitable and we only get one chance at it. We need to do it right before someone does it wrong. We can ensure that no single person or organization can exploit it by open sourcing the AI in terms of:
- code and design
- CPU/GPU compute and memory
- data
The world’s compute is strongest and most resilient when it’s distributed. The largest supercomputers are much slower than the distributed network can be, as discussed in this article. Even some Googlers (and at least one Microsoftie) agree that the open source community will win the AI race. The resiliency of distributed systems is clear when you consider cloud computing.
In order to make it very hard for the models to be used for evil, we have to make the models large enough for only our distributed system to be able to handle the training and eventually the querying.
Another benefit of decentralizing the compute is it allows anyone to contribute to the leading AI research, rather than just employees at OpenAI and Google.
To get more compute, we can incentivize it. Cryptocurrency mining has so much compute we can tap into, but the main motivation for crypto mining is currently money. The main downside to monetization is that it can lead to centralization (e.g. Coinbase). We can avoid this by keeping the entire stack decentralized. This means we must make the entire stack easy to use:
- Sufficient documentation for AI Engineers
- Sufficient documentation for software engineers
- Easy use for compute providers
- Easy interaction with the AI: both APIs and user interface
Money goes directly from customers to computers. We can copy how Bitcoin/Etherium work where only compute providers get any money. We can make the model public, but a waste of money to implement. It will be easier to just use the existing model. We want to implement both the training and access of the model via distributed compute. The API requires payment per use and the website/app have advertisements unless the user pays to remove advertisements. The first payment option will be Etherium.
Given the difficulty of defining general AI and our lack of understanding of the human brain, we may achieve strong AI simply by increasing the model size and integrating it more with the real world environment. We may even consider ChatGPT-4 to be strong AI.