Original date published: November 15, 2024

Own your own AI or it will own you

We are closer than ever to AGI, but we are more focused on capabilities than ownership. Who will benefit from the economic gains from AI systems, and how do we ensure they are evenly distributed?

2023 marked the year AI entered mainstream public discussion. This is good and has gotten the public interested in the topic, but we are still more focused on capabilities rather than ownership.

AI will create incredible economic opportunities, but these gains will not be equally distributed and will be concentrated in the hands of the few who own this technology. Fortunately, the fundamental techniques for building these systems are generally known within the public domain and shared as research papers and open-source code. Yet there is still a significant constraint on general participation: compute. Compute is the most important currency in the development of frontier AI systems and accounts for a huge fraction of the performance gains over the years. Simply making the models bigger and scaling them with more compute can significantly improve their performance. This is why we have seen one of the greatest bull runs in NVDA's stock price.

The open-source community struggles to keep up with the large compute requirements of the highest-performing frontier models and relies on the benevolence of large tech companies such as Meta and Databricks to train these large models and make them open to the community. For these reasons, the open-source community is mostly focused on inference (making the models run faster after training). This has created incredible opportunities for startups and individuals to use techniques like fine-tuning and RAG to personalize these models for their use cases and businesses. Even with this limited window of progress, there is a strong push by the biggest AI labs and providers to regulate AI, citing reasons such as safety.

Individuals deep in the topic call this "regulatory capture." Citing unfounded future claims of doom should not allow individuals to dictate the fate of such a powerful technology so early in its infancy. We need to investigate techniques to make the training and inference of these foundation models more accessible to the public through grid computing, volunteer computing, and decentralized incentivization.

"Own your own AI or it will own you."

Last updated at: November 15, 2024