AI Predictions

Long time no post!

I have been spending an ever-increasing amount of time in the AI field as of late, and I figured it might be worth measuring how accurate my understanding of the rate of change in this industry truly is. I'll make a couple predictions in the 1-5 year time frame, and for fun I'll try to predict when we will stumble across AGI. All these acronyms made me realize that this website will probably be less accessible than I'd like it to be, but I'm sorry I just want to finish this post - maybe I'll add an acronym table later. Generally, just search google like so: <acronym> AI meaning.

In a later post I may go over my reasoning for these predictions, but for now I'll just leave them here.

Predictions:

1. There will be a 1 Trillion parameter LLM in the open-source domain within two years.
   • June 14th 2022: Google open-sourced a 1.6 Trillion parameter mixture of experts LLM, beating my prediction by 18 months
2. Within three years, a combination of multiple trillion-parameter multi-modal models, each trained for specific tasks, will become world-class experts in most fields. Eg. LLM base + Physics model + Diffusion model = world-class physicist expressing knowledge through art.
3. Within five years, the above model will be abstracted to a general purpose model that is a world-class expert in all fields.

Long-term prediction:

A further abstraction of the general model mentioned above leads to AGI in 10-20 years. We have until then to figure out the AI-alignment problem to ensure this development is net-positive for society.